Levin Controls Cell Bioelectrics to Regrow Limbs Without Gene Editing

Levin利用细胞生物电调控实现无须基因编辑的肢体再生 非神经组织将形态学目标模式储存于生物电吸引子状态中，使短暂的离子通道药物治疗即可触发完整的器官再生，而无须改动基因组。 在本次主题演讲中，Michael Levin论证了生物电计算在所有体细胞组织中均有运作，其出现远早于特化的中枢神经系统且独立于后者存在。他证明，操纵内源性细胞电压模式可作为复杂形态发生子程序的软件级触发器，使研究人员能够在不编辑基因组的情况下诱导异位眼生长、使完整肢体再生，甚至跨物种屏障改变头部形状。Levin阐释了生物组织如何作为具有稳态目标设定点和稳定生物电吸引子状态的动力系统运行，且这些状态在经历剧烈的身体重塑时仍能保持。最终，他敦促计算机科学家和人工智能研究人员采纳这些基础认知与多层体细胞可塑性原理，以构建更具弹性、适应性更强的合成架构。 非神经体细胞组织利用古老的生物电网络做出大尺度的形态学决策，其运作方式犹如区别于底层基因组硬件的计算软件。 (1:35) 短暂操纵静息膜电位即可触发复杂的模块化器官发生（例如生长出形态完整的异位眼或肢体），而无须进行基因编辑或干细胞移植。 (20:57) 目标形态作为稳定的生物电吸引子状态存在，不仅在身体遭受损伤后依然存续，还能在无性繁殖再生的无限循环中保持改变后的解剖学记忆。 (24:08) 生物电回路表现出类似图灵模式的自发对称性破缺，且无须即刻的转录改变，其作为主子程序触发器，处于生化和机械级联反应的上游。 (49:36)

Michael Levin argues that biological computation is not unique to brains, but an ancient somatic mechanism where non-neural cell collectives use bioelectric networks to store target anatomical setpoints and coordinate large-scale morphogenesis. Pattern Homeostasis and Anatomical Setpoints (9:41) Developmental biology often conceptualizes pattern formation as an open-loop, feed-forward process driven by gene regulatory networks. Levin counters that biological patterning functions as a closed-loop homeostatic feedback system that actively detects and corrects deviations from an anatomical setpoint. In 'Picasso tadpoles' engineered with craniofacial organs scrambled in abnormal locations, eyes and jaws follow novel, unnatural migration paths to produce normal frog faces, demonstrating that tissues actively detect target configurations rather than following hard-coded trajectories. This homeostatic fidelity persists despite radical structural remodeling. In holometabolous insects, learned associations survive pupal metamorphosis despite the complete histolysis and restructuring of the larval brain. Similarly, planarian flatworm tail fragments retain conditioned spatial memories after decapitation and subsequent brain regeneration. For Levin, these phenomena reveal that somatic tissues maintain macroscopic anatomical representations and goal-seeking capabilities across severe physical disruptions. Bioelectric Networks as Morphological Subroutines (20:57) Non-neural cells communicate through membrane voltage gradients ($V_{ ext{mem}}$) mediated by ion channels and gap junctions, utilizing the same electrical machinery that specialized neurons later adapted for rapid signaling. Fluorescent voltage-reporter dyes demonstrate that spatial bioelectric pre-patterns appear before physical organogenesis and downstream gene expression, serving as a functional blueprint for anatomical outcomes. Manipulating these bioelectric states operates as a master subroutine call rather than micro-managing cell differentiation. For instance, pharmacologically opening specific potassium channels in a tadpole gut instructs the local tissue to construct a complete, functional eye complete with retina, lens, and optic nerve. Furthermore, altering the bioelectric state of wild-type *Dugesia japonica* fragments causes them to regenerate head anatomies characteristic of distant species diverged by roughly 150 million years, confirming that physiological software can override genomic defaults without genetic mutation. Latent Engrams and Rewritable Attractors (24:08) Multicellular bioelectric circuits store patterning instructions as dynamical attractors. Levin demonstrates that morphologically normal, single-headed planaria can carry a silent, latent bioelectric engram specifying a two-headed anatomy. The animal's physical morphology remains standard until amputation occurs, at which point the tissue deploys the encoded state to regenerate heads at both ends. This bioelectric pattern memory is both permanent and reversible. Once established, two-headed planaria cut in unmanipulated water continue regenerating two heads indefinitely across successive generations without genomic alteration. However, targeted bioelectric intervention can overwrite this memory, resetting the circuit back to the single-headed attractor state. Levin frames these bioelectric networks as connectionist systems capable of storing alternative anatomical memories independently of the underlying genome. Clinical Regeneration Without Genomic Editing (28:17) Because bioelectric states function as high-level control triggers, therapeutic interventions can be remarkably brief and avoid viral vectors or gene editing. Applying a tailored ionophore cocktail to an amputated adult frog limb for just 24 hours pushes the tissue into a regenerative attractor state, triggering nine months of autonomous tissue remodeling that yields a touch-sensitive, motile limb with claws and digits. Predictive computational simulations of ion channel networks have also been used to rescue severe craniofacial and neurological birth defects. When tadpole embryos suffer catastrophic brain malformations induced by exposure to teratogens like nicotine or alcohol, or via mutations in the developmental gene *Notch*, targeted small-molecule channel modulators restore endogenous electrical gradients. The resulting tadpoles develop normal brain architecture and recover cognitive performance in associative learning assays, illustrating that physiological signaling can compensate for underlying genetic damage. Multiscale Competency and Non-Neural AI (47:04) Levin proposes that artificial intelligence and robotics can overcome architectural brittleness by embracing basal cognition and multiscale biological competency. While engineered machines typically fail outside narrow operating envelopes, biological systems exhibit autonomous problem-solving at every scale—including molecular networks, single cells like *Lacrymaria olor*, somatic tissues, and colonial organisms. In experiments grafting ectopic eyes onto tadpole tails, the central nervous system spontaneously learns to decode the novel rear visual stream without evolutionary pre-training. Levin envisions the eventual development of biological compilers that translate high-level anatomical CAD designs into specific bioelectric inputs, directing cellular collectives to assemble synthetic living machines. Rather than restricting machine learning to neural emulation, he advocates extracting the computational principles used by somatic tissues to optimize shape, solve physiological problems, and navigate unpredictable environments. Levin concludes that non-neural somatic bioelectricity provides an accessible software layer for guiding complex pattern formation and tissue repair without viral delivery or genomic editing. However, mapping the intractable inverse problem between low-level molecular networks and high-level anatomical shapes remains the primary computational challenge requiring interdisciplinary machine-learning partnerships.

Tufts大学的Michael Levin教授在NeurIPS 2018上指出，生物计算与认知并非大脑的专属产物，非神经体细胞早在大脑演化前就通过古老的生物电网络协调形态发育与器官修复。他认为，破译并重写这种生理学“软件”不仅能革新无基因编辑的再生医学，还将为解决脆性问题、开发具自适应能力的非神经人工智能提供全新范式。 超越大脑的记忆与形态稳态 (2:49) 形态发生通常被传统生物学视作由基因表达引发的开环前向过程，但Levin主张生物形态更接近一个闭环的稳态目标控制系统。即使经历极端的物理重构，生物体依然能维持目标结构与记忆。例如，完全变态昆虫在蛹期大脑几乎完全溶解重组，成虫却保留了幼虫习得的条件反射；被斩首的涡虫在数天内完全由尾部再生出功能完整的大脑，并重现训练记忆。此外，将胚胎分割会发育成同卵双胞胎，融合两个早期小鼠胚胎则形成嵌合但比例正常的个体；蝾螈可无限再生四肢与器官并在达到正确比例后精准停长，鹿角每年以每天高达1.5厘米的速度再生带神经的骨骼，7至11岁儿童的断指也有自然再生记录。 在“毕加索蝌蚪”实验中，研究团队将蝌蚪的面部器官人为移植至异常位置（如眼睛移到后脑或下颌侧移）。变态发育过程中，这些器官并未沿基因预设的机械轨迹移动，而是主动探索全新的位移路径，最终校正并重组成结构完全正常的青蛙面孔。Levin指出，这表明基因组编码的并非僵化的解剖动作，而是一个带有误差检测能力的形态稳态基准（Set Point），只要实际物理结构与基准不符，细胞群便会持续重塑，直至误差归零。 生物电网络与器官级子程序调用 (12:39) 当代分子生物学虽能详尽绘制基因交互网络，却仍无法仅凭测序数据反推生物的三维宏观几何形态，Levin将这种瓶颈比作计算机早期的“硬件打孔卡”时代——科学家总在重组微观硬件，缺乏对高层形态控制算法的理解。其实，所有体细胞均通过离子通道和缝隙连接（Gap Junctions）维持膜电压梯度（Vmem），在宏观上形成生理生物电回路。这种电信号早于神经系统演化出现，也是真菌化合物（如裸盖菇素）能直接识别人类神经受体的原因。借助荧光电压敏感染料，研究人员可在早期胚胎发育中直接观测到指导器官定位的生物电分布图，甚至能在实体肿瘤形成前侦测到恶性细胞因去极化与电通讯阻断而退化至单细胞状态。 利用这一原理，研究人员无需微观干预每个干细胞的分化，只需调控整体电位便能触发复杂的形态构建“子程序”。例如，在蝌蚪肠道细胞中开启特定的电压门控钾通道，就能诱导整个组织重塑为包含视网膜、视神经和晶状体的完整异位眼睛；在斩首并截尾的涡虫中调整局部静息电位，即可根据指令诱导出两头或无头的异常个体。在此类干预中，外部输入的仅仅是粗粒度的高层触发信号，底层的精细装配则由组织已有的形态发育子程序自主执行。 形态记忆印迹与表型空间重写 (22:21) 生物电状态并非仅反映当前的物理结构，还以动力系统稳定吸引子（Attractor States）的形式存储着潜在的形态发生指令。实验表明，未编辑基因的正常单头涡虫经短暂生物电扰动后，其体内便写入了“双头”形态印迹。当该涡虫被切断头尾并在普通清水中多次断体再生时，它将永久地持续再生出双头个体；而通过针对性的生理学处理，又能将这种双头表型逆转回单头状态。这证明了在基因组完全一致的前提下，多细胞网络能够长久稳定地保存并随时读取不同的离散宏观形态记忆。 更具颠覆性的是，在不改动Dugesia japonica野生型基因组的情况下，仅改变截肢断面的生理网络连接，再生出的头部即可呈现出进化距离达1.5亿年的其他物种特征——如Polycelis felina的扁平头或Schmidtea mediterranea的圆形头。进一步的生物电扰动甚至能驱使涡虫长出自然选择中从未出现过的复杂形态，例如三维尖刺结构、带内部空腔的帽状构造或背部长出的管道。Levin强调，在基因硬件之下存在着一层完全受生理软件调控的广大表型空间（Morphospace）。 无基因编辑治疗与生物编译器前景 (28:17) 在再生医疗实践中，成年青蛙截肢后通常仅能长出微小的软骨刺。Levin团队仅用靶向离子通道的化学药物鸡尾酒对伤口进行24小时的局部处理，就成功诱导了自主且持续9个月的再生级联反应，长出了带脚趾、爪子且具备触觉与运动功能的完整后腿。类似地，因酒精、尼古丁等致畸剂或Notch关键基因突变而面临前脑缺失和中脑空洞等严重脑缺陷的蝌蚪，在计算机模拟推算出正常电位并通过小分子药物纠正离子通道后，不仅解剖结构得以挽救，其关联学习和认知行为指标也恢复至健康水平。 与目前主流的CRISPR基因编辑或病毒载体相比，利用内源性离子通道重塑电生理状态具有极高的临床安全性，不仅避免了基因永久突变与生物安全风险，而且干预窗口极短。展望未来，Levin提出了“解剖生物编译器”概念：设想研究人员利用计算机辅助设计（CAD）绘制出目标组织或器官的三维蓝图，由编译器自动逆向推导出实现该目标所需的最小生物电刺激参数。由于生物电网络本身具备类似图灵反应-扩散系统的自发对称性破缺能力，其不仅能作为上游指令调控BMP等下游化学形态发生素，还能直接驱动宏观对称与结构的分化。 多尺度智能与新型非神经人工智能架构 (34:35) 面对现场关于信号干扰与认知边界的提问，Levin澄清形态生物电信号是以分钟、小时乃至数天为单位缓慢演化的稳定直流状态，与毫秒级的肌肉收缩和神经动作电位处于不同的时间尺度频道，因此调控长期解剖电位并不会打乱个体的日常运动与捕食行为。他强调单细胞纤毛虫（如Lacrymaria olor）无需神经系统即可精准伸缩颈部捕食，植物根系借助生物电避障，甚至细菌生物膜也会利用类似大脑的钾离子波实现群体协调，这表明认知在生物界呈现出连续谱系（即基底认知），并不存在人类心理学设立的绝对壁垒。 当前的人工智能系统普遍存在严重的脆性，在脱离受训场景后极易崩溃；相比之下，生物系统从分子网络、单细胞、多细胞组织到器官，在每个层级均具备自主解决未知环境问题的能力（多尺度竞争力）。在合成生物学探索中，从原本语境中剥离的细胞群能自我协调构建出全新生命机器；即便把眼睛移植到蝌蚪尾部，大脑也能迅速学会解析后方传入的视觉信号。Levin呼吁计算机科学与机器人学界跳出传统的脑神经连接主义范畴，汲取非神经组织凭借动力学吸引子追求终极目标、自适应修复物理损伤的运算机制，共同开发兼具多层鲁棒性与自我重构能力的新型人工智能与机器人形态。 Levin总结认为，生命通过深层生物电软件实现了高层形态与认知控制。目前的挑战在于如何建立从低层分子参数到高层几何输出的逆向映射计算工具，以及将两栖类成果跨物种平移至哺乳动物；这为机器学习构建新型非神经架构、与发育生物学联合开发“生物编译器”提供了广阔的合作空间。

## Background 本视频由 YouTube 频道 The Artificial Intelligence Channel 于 2018 年 12 月 5 日发布，记录了塔夫茨大学艾伦探索中心主任迈克尔·莱文（Michael Levin）教授在 NeurIPS 2018 大会上的受邀特邀演讲。视频页面简介指出，莱文教授身兼范内瓦·布什终身讲席教授及发育生物学中心主任，演讲重点探讨了生命系统在神经系统之外介导的形态与行为信息处理机制。 ## TL;DR 莱文教授在演讲中指出，在真正的“大脑”演化出来之前，生物组织就已经在利用一个古老的、非神经元的跨细胞生物电网络进行计算和决策。他通过毛毛虫变态保留记忆、涡虫切块再生、毕加索畸形蝌蚪面部器官自组织重构等实例，揭示了解剖形态发生本质上是一个具备目标导向与误差校正能力的闭环稳态系统。研究团队通过调控内源性离子通道与跨膜电位，成功实现了诱导肠道长出眼睛、重写双头涡虫记忆、诱导青蛙断肢再生以及逆转基因突变引起的脑缺陷，且全程无需基因编辑或干细胞移植。最后，他向人工智能领域呼吁，打破过度模仿人类大脑的狭隘局限，借鉴多尺度生物体细胞的高弹性与自适应容错特性，以开发出更具鲁棒性的全新非神经计算与机器学习架构。 ## Atmosphere & style 整场报告呈现出一种严密理性与颠覆性想象力交织的学术震撼感。莱文教授语调冷静从容、叙述层层递进，用海量惊人的真实活体实验数据不断打破传统遗传决定论的固有认知；从五腿青蛙到双头涡虫，每一个生动例证都引发台下计算机科学家的极大兴趣。问答环节讨论深入且充满前沿思辨，在生物工程、哲学意识与人工智能交叉的前沿阵地中，洋溢着开拓全新科学范式的思想火花与智力张力。 ## Bulletin - [01:35](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=95s) Michael Levin 提出体细胞生物电网络作为一种古老的非神经计算层，早在脑结构演化出现之前，就已在生命系统的各个尺度上通过离子通道与缝隙连接指导复杂解剖决策与形态发生。 - [28:17](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1697s) 生物形态受闭环目标导向系统控制，通过短时（如24小时）药理学靶向改变内源性跨膜电压，可触发长达数月的下肢完全再生或逆转严重的脑部先天畸形，全程无需干细胞移植或基因组编辑。 - [24:06](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1446s) 涡虫实验证实形态印记具有真正的多稳态记忆特性：改写中间体段的生物电预图式后，不仅能诱导再生出双头涡虫，且在后续反复切除头部中该双头形态均永久保持并世代自维持。 - [49:36](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2976s) 非神经细胞网络在不依赖基因转录改变的前提下，纯靠生理电动力学即可自发产生对称性破缺并形成类图灵斑图，同时作为上游层级直接指挥下游的化学形态发生素梯度。 ## Key takeaways - [01:35](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=95s) 早在神经系统演化形成之前，所有体细胞就已共享离子通道、电突触与神经递质等生物电硬件，用于在组织宏观尺度上执行解剖决策。 - [02:49](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=169s) 生物系统在经历剧烈物理重构时仍能保持跨介质的信息存储，如毛毛虫变态为蝴蝶后保留学习记忆，被斩首涡虫在仅剩尾部再生新脑后仍能恢复既有记忆。 - [09:41](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=581s) 解剖形态发生并非僵化的开环前馈程序，而是一个具备误差感知与纠错能力的闭环稳态系统，拥有极其严密且可重写的目标形态设定点。 - [09:41](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=581s) “毕加索蝌蚪”实验中错位生长的面部器官能自主调整轨迹直至形成正常青蛙面容，证明形态控制由宏观解剖目标驱动而非机械设定的固定运动位移。 - [12:39](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=759s) 传统发育生物学目前陷入“只操纵微观硬件（基因与分子）而忽视宏观控制软件”的困境，亟需借助机器学习反向破译微观网络到宏观解剖形态的映射算法。 - [15:34](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=934s) 神经系统的演化本质上是对古老体细胞大尺度信息处理系统的“速度优化”，前者以毫秒级响应指导机体行为，后者以分钟至天为单位指导解剖形态决策。 - [20:51](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1251s) 研究团队通过非侵入的药理学、光遗传学调节内源性离子通道与拓扑连接，即可触发器官级子程序，如在蝌蚪肠道诱导出功能完整的眼睛，或诱导涡虫长出跨越1.5亿年演化差异的其他物种头部。 - [24:06](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1446s) 生物电图式在组织内充当形态记忆印记，具备长期稳定性、可重写性与延迟召回特性，对应于动力学状态空间中的稳定吸引子，微小瞬时扰动即可完成系统状态锁定。 - [28:17](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1697s) 仅需 24 小时的靶向离子通道药物鸡尾酒干预，即可在不使用干细胞或基因编辑的情况下，诱导青蛙截肢创口在接下来的 9 个月内自主再生出具有触觉与运动功能的完整下肢。 - [28:17](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1697s) 基于生物电动力学仿真推算组织跨膜电位，团队成功筛选小分子药物逆转了由致畸化合物或 Notch 基因突变引起的灾难性大脑畸形，完全恢复其脑解剖结构与学习记忆能力。 - [37:56](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2276s) 将眼睛移植至蝌蚪尾部后其依然能够实现正常视物并学习，展现出中枢神经系统在面对全新输入源时所具备的惊人适应性与计算可塑性。 - [47:04](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2824s) 生命系统具备自下而上的“多尺度智能”，从分子网络、单细胞到组织器官各层级均拥有自主计算、自适应与目标导向能力，为突破当前 AI 脆弱性与分布外泛化瓶颈提供了全新工程范式。 - [49:36](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2976s) 生物电网络本身就是原生的反应-扩散动力学系统，无需底层基因转录参与即可自发电学对称性破缺，并在形态发生层级上作为上游开关直接统领下游的化学形态发生素。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=0s) 讲者介绍与大会致辞 在 2018 年的 NeurIPS 大会上，来自 Facebook AI 研究院和德克萨斯大学奥斯汀分校的计算机科学教授克里斯汀·格劳曼（Kristin Grauman）作为大会程序主席之一，主持并介绍了受邀演讲嘉宾迈克尔·莱文（Michael Levin）教授。莱文教授是塔夫茨大学（Tufts University）的生物学教授，并持有范内瓦·布什（Vannevar Bush）冠名终身讲席。同时，他也是塔夫茨大学艾伦探索中心（Allen Discovery Center）的主任，该中心致力于研究生命系统介质中的计算。 莱文教授拥有计算机科学和生物学的双重学术背景，于 1996 年在哈佛医学院获得博士学位。他的研究团队处于发育生物学、认知科学和计算神经科学的交叉领域，尤其关注这些研究在再生医学中的应用。格劳曼教授指出，莱文教授的工作不仅对生物学意义重大，还能为机器学习和机器人技术带来全新的灵感。他的研究曾荣获多项大奖，包括被《自然》（Nature）杂志评为上世纪发育生物学里程碑之一的“左右不对称性分子基础”研究，以及美国心脏协会（AHA）的杰出研究者奖和国际功能性电刺激学会（IFESS）的远见科学家奖。 格劳曼教授在致辞结束时热情地表示： > “因此，我们非常激动能听到他今天的报告，请大家与我一起欢迎迈克尔·莱文教授。” --- ### [01:35](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=95s) 核心观点：无脑介导的生物计算 迈克尔·莱文教授登台并对观众的欢迎表示感谢。他直接抛出了整场演讲的核心信息：早在真正的“大脑”演化出来之前，生物体就已经在多个尺度上进行计算了。所有身体组织中的决策，实际上都是由一个古老的、非神经的电网络所介导的，这个网络由身体内的所有细胞（而不仅仅是神经元）共同组成。莱文教授认为，利用这种非神经的认知能力，是开发新型、鲁棒的人工智能技术的一个尚未被开发的兴奋前沿。他非常期待与机器学习领域的专家合作，将生物电的研究从再生医学拓展到更多令人兴奋的新应用领域。 在接下来的演讲大纲中，莱文教授规划了几个主要部分：首先是脑与身体的塑料性（Plasticity）；其次是身体认知（Somatic Cognition）的实例，即复杂解剖结构在自我编辑过程中的决策制定；接着是细胞和组织向大规模目标重塑背后的生物电机制；最后，他将展望再生医学、人工智能及相关领域的未来。 莱文教授在开场时强调： > “我想告诉大家的是，早在脑出现之前，生物就已经在多个尺度上进行计算了，实际上，所有身体组织中的决策都是由一个古老的、非神经的电网络介导的，这个网络由所有细胞组成，而不仅仅是神经元。” --- ### [02:49](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=169s) 脑与身体的塑料性：毛毛虫与涡虫的记忆奇迹 为了展示生物体令人难以置信的塑料性，莱文教授首先举了毛毛虫变态发育的例子。毛毛虫在生命初期是一个爬行并咀嚼植物的“软体机器人”。在蜕变为蝴蝶的过程中，它的神经系统（包括大脑）基本上会被液化、拆解，然后重新组装成一个完全不同的、能够控制“硬体飞行机器人”（即蝴蝶）的大脑。然而，在这个神经系统被彻底重构的过程中，记忆却奇迹般地保留了下来。毛毛虫时期学到的东西，变态发育后的蝴蝶依然记得。这向我们提出了一个挑战：在介质经历如此剧烈形变的情况下，信息是如何被存储和读取的？ 另一个更极端的例子是涡虫（Planaria）。这是一种神奇的扁形虫，如果你切掉它的头（或身体的任何部分），它都会重新长出来，包括重新长出一个完整的大脑和所有其他器官。莱文教授的团队在几年前的研究中表明，如果你训练一只涡虫记住某个特定区域是它的家，然后将其斩首，只留下尾部，当尾部重新长出一个全新的脑袋后，这只涡虫依然保留着之前的记忆。这意味着记忆以某种方式被印刻在了再生出来的新大脑中。 这一现象最早是由詹姆斯·麦康奈尔（James V. McConnell）在 20 世纪 60 年代发现的，但此后多年无人问津。当莱文教授的团队重新证实这一现象并发表论文时，在互联网上引发了巨大的轰动。 ```mermaid flowchart TD accTitle: 毛毛虫变态发育与涡虫再生的记忆保留机制 accDescr: 展示了毛毛虫和涡虫在经历身体结构剧烈重组时，记忆如何跨越介质得以保留。 subgraph CaterpillarMetamorphosis ["毛毛虫变态发育"] A["软体爬行阶段 (毛毛虫)"] -->|"学习与记忆形成"| B["特定记忆存储"] A -->|"变态发育 (脑部液化与重组)"| C["硬体飞行阶段 (蝴蝶)"] B -->|"记忆跨介质传递"| C end subgraph PlanarianRegeneration ["涡虫再生过程"] D["训练阶段 (涡虫记住家园)"] -->|"斩首 (失去原有头部和脑)"| E["仅存的尾部组织"] E -->|"再生过程 (重新长出头部和脑)"| F["拥有新脑的完整涡虫"] D -->|"记忆印刻与传递"| F end ``` 莱文教授在谈到这一现象时指出： > “在整个神经系统被完全拆解并重新组装的过程中，记忆实际上得以保留，因此毛毛虫学到的东西，蝴蝶仍然会记得。” --- ### [04:38](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=278s) 身体认知与单细胞的惊人掌控力 莱文教授指出，认知存在一个光谱，从极简单的形式一直延伸到我们希望用机器学习和人工智能模拟的复杂认知。自然界中各种各样的生物都处于这个“认知天梯”的不同位置。他展示了一个名为大泪丝虫（*Lacrymaria olor*）的单细胞生物在真实时间下的活动视频。这个单细胞生物在寻找、捕食环境中的其他微小生物时，对自身结构展现出了令人难以置信的动态控制力。 大泪丝虫在没有大脑的情况下，将行为、结构控制和生理学完美地集成在了一个单一细胞内。这表明，单细胞生物在实现各种复杂目标时，已经具备了极高的能力。而当细胞联合起来形成多细胞身体和复杂的解剖结构时，它们并没有失去这种做决策和展现复杂行为的能力。 例如，在青蛙胚胎的发育过程中，细胞协同工作以构建极其复杂的结构；在斑马鱼神经系统的发育中，所有细胞也在协同构建一个远超单个细胞物理边界的宏大结构。我们每个人都始于一个单一的受精卵，正是通过胚胎发育过程，这些复杂的形态才得以可靠地产生。 莱文教授在展示单细胞生物的视频时感叹道： > “但是行为、结构控制和生理学都集中在一个细胞中，这里并没有所谓的脑，因此单细胞生物和单个细胞在实现各种有趣的目标方面可以表现出令人难以置信的能力。” --- ### [05:58](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=358s) 胚胎发育的鲁棒性与再生之谜 莱文教授强调，理解解剖结构的构建不能仅仅依靠干细胞生物学。干细胞确实能够分化出身体所需的各种具体组织，但如果缺乏大尺度的三维结构控制，就会产生一种被称为“畸胎瘤”（Teratoma）的肿瘤。这种肿瘤内部虽然有头发、牙齿、皮肤和肌肉，但它们杂乱无章，完全失去了正常的三维空间结构。因此，我们必须理解细胞是如何“决定”要构建什么，以及它们是如何执行这一构建过程的。 发育生物学教科书中的标准观点是一个“开环前馈系统”：基因调控网络产生蛋白质，蛋白质相互作用，细胞执行功能，最终涌现出复杂的解剖结构。但实际情况要有趣得多。首先，胚胎发育具有极强的调节能力，能够应对巨大的外界干扰。例如，将早期的哺乳动物胚胎切成两半，每一半都能重新构建出缺失的另一半，从而发育出完全正常的同卵双胞胎。反过来，你也可以把两个小鼠胚胎像捏雪球一样揉合在一起，最终依然能发育出一只完全正常的小鼠。 这种塑料性不仅存在于胚胎阶段，在成年动物中同样存在。例如，美西螈（Axolotl，一种蝾螈）在成年期依然能再生出它的眼睛、下颌、四肢、脊髓、卵巢甚至部分大脑。 莱文教授指出，再生最神奇的地方不仅在于它能“启动”，更在于它知道何时“停止”： > “关于再生，最令人惊叹的不是它能够开始，而是它知道何时停止。” 一旦正确的解剖结构被重新构建出来，进一步的生长和重塑就会停止，而我们目前对这一机制的理解还非常肤浅。 此外，前文提到的涡虫甚至被切成 275 块后，每一块都能精确地重新长出缺失的部分，不多也不少，最终形成一只完美的涡虫。这些涡虫甚至克服了衰老，在个体细胞衰老死亡的同时，动物整体通过不断再生实现了“永生”。最后，莱文教授提到，这种再生能力并非低等动物专属，人类的肝脏同样具有极强的再生能力。 - **胚胎鲁棒性实例**： - 哺乳动物胚胎一分为二：发育为正常的同卵双胞胎。 - 两个小鼠胚胎融合：发育为一只正常的小鼠。 - **美西螈（Axolotl）可再生部位**：眼睛、下颌、四肢、脊髓、卵巢、部分大脑。 - **涡虫（Planarian）再生极限**：可被切成 275 块，每块均能长成完整、无衰老的永生个体。 ### [09:41](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=581s) 闭环稳态与毕加索蝌蚪：形态学计算的目标导向性 生物体在面对严重物理损伤时展现出的修复能力令人震惊。鹿每年都会重新生长出高度神经化的鹿角骨骼，其新生骨骼的生长速度可达每天1.5厘米。人类也保留着部分惊人的再生潜能：早在20世纪70年代，科学家就发现大约7岁到11岁以下的人类儿童能够完全再生被切断的指尖，不过这种再生能力随着年龄增长便会消失。演讲者幽默地提醒台下听众：“大多数在座的人可千万别亲自尝试这个。”这些现象揭示了生物形态的控制机制绝非简单的开环系统，而是一个具备误差检测与校正能力的闭环系统。一旦形态发生偏离，系统便会在转录水平以及细胞物理重建层面触发调整，通过恒定的稳态环路引导组织恢复目标形态。演讲者所在中心的核心研究目标，正是去破解并重写这种稳态过程的“解剖学设定点”（anatomical set point），以此重新指定细胞应当构建的形态结构。 ```mermaid flowchart TD accTitle: 形态稳态的闭环控制系统 accDescr: 展示生物体如何通过闭环反馈检测形态误差并进行物理重建以达到目标解剖形态 subgraph Sense["形态感知与检测"] A["当前解剖形态"] --> B["误差检测机制"] SP["形态学设定点<br>(Target Morphology)"] --> B end subgraph Actuate["纠偏与重建响应"] B -->|"偏离设定点"| C["转录水平变化"] B -->|"驱动细胞行为"| D["物理重建过程"] C --> E["细胞增殖/迁移/重塑"] D --> E end E -->|"持续反馈直至误差清零"| A ``` 生命系统按照图纸“规格”施工的能力甚至超越了常理。在蝌蚪发育为青蛙的过程中，面部会经历剧烈的重构：双眼向前并拢，鼻孔位移，整张脸型完全重组。当研究人员人为干预制造出眼睛长在背上、下颌偏向一侧的“毕加索式畸形蝌蚪”（Picasso-like tadpoles）时，令人惊奇的一幕发生了：尽管所有器官的初始位置完全错乱，它们依然能够沿着极其不寻常的轨迹自主移动，最终绝大多数仍旧重构成了完全正常的青蛙面部结构。这一发现强有力地证明了基因组并没有为每个器官设定一套死板固定的机械位移程序。Michael Levin 总结道：“它们实际上能够判断自己何时处于错误的位置，并持续移动，直到构建出正确的图案为止。”早在50多年前的蝾螈实验中也发现了类似的宏观控制能力：将蝾螈尾部嫁接到被截肢的肢体创口处，数月之内，由于系统致力于恢复整体的大尺度全局秩序，这个嫁接的尾部竟会被彻底重塑为一条正常的肢体。这种大尺度的决策并非单个细胞自行决定变成肌肉或皮肤，而是属于整个生物体层面的高级计算。 - 鹿角作为高度神经化的骨骼组织，年生长速率可达每日1.5厘米。 - 7至11岁以下的人类儿童具备指尖自然再生的生理能力。 - 毕加索蝌蚪实验表明，面部器官重塑过程属于目标驱动（goal-directed），而非执行机械设定的预设运动轨迹。 - 蝾螈尾部移植到截肢残端后，可在数月内逆转并重塑为肢体表型，证明大尺度形态控制拥有压倒局部细胞身份的纠偏能力。 --- ### [12:39](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=759s) 发展生物学的认知盲区与生物“软件”层面的缺失 如果剖析一个正常的人类躯干横截面，会发现所有内部器官与几何拓扑结构的位置都惊人地恒定。然而，生命最初仅仅是一个受精卵和一团均一的胚胎干细胞，这种精准且高度一致的宏观几何结构究竟从何而来？基因组中包含的 DNA 编码的只是蛋白质分子，它并未直接绘制几何形状，也没有附带关于形态该如何编辑的图纸。细胞群究竟如何得知自己该制造什么形状、又在何时停工？作为工程师，人们迫切想探寻能够将这一过程推向多远、生物可能性的边界又在何处。然而现实情况令人惊讶：传统发育生物学在这些关键问题上几乎一片空白。 如今的生物学极度擅长梳理低层次的分子相互作用网络，清楚哪个基因产物与哪个产物结合，却根本无法单凭阅读一段基因组序列就推断出最终有机体的三维几何形状。面对疾病或畸形时，科学家手头堆满了底层的分子通路数据，但由于这些极其复杂的交互网络，很难反向推导求解出“逆问题”（inverse problem）——即为了在系统层面实现特定的宏观解剖形态，到底需要在分子层级进行怎样的精确干涉。这也为机器学习提供了巨大机遇：除了传统的药物筛选，算法应当帮助人类破译从微观分子数据到宏观解剖表型的映射机制。 ```mermaid flowchart LR accTitle: 计算机科学与生物学控制层级的对比 accDescr: 对比计算机科学已实现软硬件解耦，而生物学当前仍停留在直接篡改底层硬件代码的阶段 subgraph CS["计算机科学 Evolution"] CS1["硬件布线调整<br>(早期重构物理插线)"] --> CS2["机器码 / 汇编"] CS2 --> CS3["高级抽象语言<br>(软硬件解耦/平台无关)"] end subgraph Bio["生物学现状与未来"] Bio1["分子/细胞操控<br>(直接重构物理硬件)"] -.->|"亟待突破的算法鸿沟"| Bio2["解剖学控制算法<br>(软件层重编程形态)"] end ``` 生物学领域目前陷入了与早期计算机发展相似的困境。早期计算机编程必须手动重构物理线路板，随后逐步发展出汇编机器码，最终走向了完全独立于具体底层硬件的高级算法与计算逻辑；而当下的生物学依然停留在“修改硬件”阶段，每次实验都忙于敲除基因或改变底层细胞结构。人类尚未建立起针对生物形态控制算法的认知框架，忽略了生物体内同样存在的“硬件与软件”的划分。形态稳态系统表现出了典型的认知特征：针对特定目标的导向性、面对外界扰动时的鲁棒性与弹性，以及将低层细胞功能整合为复杂宏观产物的能力。 > “在生物学中，我们正变得非常擅长操纵分子和细胞——那些控制复杂图案所必需的最低层级的硬件；而我们尚未真正理解的，是足以实现我刚才展示的图案编辑实例的控制算法。” —— Michael Levin --- ### [15:34](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=934s) 体细胞生物电网络与古老神经机器的起源 生命系统究竟如何执行宏观形态决策？答案在于细胞之间的生物电通讯（bioelectrical communication）。与大脑的神经运作极为相似，生物体的体细胞组织同样会构建起大尺度的电网络来进行计算与决策。区别仅在于：大脑网络所做出的决策表现为机体的行为，而体细胞电网络做出的决策则直接决定了机体的解剖形态。为了验证这些电计算模型，研究团队必须在实验室中通过理性干预诱导出解剖决策的改变。Levin 特别展示了一只活生生的五腿青蛙照片，并向观众强调这绝不是 Photoshop 图像处理软件合成的产物，实验室里生活着许多外形奇异的真实活体动物，正是用来严格验证生物电图案编辑的理论预测。 从演化史来看，大脑绝非凭空发明了电信号这一套复杂的运行机制。神经元是在多细胞生物诞生之前就已经在古老单细胞生物中运行的原始计算机制上演化而来的。早在那时，单细胞祖先就已经演化出了成套的离子通道、突触传导机制以及神经递质。正是因为这套生物电机制无比古老且跨物种高度保守，许多真菌所含的化合物——例如迷幻蘑菇中的致幻成分——即使在演化路径上与人类相隔极其遥远，却依旧能够强烈干扰人类的大脑认知。神经系统的演化本质上只是将早已存在于普通体细胞中的大尺度信息处理机制进行了“速度优化”（speed-optimized）。 ```mermaid flowchart TD accTitle: 神经元电网络与体细胞生物电网络的机制映射 accDescr: 展现大脑神经回路与体细胞电网络在底层机制上的相同硬件逻辑以及在输出目标上的差异 subgraph SharedHardware["保守的古代生物电硬件"] IC["离子通道<br>(Ion Channels)"] ES["电突触 / 缝隙连接<br>(Electrical Synapses)"] NT["神经递质系统<br>(Neurotransmitters)"] end SharedHardware --> BrainNet["大脑神经回路"] SharedHardware --> SomaticNet["体细胞生物电网络"] BrainNet -->|"毫秒级高速运算"| DecBehav["输出：动态行为 / 认知记忆"] SomaticNet -->|"分钟至小时级发育引导"| DecMorph["输出：解剖学决策 / 形态图案"] ``` 在构成原理上，全身所有体细胞与脑神经元别无二致：它们都表达相同类型的离子通道，通过电突触与相邻细胞相连，并实时产生大尺度的电学动态模式。以显微荧光成像记录的斑马鱼大脑为例，当斑马鱼处理视觉信息时，大脑神经元会迸发出剧烈的电信号活动；而在一段经过压缩加速、时长约8小时的早期青蛙胚胎发育录像中，使用电压敏感荧光报告剂（fluorescent voltage reporter）伪彩色成像同样可以清晰观察到体细胞间激烈的“电对话”。胚胎在尚未长出神经系统之前，正是依靠这套全组织的膜电位模式（$V_{\text{mem}}$ pattern）来统筹决策谁该变成前端、后端、背侧或腹侧。正如神经科学家致力于破译大脑电信号如何表征记忆与行为一样，发育生物学当下的核心任务是破译体细胞的“生物电密码”，搞清膜电位状态的空间拓扑分布究竟如何映射并决定了生命未来的宏观解剖形态。 - 体细胞利用离子通道建立跨膜电压，并通过缝隙连接形成互联的非神经元电网络。 - 真菌代谢产物之所以能改变人类认知，是因为离子通道和神经递质信号早在多细胞生物出现前就已在单细胞祖先中演化形成。 - 斑马鱼大脑处理视觉信号的电信号爆发，与青蛙胚胎早期决定体轴极性（前后、背腹）的荧光电压成像在信息交互模式上具有等价性。 - 膜电压模式（$V_{\text{mem}}$ pattern）不仅是生理副产物，更是存储宏观解剖学“目标形态”的动态媒介。 ### [19:16](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1156s) 生物电图式的可视化与调控工具 研究人员能够利用电压敏感荧光染料（voltage-sensitive fluorescent dyes）等多种技术，直观地观测到这些生物电图式。同时，通过大量的定量计算模拟，科学家们正在努力理解不同的离子通道和离子泵蛋白质是如何在多细胞组织中产生特定电信号图式的。这些图式主要分为两种：正常的（内源性）图式和病理性的图式。 以灰度显示的蛙胚发育为例，当它的面部开始聚合时，系统内就已经存在一个预图式（pre-pattern）或微型地图，预先描绘了面部基因表达和解剖结构的未来样貌。图式上清晰地标示出了眼睛和嘴巴将在何处形成。这些图式对正常发育至关重要，一旦移动或干扰它们，胚胎的解剖结构就会发生改变。相反，图式也可能呈现病理性。在诱导了肿瘤的蝌蚪身上，早在肿瘤肉眼可见之前，相关细胞就已经表现出异常的电状态。它们停止与邻近细胞沟通，仿佛退回到了单细胞生物的生存状态，并最终演变为癌细胞。 为了控制这种非神经细胞网络中的电信号，研究团队开发了一系列功能性工具。需要强调的是，这种方法绝非通过外部施加电场或使用电极来实现。 > “基本上，我们不施加电场，这不是我们的做法，这里没有使用任何电极。我们所做的是操纵允许细胞产生电状态的内源性离子通道，或者干扰细胞间的连接性。” —— 迈克尔·莱文（Michael Levin） 通过基因、药理学和光遗传学等多种技术，研究人员既能控制网络的拓扑结构，也能直接控制其电状态。 --- ### [20:51](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1251s) 跨越进化深渊的器官诱导与物种形态重塑 改变电回路可以带来令人惊叹的成果，其中之一就是能够说服细胞群分化为完全不同的器官。这并非指单一的干细胞分化，而是模块化的器官级控制。例如，在蝌蚪的肠道区域，通过开启一种特定的电压门控钾通道来改变其电学特性，一旦建立起特定的电状态，整个肠道区域就会发育出一只结构完整的眼睛，其中包含视网膜、视神经和晶状体等所有正常眼睛应有的内部组件。 这种模块化控制在涡虫（planaria）实验中同样得到了验证。如果切除涡虫的头部和尾部，并对其具有独特“头-尾”电学预图式的中间片段进行干预，就可以说服这些细胞在原本应该长出尾巴的地方长出一颗头部，从而培育出双头涡虫（甚至也可以培育出无头涡虫）。在这类实验中，研究人员并没有像传统生物工程那样去微观管理每一个细胞的布局，而是提供了一个相对少量的关键信息，告诉该区域应该成为什么，这就像是触发了一个“子程序”。 ```mermaid flowchart TD accTitle: 涡虫生物电图式记忆重写实验流程 accDescr: 展示如何通过改变电生理状态，使正常单头涡虫在受伤后永久性地再生为双头涡虫的实验步骤。 subgraph Phase1 ["第一阶段：电生理干预"] A["正常单头涡虫 (基因组未改变)"] --> B["改变电生理预图式 (写入双头记忆)"] end subgraph Phase2 ["第二阶段：首次再生"] B --> C["切除头部和尾部"] C --> D["再生出双头涡虫"] end subgraph Phase3 ["第三阶段：永久性验证"] D --> E["在普通水中静置数周"] E --> F["再次切除两个头部"] F --> G["依然再生为双头涡虫 (永久锁定新状态)"] G -->|可通过特定干预| H["重塑回单头涡虫"] end ``` 更令人惊叹的是跨物种的形态重塑。当切除一种具有三角形头部的涡虫（如多头杜氏涡虫 *Dugesia dorotocephala*）的头部后，通过干扰剩余片段的电网络，可以促使新细胞构建出属于其他涡虫物种的头部形状。例如，长出像 *Polycelis felina* 一样的扁平头部，或是像地中海涡虫（*Schmidtea mediterranea*）一样的圆形头部。 - **进化跨度**：这些物种在进化史上相隔约 1.5 亿年。研究人员成功让它们长出了属于其他物种的脑部结构和干细胞分布。 - **基因组未变**：在整个过程中，涡虫的基因组完全正常，没有进行任何基因编辑。 > “这是在与生理系统的实时‘软件’进行交互，而不是重新连接该动物体内现有的基因或蛋白质网络。” —— 迈克尔·莱文（Michael Levin） 此外，研究人员甚至能触及进化尚未涉足的“形态空间”（morphospace）荒原。通过改变电学属性，正常的扁平涡虫可以被重塑为带有尖刺的形状、带有内部空腔的三维“帽子”形状，甚至是背部长出管状物的异形。这表明在基因组之外，存在着一个极其易于操作且影响深远的生理“软件级”控制系统。 --- ### [24:06](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1446s) 生物电图式记忆的重写与永久性重塑 为了理解这些电回路以及动力学系统特性如何映射到具体的解剖学结果上，研究团队进行了大量的计算建模。如果“发育图式记忆”确实储存在生理电网络的电状态中，那么理论上这种记忆应该是可以被重写的。 在一项研究中，研究人员选取了外观和分子标记物均完全正常的单头涡虫（头部标记物仅存在于头部，尾部没有）。在正常情况下，切除其头尾后必然会再生出一条普通的单头涡虫。然而，如果在切除前改变其体内的电学预图式，那么即使这只涡虫在外观和分子层面上看起来依然是普通的单头涡虫，它在被切除头尾后却会再生出一条双头涡虫。 > “电信号图式并不是对当前解剖结构的读取，它是一段编码信息，一个发育图式印记（patterning Engram），在动物未来受伤时用于引导形态发生。” —— 迈克尔·莱文（Michael Levin） 这表明，单个涡虫身体可以同时存储至少两种不同的图式记忆（单头或双头）。更重要的是，这种改变具有长期的稳定性。如果将上述实验中得到的双头涡虫放在普通的清水中静置数周（不施加任何进一步的电生理干预），然后再次切除它的两个头部，按照传统的基因决定论，它应该会恢复成由其正常基因组决定的单头状态。但事实并非如此： - **永久性重塑**：一旦电网络被转移到“双头状态”，该图式记忆就会永久锁定。在不进行任何后续干预的情况下，无论对其进行多少次切除，它都会持续再生为双头涡虫。 - **可逆性**：研究人员目前也已掌握了通过特定手段将双头涡虫重新塑造回单头涡虫的方法。 这种生物电图式表现出了记忆的几个核心特征：长期稳定性（long-term stability）、可重写性（rewritability/lability）、延迟或条件性召回（latency/conditional recall，即平时不显现，仅在受伤需要再生时召回），以及离散的可能结果（单头或双头）。 目前，研究团队正致力于将联结主义模型（connectionist models）引入该领域，以帮助预测和引导实验。这些电回路状态空间中的稳定吸引子（stable attractors）对应着下游具体的解剖学结构。掌握这种生物电编码不仅对基础科学至关重要，在再生医学领域也展现出巨大的应用前景。例如，蝾螈可以自然再生肢体，而青蛙在正常情况下则无法再生其腿部，但通过生物电干预，这一局限正在被打破。 ### [28:17](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1697s) 生物电诱导肢体再生与逆转先天畸形 Michael Levin 博士展示了他们在再生医学领域的突破性进展。在正常情况下，当青蛙的腿被截肢后，它们“什么都不会长，或者只能长出一根非常薄的软骨刺”。然而，研究团队研制出了一种药物鸡尾酒配方，能够改变伤口处细胞的生物电状态，进而“说服”这些细胞去再生一条全新的肢体。实验图像清晰记录了这一过程：截肢后仅 45 天，肢端就开始长出脚趾甚至脚趾甲；随着时间推移，青蛙最终拥有了一条机能完备的新腿，不仅对外界触碰具有灵敏的触觉感知，而且行动自如，能够供动物正常活动与行走。随后，组织结构的模式形态完全长成，只需等待其体积继续长大。Levin 强调，这为复杂的再生医学开辟了全新路径：它完全不需要植入干细胞，也不需要进行任何基因组编辑（genomic editing），其本质是“纯粹通过触发这些动物在电学软件中早已内置好的形态发生级联反应和子程序”（patterning subroutines）。目前团队正将该技术从青蛙逐步推进到哺乳动物实验，目标是最终服务于人类医学。 随后，Levin 介绍了如何利用生物电信号的预测性模拟来修复各种严重的先天畸形（birth defects）。正常的蝌蚪脑部结构清晰，包含分化明确的前脑、中脑和后脑；但若暴露于致畸物（teratogens，如酒精、尼古丁等有害化学品），或发生关键基因（如 Notch 基因）的突变，脑部发育就会遭遇灾难性破坏——前脑完全缺失、中脑退化成气泡状、后脑陷入混乱。研究团队通过生物电仿真系统推算出需要恢复什么样的电学状态才能重建正常的大脑，并根据计算结果筛选出能够打开或关闭特定离子通道的候选药物。实验证明，即使动物携带严重致畸的 Notch 基因突变，经过靶向离子通道的药物调控后，其大脑结构、基因表达，乃至学习记忆功能全部得到了恢复。Levin 略带自豪地调侃道：“我们甚至给它们进行了智商测试，信不信由你，你确实可以在蝌蚪身上测定学习和记忆能力。” ```mermaid flowchart TD accTitle: 生物电干预逆转脑部畸形流程图 accDescr: 展示致畸影响通过生物电仿真预测再到靶向药物调控修复大脑的完整流程 subgraph Defect["病理状态"] A["基因突变 / 致畸物暴露<br>(Notch 突变、酒精、尼古丁)"] --> B["脑部严重畸形<br>(前脑缺失、中脑呈气泡状)"] end subgraph Intervention["生物电靶向修复"] B --> C["生物电动力学仿真<br>(推算目标生理电状态)"] C --> D["筛选靶向离子通道药物鸡尾酒<br>(无需基因编辑或干细胞)"] D --> E["重塑创口/组织跨膜电位"] end subgraph Outcome["表型与功能恢复"] E --> F["脑解剖形态重建<br>(前脑、中脑、后脑分化正常)"] E --> G["基因表达与认知功能恢复<br>(学习与记忆测试达到正常标准)"] end ``` - “This does not require implantation of stem cells, it does not require genomic editing, this is purely about triggering the kinds of cascades, the kinds of patterning subroutines that these animals already have embedded in their electrical software.” —— Michael Levin - “We really think that this bioelectrical network is an incredibly tractable layer that allows us to intervene in the default state that is encoded by the genome.” —— Michael Levin - 肢体再生在截肢 45 天左右出现可见的趾部和指甲结构，随后形成具备触觉与运动能力的完整下肢。 - 逆转脑缺陷的实验不仅覆盖环境致畸物（酒精、尼古丁），还针对严重的遗传性基因突变（如 Notch 突变）。 --- ### [30:48](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1848s) 非神经计算构架与新型机器学习的未来 Levin 展望了生物电科学的“终局目标”（the endgame）：构建一种“生物学编译器”（biological compiler）。如同当今工程师在计算机辅助设计（CAD）或 3D 打印中直接定义最终外形一样，未来的研究人员将能够直接在解剖学层面设计生物体的形态结构，而编译器能将其自动转化为引导组织生长成型的指令。进化在极其久远的早期就发现了利用电路进行信息处理的巨大优势，因此破解生物电代码不仅将彻底革新再生医学，更将深刻推动合成生物工程与形态计算（morphological computation）的发展。 以此为基础，Levin 向现场的人工智能学者提出了一个引人深思的核心设想：当前机器学习的诸多瓶颈，是否源于研究界过度聚焦于模仿大脑（尤其是人类大脑）的架构？他呼吁大家退后一步，关注那些甚至在生物演化出大脑之前就已经存在的基本信息处理方式。研究表明，骨骼、心脏乃至胰腺细胞均展现出学习能力，且极擅长对从未经历过的新颖环境做出极其稳健的自适应调整。这种对未知状况的极高适应弹性，正是强健机器人学（robust robotics）与人工智能所梦寐以求的特质。团队正在尝试将联结主义（connectionism）和神经网络的数学框架拓展至非神经生物网络，并结合机器学习开发“机器人科学家”（robot scientist）系统，用算法直接逆向工程出复杂生物体系的动态模型。 - “Is it possible that some of the limitations of current machine learning is because people have really been focusing on imitating brains, and especially human brain architectures, and what would happen if we stepped way back and based our attempts at this on cells and tissues that are doing much more fundamental information processing than what we see in brains?” —— Michael Levin - 资助机构与致谢：Paul G. Allen Frontiers Group 以及 DARPA（美国国防高级研究计划局）。 - 演讲结尾展示了具有稳定双头结构真涡虫的活体运动视频，并面向计算机科学与生物工程交叉领域公开招募博士后及访问学者，合作研发非神经 AI 平台及生物驱动的机器人躯体。 --- ### [34:35](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2075s) 现场问答：时间尺度分离与多层级并行计算 在问答环节，现场观众提出了两个关键的技术问题：生物体内既有维持日常运动的短暂电活动，又有负责组织再生等发育决定的电信号，研究人员如何区分二者？又如何对再生所需的电信号实施精准编辑？ Levin 解释道，两者的核心差异在于时间尺度（time scale）与频率特征。负责动物行为、肌肉收缩和神经动作电位的电信号发生极快，处于“毫秒级别”（milliseconds）；而指导解剖形态决策与组织模式生成的生物电模式极其稳定，其变化是以“分钟、小时甚至数天”为单位演进的。在实验干预手段上，实验室的基础研究通常使用可控的离子通道敲入技术（如光控或特定药物调控的通道蛋白）；但在面向患者的生物医学应用中，为了规避基因治疗的风险，团队采用计算生物学工具筛选针对患者组织内天然存在的离子通道/泵的小分子药物鸡尾酒，以此诱导目标生物电状态。 当听众进一步追问调节长期电信号是否会干扰生物体即时的短期行为与神经功能时，Levin 回顾了该领域早期的学术阻力：“大约在 2000 年我们刚提出这一观点时，它被认为是彻底疯狂的；人们断言细胞的电学参数纯属基础代谢特征（housekeeping parameters），一旦触碰，所有细胞都会死掉，引发无法解释的严重毒性。”然而事实证明这完全是错误的。细胞拥有极其精妙的能力去甄别不同时间尺度的电学信号，因此只要不使用过度极端的药物剂量破坏整个系统，长期发育调控与短期行为控制完全是并行运作的。在涡虫或蝾螈的再生过程中，动物完全可以一边到处游走、进食、开展正常日常行为，而底层的细胞网络则同时在另一尺度上从容推演解剖形态重建，两者在多尺度的层级系统中互不干扰。随后，下一位提问者走上麦克风，就合成生物学、昆虫神经电刺激控制以及真正的强人工智能路径，向 Levin 提出了更广义的探索方向探讨。 - “The kinds of electrical states that are important for decision-making at the anatomical level have a very different time scale than the kinds of things that guide behavior... anything that has to do with muscle motion, with neural activity, the time scale you're talking about is milliseconds... the electrical activity that we deal with, these patterns are stable and they change over a scale of minutes, hours, in some cases days.” —— Michael Levin - “In all of these organisms, the regenerating planarian or the regenerating salamander, this is running around, eating, behaving, doing all that stuff while all the cells are figuring out what to do. So there's a multi-scale set of problem solving that does not interfere with each other.” —— Michael Levin - 过去 15 年间，Levin 团队核心攻关方向之一正是开发将副作用降到最低的药物干预协议，确保生物电重编程与肌体即时生理活动相兼容。 ### [37:56](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2276s) 合成形态学与大脑的可塑性 Michael Levin 深入介绍了其实验室在“合成形态学”（Synthetic Morphology）方向的前沿探索。尽管许多惊人的数据尚未正式发表，但他透露，研究团队已经发现细胞具有在完全不同构型下学会协同工作的能力。通过涌现性（emergence）与引导性自组装（guided self-assembly）的结合，科学家能够制造出与供体动物截然不同的“人工活体机器”（artificial living machines）。另一个极具前景的方向是从细胞和组织中提取决策策略，并将其部署于硬件中或用于构建生物计算接口。 为了说明生物系统令人难以置信的适应性，Levin 分享了一个因时间限制未能在正片详述的实验：团队将眼睛移植到蝌蚪的尾部而非头部。结果表明，这些蝌蚪完全能够通过长在尾巴上的眼睛视物。尽管大脑在数百万年的演化中一直预期从头部特定区域接收视觉信号，但当尾部这个“奇怪的肉斑”突然开始传递神经信号时，大脑依然能迅速理解这些输入属于视觉数据并建立学习闭环。 > “The brain has no trouble figuring out that that's visual data and being able to learn so those plasticities is incredible and we need to we need to understand how this works and then I think I think the artifacts that we make are going to be just transformative,” Michael Levin 总结道。 - **人工活体机器**：利用细胞间新型的合作机制，实现与原有生物体形态脱钩的形态构建。 - **异位视觉功能恢复**：尾部植入眼睛的蝌蚪依然能处理视觉信号，展现了中枢神经系统惊人的计算可塑性。 --- ### [39:27](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2367s) 植物电信号与多物理场调控网络 在回答关于植物以及较慢的机械力信号（mechanical signaling）的提问时，Levin 指出，尽管自己的实验室不直接研究植物，但已有其他团队证实植物中同样存在高度相似的生物电机制。这种保守性极为惊人，因为植物与动物的多细胞特性是各自独立演化而来的。事实上，生物电通信的起源可一路追溯到细菌祖先——即便细菌生物膜（bacterial biofilms）也会传播类似脑电波的钾离子波。植物根系在土壤中导航探索与器官形态发生的过程中，都深度依赖生物电模式决策。 机械力信号与生物电信号之间存在紧密的双向耦合反馈：细胞膜上分布着对拉伸和压力敏感的机械门控离子通道（stretch- and pressure-sensitive ion channels），能将物理机械张力转化为电信号；反之，也存在能在跨膜电位梯度驱动下产生应变与应力的蛋白质。这两种物理场再与经典的遗传和生物化学信号通路交织在一起，构成了一个高度动态的复杂系统。 ```mermaid flowchart TD accTitle: 多模式生物信号交互网络 accDescr: 展示机械力、生物电以及基因生化信号之间的双向互动，以及在生物电层级进行高效干预 subgraph Signals ["多物理场动力学系统"] Mech["机械力信号<br>(压力 / 拉伸)"] BioE["生物电信号<br>(静息电位 / 离子流)"] BioChem["生化与基因调控网络<br>(生长因子 / 转录)"] end Mech <-->|"机械敏感通道 / 电致应变蛋白"| BioE BioE <-->|"信号转导 / 离子通道表达"| BioChem Mech <-->|"力化学转导"| BioChem Intervention["干预切入点"] -.->|"触发高阶模式子程序 (最高效)"| BioE Intervention -.->|"逐个细胞装配 (传统生化干预)"| BioChem ``` 面对这一复杂的控制架构，关键在于寻找最高效的干预切入点。Levin 强调，多数同行目前聚焦于生化与遗传层面的分子操作（如生长因子与扩散信号），试图逐个细胞地搭建组织；然而，从生物电层级切入提供了一种如同“调用完整子程序”的高阶控制机制。 > “Intervening at the electrical level gives extremely efficient control so if you want an eye you can try to assemble it cell by cell but it's really much more efficient to trigger the whole subroutine as it were,” Michael Levin 强调指出。 --- ### [41:55](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2515s) 意识的本质与“基础认知”的连续谱 有现场听众提问：从肢体再生细胞做决策、植物根系在土壤中寻找路径，到人类大脑做出复杂抉择，意识（consciousness）究竟在何处萌发？它是一条连续的渐变谱带，还是存在明确的分界线？ Levin 幽默地回应称，自己在细胞层面上谈论“决策”就已经引来了足够多的争议，因此不想在“意识”这一哲学与神经科学的火药桶上妄下断言。但他着重介绍了正在兴起的“基础认知”（Basal Cognition）研究领域。该学科探讨从细菌、植物、动物单细胞乃至蚁群等群落生物中所体现出的记忆、学习、感知和创造力。例如，蚁群作为一个整体所掌握的信息，往往是单个蚂蚁并不具备的。 > “I haven't anything profound to say specifically about consciousness but I do feel very strongly that all of this is a continuum that there is nothing radically special about it,” Michael Levin 肯定地指出。 Levin 严厉批评了将认知术语纯粹局限于人类心理学范畴的狭隘观念，认为这在演化生物学上是一个重大错误。在演化长河中不存在绝对的原则性区隔，只有量级上的渐变。例如，单细胞酵母菌（yeast）就能学会简单的环境模式并展现出“预判”（anticipation）行为。关于界线的划分，Levin 提及一位同事的观点：只有当生物具备“无限联想学习”（Unlimited Associative Learning, UAL）能力时，意识才真正诞生。但不管如何定义，功能性认知能力从生命之树的最底层便已存在。 - **基础认知（Basal Cognition）**：打破“大脑中心主义”，在没有传统神经系统的单细胞、真菌、植物与集群生命中研究智能行为。 - **酵母菌学习行为**：单细胞真核生物能够根据过往刺激模式做出预期反应。 - **意识连续谱理论**：不存在从无到有的突变界限，各种认知能力表现为跨物种的连续谱系。 --- ### [44:07](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2647s) 生物电调控方法与吸引子动力学 一位听众对生物电操作的具体实验手段表示好奇，询问是否需要将电极刺入组织或施加特定频率的外部交变电场，以及信号需要维持多久。 Levin 明确澄清，团队从不使用外加电场或插入式电极，也不依赖特定电频率。其实际操作策略是通过干预细胞内源性表达的离子通道蛋白，打开或关闭特定通道，从而调控细胞膜的跨膜静息电位（resting potential）。具体技术路线包括： 1. **药理学方法**：在再生医学应用中，通过筛选小分子药物直接调节离子通道； 2. **遗传学与光遗传学（Optogenetics）**：例如在蝌蚪模型中表达通道视紫红质（channelrhodopsin），利用空间光图案精确控制特定区域的电位空间分布。 在时间尺度方面，这些干预手段的作用时间短得令人惊讶，本质上只是触发形态发生的“子程序调用”。Levin 以青蛙断肢再生为例：对截肢伤口仅需进行 24 小时的药理学刺激，随后彻底撤除干预，组织便会在接下来的 9 个月内自主生长出完整的蛙腿。 > “The treatments are often incredibly brief because these one of the properties of these electrical circuits is they have these stable attractors that once you knock them into that sort of neighborhood of the state space you don't need to hold them there they're gonna maintain state,” Michael Levin 解释道。 这种现象背后的物理机制是生物电回路中具有稳定的“吸引子”（attractors）。只要将系统推入状态空间的特定吸引子邻域，回路就会自我维持该状态，无需外界持续维持刺激。团队的核心目标正是测绘出这些状态空间图谱，以最小的外部干预换取最大化的生物形态产出。 - **无外加电场策略**：摒弃传统电极刺激，依托小分子药物或光遗传学靶向内源性离子通道。 - **24 小时对 9 个月**：蛙腿断肢再生仅需 24 小时诱导，即启动长达 9 个月的自主形态发生程序。 - **状态空间吸引子**：生物电网络具备多稳态记忆能力，微小瞬时扰动即可完成系统状态的锁定与维持。 --- ### [46:14](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2774s) 面向未来的人工智能与非神经计算架构 针对提问者关于“非神经计算架构（non-neural architectures）能为计算与工程带来什么、是否能带来有感知或更高组合性的系统”的探讨，Levin 展开了思考。尽管目前无法判断这种架构是否能创造出“具有感受体验”（feelings）的系统，但其最具颠覆性的价值在于带来极高鲁棒性与自适应能力的系统。 生物系统最令人惊叹的核心特征，在于能够面对完全陌生的未曾演化出的全新环境，并自主做出灵活适应。现有人工智能与机器学习算法在泛化和应对极端分布外（OOD）扰动时往往表现脆弱，而非神经计算在形态发生过程中展现出的拓扑自适应与纠错能力，为突破这一瓶颈提供了全新的工程思路与理论基石。 ### [47:04](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2824s) 生物多尺度智能与人工系统设计的启示 在问答环节的开头，Michael Levin 深入探讨了生物系统与传统工程/AI系统之间的核心差异。他指出，当前工程系统普遍面临的一个致命瓶颈在于：在一个特定任务上表现极佳的系统，一旦环境或任务发生迁移，往往表现得极其糟糕。相比之下，生物结构具有极强的适应性，细胞和组织在面对极其陌生的情境时依然能做出高度弹性的响应。Levin 强调，生物系统最独特之处在于它在“每一个尺度上都具备能力”（competency at every level）：不仅单独的分子网络具有记忆和灵活性，细胞、组织以及器官也各自拥有独立的自主计算与调节能力。换言之，生物体完全是由“具备高度智能的组件”构建起来的。 这种自下而上的多层级智能为解决工程难题提供了全新视角。Levin 概括道，整个生命系统的本质就是在危机四伏、充满不确定性的环境中坚决追求目标：“These biological systems have very specific goals, the environment throws all kinds of crazy scenarios at them, and they are really good at navigating all of that, and I think we don't really know how to do that, so that's what we'll get.” 当观众追问如何着手构建这种具备多尺度适应能力的系统时，Levin 表示自己已经有了许多初步构想与方案，并热情邀请提问者在演讲结束后私下深入交流细节。 - **多层级能力（Competency at Every Level）**：从分子网络、细胞到器官，各级结构均具备记忆、计算和适应性，而不是由呆板无知觉的基础构件机械拼装而成。 - **目标导向（Goal-Directed Navigation）**：生物系统擅长在环境施加不可预测干扰的情况下，持续调整并朝向稳态或发育目标演进。 --- ### [48:20](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2900s) 生物电调控的伦理考量与再生医学路径 随后，一位听众对该技术的伦理问题提出了疑问。听众指出，在两栖类等小型模式动物上操作可能争议不大，但如果未来拓展到人体临床并引发系统性的全身效应，公众可能会产生强烈反对，因此询问 Levin 对防范此类伦理风险有何思考。Levin 明确回应称，在生命科学现有的各种高风险前沿领域中，生物电调控的伦理风险实际上处在非常低的位置。 Levin 将生物电方法与当下主流的生物技术进行了对比。他指出，诸如全基因组编辑、制造新型病毒的感染性疾病研究等现有技术，其潜在危险与不可控性“远远超出了我们在这里讨论的范畴”。相比之下，生物电模式控制是一种极其温和、微妙的干预手段。Levin 最核心的技术构想是：“I think we can develop a strategy for regenerative medicine that does not require manipulating the genome, it does not require any kind of viral delivery vectors or all this.” 这种方法既无需篡改人类基因组，也不需要依赖病毒载体递送，仅通过调控细胞间的生理电学通信即可重塑形态，因此从伦理和生物安全性来看，它不仅不激进，反而是一种回避基因安全隐患的更优替代方案。 --- ### [49:36](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2976s) 信号转导网络建模与反应-扩散系统 在提问的最后阶段，两位研究人员从机器学习与计算生物学角度探讨了底层机制。第一位听众提到，机器学习领域目前在模拟 T 细胞等免疫细胞的信号转导网络，而计算系统生物学也建立了从常微分方程（ODE）、偏微分方程（PDE）到布尔网络的整细胞模型，询问生物电离子通道网络与这些传统信号网络之间的内在联系。Levin 肯定了该问题的广度与深度，表示这个问题涉及庞大的机制细节，非常适合在会后进行长篇技术交流。 紧接着，另一位学者就经典的图灵“反应-扩散系统”（Reaction-Diffusion Models，如已被证实的 BMP 骨形态发生蛋白信号）向 Levin 发问：生物电系统究竟是位于反应-扩散系统的上游、干扰了它，还是说生物电网络本身就是一个反应-扩散系统？Levin 给予了双重肯定的解答：一方面，从现有实验数据来看，生物电处于上游，它负责指导和调控下游的化学形态发生素（chemical morphogens）；另一方面，生物电网络自身完全就是一个原生的反应-扩散系统。他特别指出，即使在底层没有发生任何基因转录变化的前提下，一组空间分布的细胞网络纯粹在生物电水平上就能够实现自发对称性破缺（spontaneous symmetry breaking），自发产生典型的类图灵图案（Turing-like patterns）。演讲在热烈的掌声与对 Levin 精彩报告的感谢声中圆满结束。 ```mermaid flowchart TD accTitle: 生物电与化学形态发生系统的双层关系 accDescr: 展示生物电网络如何既作为上游信号指导化学形态发生素，又自发构成反应-扩散系统产生对称性破缺。 subgraph BioelectricLevel["生物电层面 (Bioelectric Level)"] IonNet["跨细胞离子通道与间隙连接网络"] TuringPattern["纯电学自发对称性破缺\n(类图灵模式 / Turing-like patterns)"] IonNet -->|"纯生理电学动力学\n(无需底层基因转录改变)"| TuringPattern end subgraph ChemicalLevel["化学形态发生层面 (Chemical Morphogen Level)"] BMP["化学形态发生素梯度\n(如 BMP 信号系统)"] DownstreamPattern["组织与器官宏观图案化发生"] BMP --> DownstreamPattern end BioelectricLevel -->|"作为上游系统指导与调控"| BMP ``` - **纯电学对称性破缺**：空间分布的细胞电网络不需要转录重编程，即可独立完成对称性破缺并形成图灵斑图。 - **与经典生物化学的层级关系**：生物电不仅作为物理级反应-扩散系统存在，还在形态发生层级上作为上游开关，直接统领化学形态发生素（如 BMP）的释放与空间分布。 ## Notable quotes > [01:35](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=95s) 我想告诉大家的是，早在脑出现之前，生物就已经在多个尺度上进行计算了，实际上，所有身体组织中的决策都是由一个古老的、非神经的电网络介导的，这个网络由所有细胞组成，而不仅仅是神经元。 — Michael Levin > [12:39](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=759s) 在生物学中，我们正变得非常擅长操纵分子和细胞——那些控制复杂图案所必需的最低层级的硬件；而我们尚未真正理解的，是足以实现我刚才展示的图案编辑实例的控制算法。 — Michael Levin > [28:17](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1697s) This does not require implantation of stem cells, it does not require genomic editing, this is purely about triggering the kinds of cascades, the kinds of patterning subroutines that these animals already have embedded in their electrical software. — Michael Levin > [28:17](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1697s) We really think that this bioelectrical network is an incredibly tractable layer that allows us to intervene in the default state that is encoded by the genome. — Michael Levin > [39:27](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2367s) Intervening at the electrical level gives extremely efficient control so if you want an eye you can try to assemble it cell by cell but it's really much more efficient to trigger the whole subroutine as it were — Michael Levin > [44:07](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2647s) The treatments are often incredibly brief because these one of the properties of these electrical circuits is they have these stable attractors that once you knock them into that sort of neighborhood of the state space you don't need to hold them there they're gonna maintain state — Michael Levin > [48:20](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=2900s) I think we can develop a strategy for regenerative medicine that does not require manipulating the genome, it does not require any kind of viral delivery vectors or all this. — Michael Levin > [24:06](https://www.youtube.com/watch?v=RjD1aLm4Thg&t=1446s) 电信号图式并不是对当前解剖结构的读取，它是一段编码信息，一个发育图式印记（patterning Engram），在动物未来受伤时用于引导形态发生。 — Michael Levin ## Mentioned NeurIPS, Michael Levin, Kristin Grauman, Facebook AI Research, University of Texas at Austin, Tufts University, Allen Discovery Center, Harvard Medical School, Nature, American Heart Association, International Functional Electrical Stimulation Society, James V. McConnell, Paul G. Allen Frontiers Group, DARPA ## Video description Presented December 4th 2018 by Prof. Michael Levin (Allen Discovery Center at Tufts University) Michael Levin Vannevar Bush Professor Director, Allen Discovery Center at Tufts Director, Tufts Center for Regenerative and Developmental Biology Morphological and behavioral information processing in living systems [Source](https://www.youtube.com/watch?v=RjD1aLm4Thg)

[1:35] Michael Levin: 我想告诉大家的是，早在脑出现之前，生物就已经在多个尺度上进行计算了，实际上，所有身体组织中的决策都是由一个古老的、非神经的电网络介导的，这个网络由所有细胞组成，而不仅仅是神经元。 [12:39] Michael Levin: 在生物学中，我们正变得非常擅长操纵分子和细胞——那些控制复杂图案所必需的最低层级的硬件；而我们尚未真正理解的，是足以实现我刚才展示的图案编辑实例的控制算法。 [28:17] Michael Levin: This does not require implantation of stem cells, it does not require genomic editing, this is purely about triggering the kinds of cascades, the kinds of patterning subroutines that these animals already have embedded in their electrical software. [28:17] Michael Levin: We really think that this bioelectrical network is an incredibly tractable layer that allows us to intervene in the default state that is encoded by the genome. [39:27] Michael Levin: Intervening at the electrical level gives extremely efficient control so if you want an eye you can try to assemble it cell by cell but it's really much more efficient to trigger the whole subroutine as it were [44:07] Michael Levin: The treatments are often incredibly brief because these one of the properties of these electrical circuits is they have these stable attractors that once you knock them into that sort of neighborhood of the state space you don't need to hold them there they're gonna maintain state [48:20] Michael Levin: I think we can develop a strategy for regenerative medicine that does not require manipulating the genome, it does not require any kind of viral delivery vectors or all this. [24:06] Michael Levin: 电信号图式并不是对当前解剖结构的读取，它是一段编码信息，一个发育图式印记（patterning Engram），在动物未来受伤时用于引导形态发生。

[7:18] Michael Levin: The most amazing thing about regeneration is not that it's able to start it's that it knows when to stop.

[0:00] good afternoon everyone I'm Kristin [0:04] Grauman I'm a researcher at Facebook AI [0:06] research and a professor at UT Austin [0:08] and is one of the program chairs for [0:10] nerps this year I have the pleasure to [0:11] introduce our next invited speaker [0:13] professor Michael Levin Michael is a [0:16] professor biology at Tufts University [0:18] and he holds a Vannevar Bush endowed [0:20] chair he's also a director of the Allen [0:23] Discovery Center at Tufts which examines [0:26] computation in the medium of living [0:28] systems his academic background is both [0:31] computer science and biology he received [0:34] his PhD from Harvard Medical School in [0:36] 1996 Michaels group does work at the [0:39] intersection of developmental biology [0:41] cognition and computational neuroscience [0:44] they are especially interested in [0:46] applications in regenerative medicine [0:48] and as we'll see in his talk today these [0:51] things can offer inspiration for new [0:52] machine learning and robotics approaches [0:55] professor Levin's research has been [0:57] recognized with a number of different [0:59] awards I'll mention just a few [1:00] one is recognition as having made a [1:04] milestone in developmental biology in [1:06] the last century from the journal Nature [1:08] and that was for his work on molecular [1:10] basis of left-right asymmetry he's also [1:13] received the established investigator [1:15] award from the American Heart [1:16] Association and the scientist of vision [1:19] award from the International functional [1:21] electrical stimulation society so we're [1:24] very excited to hear about his work [1:25] today so please join me in welcoming [1:27] professor Michael Levin [1:29] [Applause] [1:35] thank you so much and I'm really excited [1:38] to be able to present this work to a new [1:40] audience and I'm gonna give you the main [1:43] message of my talk right up front what I [1:45] would like to tell you about is that [1:47] biology has been computing at many [1:49] scales long before brains came on the [1:51] scene and then actually decision-making [1:53] and all body tissues are mediated by an [1:56] ancient pre neural electrical network [1:58] that is made up of all cells not just [2:00] neurons and so I think that exploiting [2:02] this kind of non neural cognition is an [2:05] exciting untapped frontier for the [2:06] development of new robust AI [2:08] technologies and I'm very interested in [2:10] talking to people about these [2:11] possibilities we're looking for experts [2:13] in machine learning to collaborate with [2:14] us to take dielectrics beyond [2:16] regenerative medicine and towards many [2:18] new exciting applications so the first [2:20] thing I would like to do is talk to you [2:22] a little bit about brain body plasticity [2:24] - and then to bring to you some examples [2:28] of somatic cognition in the body [2:30] decision-making during the self editing [2:33] of complex anatomies and then we'll talk [2:35] about by electric mechanisms that [2:37] underlie the ability of cells and [2:39] tissues to remodel towards large-scale [2:41] goals and then I'll talk about that what [2:43] I think is the future of regenerative [2:45] medicine and Nai and related issues so [2:49] let's talk about plasticity caterpillar [2:52] start out life as a soft body robot that [2:54] crawls and choose plants during [2:57] metamorphosis its nervous system and [2:59] including its brain is basically [3:00] liquefied it's taken apart and then [3:02] reassembled into a completely different [3:04] brain that is able to run a hard robot [3:07] which flies and drinks nectar during [3:10] this process of complete nervous system [3:12] take up being taken apart and putting [3:14] back together the memories in fact [3:16] remain so things that the butter that [3:18] the caterpillar has learned the [3:20] butterfly will still remember and this [3:22] has been this has been as well [3:23] established and so it's very interesting [3:25] to be to start to think about how it is [3:27] that information can be stored in the [3:29] medium that is undergoing radical [3:31] deformation in work that we've done a [3:34] few years ago we showed that planaria [3:36] these are the flatworms that you see [3:38] here they have this amazing property [3:39] that if you cut off their head or in [3:41] fact any part of their body they will [3:43] regenerate it they regrow their brain [3:45] and all the other organs so if you train [3:47] a flatworm to wreck [3:48] a particular location as being its home [3:50] environment and then you amputate the [3:52] brain you leave the tail and let it [3:54] regenerate a new head here what you [3:56] discover is that again these these [3:58] memories are intact so the memory of [4:00] this in this animal is in some way [4:03] imprinted on the new brain that develops [4:05] after regeneration and it challenges us [4:07] to try to understand how encoding and [4:10] decoding of complex complex memories and [4:14] learned behaviors for is able to be [4:16] propagated across this kind of radical [4:18] reconstruction of the of the memory [4:20] medium and in fact when we publish this [4:22] a few years ago actually we weren't the [4:24] ones to discover this this was first [4:26] discovered by a guy named McConnell in [4:28] the 60s but we people hadn't really [4:31] looked at this for years and people got [4:32] pretty excited as you can see here on [4:35] the internet when we've when we first [4:36] showed that this that this was possible [4:38] and going beyond the sort of brain body [4:41] interface and the question of how the [4:44] programs of the of though the behavioral [4:47] programs of the brain interact with the [4:49] structures of the body I want to talk to [4:51] you about this kind of self editing of [4:53] anatomy there are many different types [4:56] of cognition from from very simple ones [4:59] to all the way through the kinds of [5:01] complex things that we hope to mimic [5:03] with machine learning and artificial [5:05] intelligence and there are all sorts of [5:07] living creatures that are able to [5:10] perform functions that are at various [5:13] places on this ladder of cognition so [5:16] one example for is here shown by this [5:18] unicellular creature called the lac [5:20] romário if you're interested in soft [5:22] body robotics take a look at this this [5:23] is a single cell okay and this organism [5:27] handles its physiology its patterning [5:29] look at the incredible control it has [5:31] this is real-time by the way this is not [5:33] sped up so look at the amazing control [5:35] that this has over its structure as it [5:38] collects bacteria and as it's basically [5:40] hunting and eating at the different [5:42] other microorganisms in this environment [5:44] but behavior structural control and [5:47] physiology all in one cell there is no [5:49] brain here as such and so unicellular [5:51] organisms and single cells can be [5:54] incredibly competent at all sorts of [5:56] interesting goals now when cells joined [6:00] to become [6:01] multicellular bodies complex anatomical [6:04] structures they did not lose this [6:06] ability to make decisions and to have [6:08] complex behaviors what you're seeing [6:09] here on the left is a frog embryo [6:11] developing and so these cells cooperate [6:13] to build a very complex structure on the [6:16] right you're seeing the development of a [6:18] nervous system in zebrafish and so all [6:20] of these cells are now working together [6:22] to build a structure that's much larger [6:25] than the single cell boundary of their [6:27] immediate surface and we need to [6:30] understand how this works [6:31] all of us started life as a single cell [6:33] a single fertilized egg and through this [6:36] process of embryonic development these [6:39] complex forms are able to be reliably [6:42] generated so all of these remarkable [6:44] complex anatomies and it's very [6:47] important that we understand that this [6:48] is not just a process that's going to be [6:50] solved by stem cell biology because stem [6:52] cells are able to generate all of the [6:55] individual tissues that are needed [6:57] foreign organisms of what you see here [6:58] is something called a teratoma this is a [7:00] tumor that has hair and teeth and skin [7:03] and muscle and so all of this stem cell [7:06] differentiation has gone okay but what's [7:08] missing is the large scale 3-dimensional [7:10] structure so one of the important things [7:12] we need to understand is how do cells [7:14] decide what they're going to build and [7:16] how do they implement that construction [7:18] and so the sort of standard story that [7:21] people are taught in developmental [7:23] biology is that there's this kind of [7:25] open-loop feed-forward system where gene [7:27] regulatory networks give rise to [7:29] proteins and then through some physics [7:31] when these proteins interact with each [7:32] other in the cells perform various [7:35] functions there's this emergent process [7:37] that give rise to complex anatomy but [7:40] it's actually much more interesting than [7:41] that first of all all kinds of embryonic [7:44] species are highly have highly regulated [7:49] development in the sense that if there [7:51] is a massive perturbation they can make [7:53] up for it so for example here is an [7:54] early mammalian embryo you can cut it in [7:57] half and each half will rebuild the [7:59] other missing half this is how you get [8:00] perfectly normal monozygotic twins the [8:03] opposite is possible too you can take [8:05] two or mammalian embryos mush them [8:07] together like a snowball here these [8:08] mouse embryos and what you get is a [8:10] perfectly normal Mouse we would love to [8:12] have technology obviously that has this [8:14] degree of plasticity [8:15] that can make up for such radical [8:17] deformations now this is not only an [8:19] embryonic phenomenon the process of [8:22] regeneration such as the salamanders [8:24] have is able to rebuild complex large [8:27] structures during adulthood [8:28] so here's an axolotl a salamander that [8:30] can regenerate its eyes its jaws its [8:33] limbs its spinal cord its ovaries and [8:36] portions of the brain so as an adult [8:38] this thing can rebuild very complex [8:40] structures throughout its life the most [8:42] amazing thing about regeneration is not [8:44] that it's able to start it's that it [8:46] knows when to stop all of these [8:48] structures after the correct anatomical [8:50] component has been rebuilt than further [8:52] growth and remodeling ceases and this is [8:55] very poorly understood how it knows when [8:57] to stop I want to introduce you to [9:00] another animal that you'll see more of [9:01] later in the talk this is a planarian [9:03] that same flatworm i mentioned a minute [9:04] ago [9:05] these can be cut into pieces the record [9:08] is 275 and each piece they're they're [9:11] about a 2 centimeters long by the way [9:12] and each of these pieces will regrow [9:14] exactly what's needed no more no less [9:17] to make a perfect worm every piece of [9:19] this animal knows exactly what the [9:21] correct worm looks like and in fact the [9:22] regenerative properties are so [9:24] remarkable that they've conquered aging [9:26] these worms are literally immortal the [9:28] individual cells will senescent die but [9:30] the animal as a whole continues to [9:32] regenerate them and there is no such [9:33] thing as an old planarian so now [9:36] regeneration is not just for so-called [9:38] lower animals the human liver is highly [9:41] regenerative deer will regrow their [9:43] antlers or bone highly innervated bone [9:46] up to a rate of a centimeter and a half [9:48] per day of new bone growth every year [9:50] and you may or may not know that human [9:52] children below the age of approximately [9:54] 7 to 11 are able to regenerate their [9:57] fingertips and this was discovered in [9:58] the 70s but then this ability ceases [10:02] with age so most of you shouldn't [10:04] shouldn't try this so what we what we [10:08] see here is that actually the control [10:10] structure of this process is not an [10:12] open-loop system but actually is a [10:14] closed loop that has some very important [10:16] error detection and control functions [10:19] here where deviations from this target [10:22] morphology lead to changes both at the [10:25] transcriptional level but also at the in [10:27] the physicals in the physical [10:29] processes that allow sales to rebuild [10:31] and so there's this constant homeostatic [10:33] loop and one of the goals of my Center [10:35] is to target the setpoint the anatomical [10:38] set point of this homeostatic process to [10:40] be able to recess off' I what the cells [10:42] should build this is what we're working [10:44] towards we discovered several years ago [10:46] that really the capabilities of [10:50] biological things to build towards a [10:52] spec are really amazing so in in the [10:54] journey of from from a tadpole to a frog [10:56] that all kinds of remodeling has to [10:58] happen so the eyes have to move together [11:00] and forward the nostrils have to move [11:02] the whole face remodels and when we made [11:04] these sort of Picasso like tadpoles [11:06] where everything was in the wrong place [11:07] the I was in the back and the jaws were [11:09] off to the side the starting positions [11:11] were all off they largely still [11:13] remodeled into a perfectly normal frog [11:15] all of the organs moved in unnatural [11:19] configurations until they reached this [11:21] normal frog morphology even though their [11:23] starting positions were not the same and [11:25] so we know that what the genome encodes [11:28] is not a rote set of movements that [11:30] every organ can simply undergo and give [11:33] rise to whatever it give rise to but [11:34] they can actually tell when they're in [11:36] the wrong configuration and it will keep [11:37] moving until the correct pattern is [11:39] constructed and tanning this process of [11:43] taking advantage of it is extremely [11:44] important over over 50 years ago people [11:47] realize that if you take a salamander [11:50] tail and graft it onto the edge of a [11:53] limb amputation so you put a tail where [11:56] the limb ought to be over the coming [11:57] months this this structure is actually [12:00] going to be remodeled from a tail into a [12:02] limb as this whole system seeks to [12:04] restore its large-scale global order so [12:06] understand what this means this is not [12:08] an individual cell deciding it's going [12:10] to be muscle or skin or anything else [12:12] this is large-scale decision making at [12:14] the level of the entire organism and so [12:18] fundamentally this is a computational [12:20] problem and we would like to use this as [12:24] inspiration for technology that can [12:26] implement this kind of highly robust [12:28] plastic behavior in the face of radical [12:31] deformation very new kinds of [12:33] environmental perturbations these [12:35] systems are able to regain their [12:37] structure and function now what [12:39] determines patterning we start life as [12:42] as an egg and than a collection of [12:43] embryonic stem cells and eventually this [12:46] is a cross-section taken through a [12:47] normal human torso you can see the [12:49] amazing invariant locations of all the [12:52] organs and the and the geometry of all [12:54] the structures where does this come from [12:56] DNA after all the genome specifies [12:58] proteins it does not say directly [13:00] anything about this this shape or how [13:02] it's edited how do cell groups know what [13:05] to make and when to stop and then as [13:07] engineers of course we'd like to know [13:08] how far can we push this process what's [13:11] possible to build what are the edges of [13:12] biological possibility and so you might [13:16] be surprised to know that as [13:17] developmental biologists we basically [13:19] don't know the answer to any of these [13:21] questions we have this significant [13:23] knowledge gap that while we're very good [13:25] at constructing these kinds of networks [13:28] we're in the we're we find out which [13:30] gene products interact with which other [13:31] gene products we can't yet read a genome [13:34] and find out what shape the resulting [13:36] organ M is going to be we can compare it [13:38] to other genomes where we already know [13:40] what it looks like but we can't read a [13:41] genome and figure out what the what [13:42] geometry had specifies what's what's [13:45] even worse in terms of medicine is that [13:47] if we wanted to fix an anatomical [13:49] structure here what we have our [13:51] molecular information of this kind we [13:54] have these networks that make it very [13:55] difficult to solve the inverse problem [13:57] that that will allow us to know what to [14:00] change at this level to make to have [14:03] outcomes desirable outcomes at the [14:05] system level and so we have here huge [14:07] opportunities for machine learning to [14:09] impact regenerative medicine beyond drug [14:11] discovery to help us be able to take [14:13] this kind of information and understand [14:15] how it maps to specific anatomy in [14:18] biology we're becoming very good at [14:20] manipulating molecules and cells the [14:22] very the very lowest level hardware that [14:25] is that is necessary for the control of [14:28] complex pattern what we don't really [14:30] understand yet are the algorithms that [14:31] are sufficient for the kind of pattern [14:33] editing examples that I just showed you [14:35] and so can we move biology beyond this [14:38] kind of machine code and I visualize it [14:40] like this where computer sciences have [14:42] this wonderful journey from from having [14:44] to alter the hardware when you were [14:45] programming all the way up to now [14:47] considerable platform independence where [14:49] you can really just focus on the control [14:51] algorithms and the computations biology [14:54] is still here everything we do [14:56] in our experiments is is rewiring the [14:58] actual hardware of the organism we have [15:00] not yet gotten to the point where we can [15:02] we can focus on the control algorithms [15:04] and really take advantage of the [15:06] hardware software distinction which I do [15:08] believe also exists in biology and this [15:10] is what my group has been trying to do [15:11] so pattern homeostasis has a number of [15:15] properties that are in a certain sense [15:17] cognitive they show goal directed [15:19] behavior towards specific outcomes they [15:21] show flexibility and robustness under [15:24] variable conditions and they allow [15:25] integration of self and low-level [15:27] functions into complex large-scale [15:30] outcomes and so what I want to talk to [15:34] you about now [15:34] are some of the progress that we've made [15:36] about understanding how electrical [15:38] mechanisms that is electrical [15:39] communication among cells allow this [15:42] kind of control and basically what I'm [15:46] going to show you everything in the next [15:48] section can be boiled down to one main [15:50] message that very much like the brain [15:52] somatic tissues form electrical networks [15:55] that make decisions the decisions they [15:57] make are not about behavior they're [15:58] about anatomy and we now can target the [16:00] system for some control of large-scale [16:02] pattern editing and so I show you this a [16:05] five legged frog here to point out that [16:07] this is not Photoshop I'm gonna show you [16:09] a bunch of really strange-looking [16:10] animals in a minute and this is these [16:11] are all real live creatures that are [16:13] that live in our lab because the way [16:16] that we test our various models of [16:19] whether we've understood how this works [16:21] is to try to make rational perturbations [16:23] we try to control anatomical [16:25] decision-making so it's important to [16:27] realize that brains did not invent their [16:29] tricks from scratch basically neurons [16:31] evolved from very ancient cells that [16:35] were around and performing computations [16:36] and doing somatic decision-making long [16:40] before neurons and brains showed up and [16:42] in fact our unicellular ancestors before [16:45] even the origin of multicellularity [16:46] already had synaptic machinery ion [16:50] channels neurotransmitters this by the [16:52] way if you've ever wondered why it is [16:53] that fungi which are very very distant [16:56] organism from ours from ourselves why [16:59] fungi have compounds in them like these [17:02] hallucinogenic mushrooms that are able [17:03] to affect human cognition it's because [17:05] the basic mechanisms are extremely [17:07] ancient there [17:08] very old and that basically in becoming [17:11] brains nervous systems speed optimized [17:15] functions that were already here long [17:17] before that and so we all know that in [17:20] the brain you have hardware that's [17:22] produced by these these ion channels [17:25] that enable cells in the brain to [17:28] communicate electrically with each other [17:30] so they mediated mediate electrical [17:32] state change or they can be propagated [17:33] through a network and this kind of [17:36] hardware gives rise to some very [17:37] interesting software that we would all [17:39] like to understand and take advantage of [17:41] so the cognitive and behavioral [17:43] processing that goes on in the [17:45] electrical network that is encoded and [17:47] enabled by this hardware well it turns [17:49] out that the rest of the body is not [17:50] very much different all cells not just [17:53] neurons have the exact same types of ion [17:55] channels they make electrical synapses [17:57] to their neighbors and they produce [17:59] electrical paths the electrical patterns [18:02] such as here what you're seeing here is [18:04] an early frog embryo this is this is [18:06] sped up this is about eight hours of an [18:08] early frog embryo and much like here you [18:11] can see the electrical activity of the [18:12] zebrafish brain as the zebrafish [18:14] processes visual information what you [18:17] can see here with pseudo color are all [18:19] the electrical conversations that these [18:21] cells are having with each other as they [18:23] figure out who's going to be front who's [18:25] going to be back left-right dorsal [18:26] ventral and so on so as people try to [18:30] understand how this to decode how the [18:34] electrical states of the brain stand for [18:37] different of memories behaviors and so [18:40] on we have to understand and we have to [18:42] crack this by electrical code how it is [18:44] that the changing electrical properties [18:47] of the tissues implement different [18:49] future anatomical States and so here's [18:52] an example this is just a snapshot of an [18:54] early frog tadpole like this this is the [18:56] mid flank the different colors are a [18:58] fluorescent voltage reporter that shows [19:01] us in living tissue what all the [19:02] electrical states are in each one of [19:04] these little circles is a single cell [19:06] and this cell has a variety of ion [19:07] channels just like neurons do that [19:09] enable them to maintain this particular [19:11] structure so this is what we call a V [19:13] mem or a membrane voltage pattern and [19:16] these patterns we're able to see them [19:18] using a variety of technologies such as [19:20] voltage sensitive fluorescent dyes [19:22] and we do a lot of quantitative [19:23] simulations to try and understand how [19:26] the different ion channel and pump [19:28] proteins give rise to specific patterns [19:30] in multicellular tissues these patterns [19:34] come in two flavors there are normal or [19:36] endogenous patterns so this is again a [19:37] frog embryo this time in greyscale and [19:39] you can see here that as it's putting [19:41] its face together here's basically a pre [19:43] pattern or a little map of what the gene [19:45] expression and the anatomy of the face [19:47] is going to look like so here's where [19:48] the eye is gonna form here's where the [19:50] mouth is going to be and so on so these [19:51] are patterns that are critical for [19:53] normal development if you move or [19:54] perturb these patterns the the anatomy [19:57] will be will be altered these patterns [19:59] can also be pathological here you see [20:01] the earliest signs of cancer we induced [20:03] a tumor in this temple but before the [20:06] tumor shows up these cells right here [20:07] you can already see their abnormal [20:09] electrical state as they stop [20:12] communicating with their neighbors and [20:13] sort of revert back to an almost [20:15] unicellular like existence and and [20:17] become become cancer cells so [20:19] functionally we've developed some tools [20:21] to control this type of signaling in [20:24] this non neural cell network basically [20:26] we do not apply electrical fields that [20:28] is that's not how this is done there's [20:30] no no electrical electrodes being used [20:32] here what we do is we manipulate the [20:35] endogenous ion channels that allow cells [20:37] to produce electrical States or we [20:39] perturb the connectivity so we can [20:41] either control the topology of the [20:43] network or actually the electrical state [20:45] and we use a variety of genetic [20:46] pharmacological optogenetic and those [20:49] kinds of techniques so I want to show [20:51] you some examples of why this is [20:53] important what you can do when you when [20:55] you alter the electrical circuits so one [20:57] thing you can do is you can convince [21:00] cell groups to become different organs [21:03] and I don't mean to control individual [21:04] cell types this is not about stem cell [21:06] differentiation but for example we can [21:08] take a region of this gut so this whole [21:10] thing is the gut of the temple we can [21:11] take this region we can change its [21:13] electrical properties by by opening a [21:16] particular kind of channel this [21:17] voltage-gated potassium channel and what [21:19] will happen is this whole once we [21:21] establish a particular electrical state [21:23] this whole region becomes an eye and if [21:25] you section this is a m-- interior [21:27] interior components that you would find [21:29] in a normal eye so retina optic nerve [21:31] lens all of these things are in there so [21:33] this is a modular control that [21:35] right here you're going to form an eye [21:37] the same thing can happen in planaria [21:39] where if we cut off the head in the tail [21:41] and we take this middle fragment which [21:43] has this very interesting head tail [21:44] electrical pre pattern what we can do is [21:47] take these cells and convince them that [21:48] they should make a head instead of a [21:50] tail and you see that here this is a [21:51] two-headed worm in fact you could go in [21:53] the other direction you can make a note [21:54] headed room and in all cases this is [21:57] very important this isn't this is [21:58] modular because what we're not doing is [22:01] going in and micromanaging the process [22:02] we are not trying to layout cell by cell [22:04] as you might if you are doing [22:06] bioengineering we're not trying to [22:07] layout cell by cell what the structure [22:09] is we're providing actually a fairly [22:11] small amount of information that tells [22:13] this region what it should be it's it's [22:15] basically a subroutine trigger at that [22:17] point and so you could use that using [22:19] this process some of the other things [22:21] you can do we can take this flat worm [22:22] that has a triangular head and when we [22:24] cut off the head we can treat the [22:26] remaining fragment in a way that [22:27] convinces the new cells to build heads [22:30] that belong to other species of flat [22:31] worms so here's a nice flat head like [22:33] this pea Felina or you can make a round [22:35] head like an S Mediterranean or a [22:37] triangular head like the original should [22:39] be notice that first of all it's about [22:41] 150 million years of evolution that [22:44] separate these animals so you're making [22:46] heads with brain structures and stem [22:48] cell distributions that belong to these [22:51] other species across about 150 million [22:54] years of evolution also note that the [22:57] genome is here is normal we have not [22:59] done any genomic editing this is not [23:00] about changing the genome this is this [23:02] is purely a physiological change this is [23:05] interacting with the real-time software [23:08] as it were of the physiological system [23:10] not rewiring the genes or the or the the [23:14] protein networks that are existing in [23:15] this animal so all of these animals with [23:17] heads belonging to other species have a [23:19] perfectly normal wild-type genomic [23:21] sequence that belongs to Dez adhirata [23:23] safale so you can go further than this [23:26] and you can reach regions of the [23:28] planarian morphus space that evolution [23:30] isn't using so from a normal flat worm [23:32] you can make things like this that are [23:34] sort of spiky you can make these that [23:36] look like hats they're three-dimensional [23:37] they have an interior compartment or you [23:39] can make ones that look like this with [23:41] tubes coming out of their backs so we're [23:43] just now beginning to really understand [23:45] the way that different anatomical [23:47] structures are encoded in the electrical [23:49] properties of an F of the network but [23:51] it's very clear that there's a very [23:52] tractable a profoundly important level [23:55] of software control that can be [23:57] implemented beyond the genome all of [23:59] these animals have the same genomic [24:01] sequence and so this is really really [24:03] important to begin to learn how to [24:04] target this for our various applications [24:06] we've done lots of computational [24:08] modeling to try to understand these [24:10] electrical circuits and how the [24:12] dynamical systems properties of these [24:14] circuits map onto specific anatomical [24:18] outcomes and so one of the predictions [24:21] of this kind of unusual view that these [24:24] are all that these pattern memories are [24:26] states of physiological electrical [24:30] network is that we ought to be able to [24:31] rewrite this pattern memory and I want [24:33] to show you one example that we just [24:35] published last year which is that if we [24:37] take these normal-looking planaria so [24:40] one head one tail and in fact [24:42] molecularly if we look at the markers so [24:44] the head markers are in the head there [24:45] are no head markers in the tail and so [24:47] on these will give rise to a one-headed [24:50] worm if we chop off the head in the tail [24:52] these will give rise to a two-headed [24:54] worm why would an animal that has a [24:55] perfectly normal one headed anatomy and [24:57] apparently normal molecular molecular [25:01] expression pattern why would they give [25:03] rise to two-headed worms and the reason [25:06] and again this is not anything to do [25:07] with genomic editing and the reason is [25:09] because what we did was we went in and [25:11] we altered the electrical pre pattern [25:13] that dictates what this at what this [25:15] tissue is going to do if it is injured [25:18] in the future so this one had an animal [25:20] has an is bearing an electrical memory [25:23] that says that if I get injured in the [25:25] future what I'm going to do is build two [25:26] heads instead of one so the electrical [25:28] pattern is not a readout of what the [25:30] anatomy is now it is a piece of encoded [25:33] information a patterning Engram that is [25:36] used if if the animal is injured and it [25:39] is used to cut to to to guide [25:41] morphogenesis in the future so what we [25:43] see is that this body a single body like [25:46] this [25:46] can store at least one of two different [25:48] pattern memories it can store one headed [25:50] or a two-headed pattern memory no doubt [25:52] probably more we've only found been able [25:54] to access to so far now this this is [25:58] this is really a very important example [26:02] because it is not just [26:03] about changing the way that regeneration [26:06] progresses after the injury this is a [26:08] long-term phenomenon and by that what I [26:10] mean is that if we take this two-headed [26:12] worm and so so so we start with a [26:14] one-headed worm we amputate the head in [26:16] the tail we perturb the electrical [26:18] network you get a two-headed worm and [26:19] now we take this two-headed worm and we [26:21] leave him alone for a couple of weeks in [26:23] plain old water so normal or [26:25] manipulation of any kind we amputate the [26:27] original head we amputate this weird [26:30] ectopic posterior head and we ask what's [26:32] going to happen and when we first did [26:34] this experiment of course the standard [26:35] paradigm said that well the genomics of [26:38] the worm are normal and you're going to [26:39] cut off the CICC topic head so of course [26:41] the rest of it is just going to go back [26:42] to normal and you're gonna get this [26:43] normal one headed worm right and of [26:46] course I'm showing it to you because [26:46] that's not at all what happens what [26:48] happens is that in perpetuity once [26:50] you've shifted that electrical Network [26:52] that pattern memory has now been shifted [26:54] to a 2-headed state it is now [26:55] permanently two-headed so if you keep [26:58] cutting this animal again and again and [27:00] with no further manipulations it will [27:02] continue to regenerate to that new [27:05] specification of what a worm [27:06] should look like and and you can by the [27:08] way we have now ways to I take it back [27:10] and take a two-headed worm and remodel [27:12] it back to one head so this is very [27:14] interesting this has this has a number [27:16] of key properties of memory it's its [27:18] long-term stable its rewritable or [27:21] labile it has latency or conditional [27:23] recall as I just showed you we can do it [27:25] so that nothing happens until the animal [27:27] needs to recall it by from injury and it [27:29] has discrete possible outcomes one [27:31] headed or two headed so what we've been [27:34] doing is trying to develop techniques to [27:37] basically extend certain connectionist [27:39] models to this to this to help us guide [27:43] these experiments where stable [27:45] attractors in the state space of these [27:48] electrical circuits correspond to [27:49] specific downstream downstream anatomies [27:53] of course mediated by all the different [27:55] genes that they turn on and off so this [27:57] is really critical for us as now is to [27:59] understand the large-scale properties of [28:01] this bioelectric code and how these [28:03] pattern memories are literally encoded [28:05] in the stable states of tissues it's [28:08] important not only for the basic science [28:10] of it but for applications in [28:11] regenerative medicine [28:12] so unlike salamanders frogs do not [28:15] normally regenerate their legs so the [28:17] they grow nothing or a very thin card [28:19] latch in a spike when you amputate the [28:20] leg and what we've been able to do is to [28:22] come up with a cocktail that alters the [28:24] bioelectric state of the cells at the [28:26] wound to convince them to regenerate a [28:28] new leg so here you can see in 45 days [28:31] it's starting to make some toes here's a [28:32] toenail and then eventually it's a [28:34] pretty good leg that can feel touch and [28:36] it's both sensitive and motile so the [28:39] animal can use these legs and then [28:40] eventually you get some you get some [28:41] really nice you get some really nice [28:43] patterning and then you have to wait for [28:44] it to grow so the idea is that this is [28:46] the first steps towards a program for [28:50] regenerative medicine where we will be [28:51] able to induce cells and tissues to [28:54] regrow various complex organs this does [28:58] not require implantation of stem cells [28:59] it does not require genomic editing this [29:01] is purely about triggering the kinds of [29:04] cascades the kinds of patterning [29:07] subroutines that these animals already [29:08] have embedded in their electrical [29:10] software and of course next after this [29:12] frog is mammalian application so we're [29:14] trying to trying to work up to human [29:15] medicine at some point okay so one of [29:18] the other things we've been able to show [29:19] is to use some of our priyad predictive [29:22] with simulations of the by electrical [29:25] signaling to reverse a birth defects of [29:28] different types so here's what a tadpole [29:30] brain a normal tadpole brain should look [29:32] like you see here the the forebrain the [29:34] midbrain and hindbrain and under the [29:37] exposure to various teratogen sand we're [29:39] talking about things like alcohol [29:40] nicotine various other nasty chemicals [29:42] or even mutations in really important [29:44] genes such as knotch what you get are [29:46] severe brain defects so the forebrain is [29:48] basically missing the midbrain as a [29:50] bubble the hind brains all messed up [29:51] what we've been able to do is to produce [29:54] by electrical simulations that tell us [29:56] what electrical state we would have to [29:58] restore in order to get back a normal [30:01] brain and in fact from these simulations [30:03] we can extract information for candidate [30:06] drugs that will open and close various [30:09] ion channels and you can see here a [30:11] successful applications of this this [30:12] animal has a mutation it's expressing a [30:16] mutation in a gene called notch which [30:17] normally causes this terrible [30:19] malformation of the brain but the [30:21] structure the gene expression and even [30:23] the function of the brain so we gave [30:24] them IQ tests believe it or not you can [30:26] do that in that in tadpoles you can test [30:28] them for learning and memory and they [30:30] get there they get [30:30] memories back despite this this really [30:33] terrible genetic mutation so we really [30:35] think that this is this by electrical [30:37] network is an incredibly tractable layer [30:40] that allows us to intervene in the [30:42] default state that is encoded by the [30:44] genome and so in the future what we [30:48] would like to do this is kind of the [30:49] endgame for this field which is of a [30:51] very long way away but the idea is to [30:54] have a kind of biological compiler where [30:56] we would be able to design our creatures [30:58] the way people now design things in CAD [31:01] or for 3d printing where you can design [31:02] at the level of anatomy and then we [31:04] would understand well enough what you [31:06] would have to do to make this to make [31:08] this structure actually grow and [31:09] evolution learned very early on to [31:12] exploit the computational properties of [31:14] electrical circuits electrical circuits [31:16] as some very convenient properties in [31:19] terms of information processing and [31:20] evolution caught on to this very early [31:23] and so under cracking this by electrical [31:25] code will have impacts not only on [31:27] regenerative medicine but also in [31:30] synthetic by engineering morphological [31:32] computation and so on this is really I [31:34] think a very important emerging field [31:36] and so what I want to are close now with [31:38] is to talk a little bit about the future [31:40] and this idea that perhaps we could make [31:43] a highly robust machine learning [31:45] platform technology that is based on non [31:48] neural architectures and we really are [31:50] interested in collaborators for this is [31:52] it possible that some of the limitations [31:54] of current machine learning is because [31:56] people have really been focusing on [31:58] imitating brains and especially human [32:00] brain architectures and what would [32:02] happen if we stepped way back and based [32:05] our attempts at this on send on cells [32:09] and tissues that are doing much more [32:12] much more fundamental information [32:13] processing than than than what we see in [32:15] brains so in fact some people have [32:18] started looking at this in other [32:19] structures so so bone heart pancreatic [32:22] cells all of these cells have been shown [32:24] to exhibit aspects of learning and in [32:26] particular robust adjustment to novel [32:28] conditions one of the most amazing [32:30] things about living things is how good [32:33] they are at adapting to novel [32:34] circumstances that they've never seen [32:36] before and cracking this problem will be [32:38] very important for robust robotics and [32:41] artificial intelligence but it's [32:44] to note that this is already present in [32:46] cells long before they Accord the [32:48] animals acquired brains so one of the [32:50] things that we're doing is to try to [32:53] extend the basic results of [32:56] connectionism neural networks and all of [32:58] those kinds of things to beyond the [33:00] brain try to try to develop a kind of [33:04] mathematics and and training algorithms [33:07] that can be used on non neural networks [33:09] and the other thing that we're doing is [33:12] we've we've had a little bit of success [33:14] coming up with contributions to what's [33:16] called the robot scientist field which [33:18] is basically which is basically [33:20] platforms that use machine learning to [33:23] help us actually understand what's going [33:25] on and to discover not just not just to [33:29] do number crunching but actually [33:30] discover a models of the kinds of [33:32] systems that were working with and so [33:35] what I would like to do now is to thank [33:37] the all the people that have done some [33:39] of the work that I've shown you today I [33:40] want to thank my funders in particularly [33:42] the Paul G Allen frontiers group and [33:44] DARPA and if you wanted to see what [33:47] two-headed worms actually look like and [33:50] behave like this is what it is here you [33:52] see them and we now we actually have [33:54] openings both for postdoctoral work and [33:56] for visiting scientists so if anybody is [33:58] interested in making robotic bodies for [34:01] various biological systems new AI [34:03] platforms based on non neural [34:06] architectures machine learning and all [34:08] kinds of computer science applications [34:10] we would love to collaborate and you can [34:12] get in touch with me here and I will [34:13] take questions thank you [34:35] hey how do you determine which [34:39] electrical activity is just due to like [34:43] functional movement and things like this [34:45] in the organisms day to day life and [34:47] which aspects are going to be used for [34:50] something like regeneration and then how [34:52] do you actually edit it yeah so okay so [34:55] so the the convenient thing is that the [34:57] kinds of electrical states that are [34:59] important for decision making at the [35:01] anatomical level have a very different [35:03] time scale than the kinds of things that [35:05] guide behavior so anything that has to [35:07] do with muscle motion with neural [35:09] activity the time scale you're talking [35:11] about is milliseconds okay the [35:12] electrical activity that we deal with [35:14] these patterns are stable and they [35:16] change over a scale of minutes hours in [35:18] some cases days so it's an entirely [35:20] different set of if you will frequencies [35:23] that you deal with and the way we I [35:25] think your second question was how do we [35:26] edit it yeah so so there's there's a [35:28] couple of different approaches that we [35:30] take in our model systems that we use at [35:32] the bench the best way to edit it is to [35:35] knock in controllable ion channels so we [35:38] will put in novel maybe light gated [35:40] maybe drug gate at different ion [35:41] channels that allow us to have very fine [35:43] control over all the electrical [35:45] properties in biomedical applications [35:47] you this kind of gene therapy is [35:48] probably best avoided so what we do [35:50] there is to develop cocktails of drugs [35:52] that will open and close specific iron [35:54] channels that are already in the patient [35:56] so we have a variety of computational [35:58] tools that let us look at what channels [36:00] and pumps are expressed in the tissue of [36:02] interest and then try to come up with a [36:04] strategy that's going to get us to the [36:05] electrical signaling that we want does [36:08] that interfere with the like short-term [36:10] electrical signaling it really doesn't [36:14] one of the one of the really remarkable [36:15] things that we found so I should I [36:17] should point out that back in about 2000 [36:19] when we first suggested this kind of [36:20] idea this was considered completely [36:21] crazy people said that electrical [36:23] properties of cells are housekeeping [36:25] parameters and that if we mess with it [36:26] everything's gonna die that you're gonna [36:28] get uninterpretable you know toxicity [36:30] none of that is true it turns out that [36:32] cells are very very good at [36:34] distinguishing the different types of [36:37] electrical signals and things that we do [36:40] at the long-term scale do not affect [36:42] behavior they can of course you can [36:44] overdo it I mean part of what we've been [36:46] doing for the last 15 years is coming up [36:47] with protocols [36:48] that have the least amount of side [36:49] effects possible right so you certainly [36:51] can overwhelm the system but these these [36:53] are designed to work in parallel and in [36:55] fact in all of these organisms the [36:57] regenerating planarian or the [36:58] regenerating salamander this is running [37:01] around eating behaving doing all that [37:02] stuff [37:03] while all the cells are figuring out [37:05] what to do so there's a multi scale set [37:07] of problem solving that does not [37:09] interfere with each other [37:10] phenomenal talk thank you okay thank you [37:15] very much for your a very interesting [37:16] talk actually you know I was very [37:18] excited about your biological inspired [37:21] work and then I think you know there's a [37:23] research area called synthetic biology [37:25] and then people are trying to create [37:27] actually you know people can create [37:29] artificial cell from scratch and also [37:31] you know now if you attach electrode to [37:33] the small insects you can you know [37:35] control the insects by sending a signal [37:37] to them and also you know people are [37:39] trying to create truly intelligent [37:41] robots by from robotics perspective and [37:44] also in an algorithmic perspective so [37:46] I'd like to know you're in a personal [37:49] opinion about what is the correct [37:51] direction of true and creating truly [37:54] intelligent agent or creatures or [37:56] whatever I mean yeah all right well I [37:59] can say a couple things we do have a [38:01] section of my group it works on [38:03] synthetic morphology so we have some [38:05] really promising it's not published yet [38:07] so I didn't bring it up here but there's [38:08] the way we have some pretty amazing new [38:10] data on the ability of cells to learn to [38:14] cooperate in completely different [38:16] configurations so you can make [38:17] artificial living machines that are [38:19] nothing like the animal that from which [38:21] they came and they're able to learn to [38:23] cooperate and through some sort of a [38:25] combination of emergence and guided [38:27] self-assembly we can make some some [38:29] machines that are just you know just [38:30] amazing I think that's a very exciting [38:32] exciting field of course the other thing [38:34] that is really interesting is to try to [38:36] extract some of the policies that cells [38:38] and tissues use and try to implement [38:41] them either directly in the hardware or [38:43] to bio interface so so we've done a [38:45] little bit of that to is try to [38:46] interface some you know the UM one of [38:49] the one of the things I didn't bring up [38:51] here for for lack of time is we found [38:54] out that if you if you make a tadpole [38:56] where the eyes are on the tail instead [38:58] of in the head those tadpoles can see [39:00] out of that are those eyes perfectly [39:01] well if [39:02] the brain has you know for millions of [39:03] years its benefit it's evolved to expect [39:05] visual input from this region up in the [39:07] head and all of a sudden now there's [39:08] this weird itchy patch of tissue in the [39:10] back that sort of sending you some [39:11] signals the brain has no trouble [39:13] figuring out that that's visual data and [39:14] being able to learn so those plasticity [39:16] ziz is incredible and we need to we need [39:18] to understand how this works and then I [39:20] think I think the artifacts that we make [39:22] are going to be just transformative [39:24] thank you [39:27] great talk yes do we know something [39:31] about plants in terms of electrical [39:33] signaling and also there is mechanical [39:36] signaling which is very slow relatively [39:39] slow so do you know anything about [39:40] interactions between these different [39:42] modes yeah [39:43] great questions so plants yes my group [39:46] doesn't work on plants but there's a [39:47] couple of other labs that do and it's [39:50] it's very interesting plants do very [39:52] similar things with electrical signaling [39:55] as a lot of what I showed you here is [39:57] true in plants as well which is [39:59] remarkable because plants and animals [40:00] shared different origins of [40:02] multicellularity right so they both [40:04] discovered this independently and that's [40:06] because they're the ancestors even from [40:08] bacteria so even bacterial biofilms [40:10] propagate potassium waves that are brain [40:12] like and there's been some interesting [40:14] work on that so the decision-making for [40:16] example and plant roots as they as they [40:18] go through the soil and figure out where [40:19] they're gonna go all the way through [40:20] morphogenesis of the plant itself a lot [40:23] of what I've said here it holds holds [40:25] for plants as well mechanical forces [40:28] yeah mechanical forces are absolutely [40:30] also important in all of this and [40:32] mechanics interacts tightly with bio [40:36] electrics there are stretch and [40:38] pressure-sensitive ion channels so they [40:40] will convert mechanical forces into [40:42] electrical forces and there are proteins [40:44] that do the reverse that basically under [40:46] electrical gradients across the membrane [40:47] they will exert stress and strain so so [40:50] those systems interact also of course [40:52] both of them interact with more familiar [40:54] genetic and biochemical signaling and so [40:57] these are three pieces of the puzzle and [40:59] the whole system is kind of this this [41:03] dynamical system where all of these [41:04] things constantly interact the one [41:06] question that we always ask ourselves is [41:08] at what point in that control structure [41:11] is it most efficient to intervene so [41:13] right now the majority of my colleagues [41:15] are [41:15] are working on biochemistry and genetics [41:17] oh so they're intervening at the [41:18] molecular level so they're looking for [41:20] growth factors and diffusible signals [41:22] what we have seen is that intervening [41:25] and the electrical level gives extremely [41:27] efficient control so if you want an eye [41:29] you can try to assemble it cell by cell [41:32] but it's really much more efficient to [41:33] trigger the whole the whole subroutine [41:35] as it were and so I think that a lot of [41:37] the important decisions are made at the [41:39] electrical level but it's absolutely [41:41] true that mechanical forces and did you [41:44] know transcriptional and chemical [41:46] networks are downstream as part of how [41:48] this stuff gets executed in the actual [41:50] tissue hi thank you for your talk it was [41:55] very fascinating so you talked a lot [41:58] about how on-like cells that regenerate [42:02] limbs are making some sort of decision [42:03] and you said that plant roots they make [42:05] decisions as they navigate through the [42:07] ground and somewhere on a higher level [42:10] we also make decisions in our brain so [42:11] somewhere between there this thing [42:13] called consciousness arises is there any [42:16] sort of real definition of consciousness [42:18] and if so is it a gradient is it a line [42:23] dude can you speak to that at all yeah [42:25] III get into enough trouble talking [42:28] about decision making at the level of [42:30] cells so I don't think I'm gonna void [42:32] opinion on cautiousness per se because [42:33] there's a lot of people that work on [42:35] this and argue about that stuff but I [42:37] will say that there is there is an [42:39] emerging field of basal cognition and [42:41] these are groups of people that talk [42:43] that talk about memory learning sensing [42:45] creativity all the kinds of things all [42:47] the way from bacteria plants animals [42:50] cells and and in fact even group [42:52] colonial organisms like ant hills you [42:54] know there are things that an anthill [42:55] knows that the individual ants don't [42:56] know and so on so there's lots of people [42:58] that worry about that kind of stuff and [42:59] we discuss all these things at meetings [43:01] so I don't think I'm I haven't eating [43:03] profound to say specifically about [43:04] consciousness but I do feel very [43:06] strongly that all of this is a continuum [43:08] that there is nothing radically special [43:11] about it there's a lot of people that [43:12] feel that some of these terms should be [43:14] left sort of exclusively in the realm of [43:16] human psychology and I think that's a [43:17] profound mistake I think that's a that's [43:19] a mistake both on evolutionary grounds [43:21] and you know just in there is no [43:24] principled distinction there are only [43:25] shades of the shades of quantitative [43:27] difference [43:27] and so you can see things like [43:29] anticipation yeast can anticipate you [43:31] know yeast can learn simple patterns and [43:33] they'll anticipate to them there these [43:35] things go all the way down to the to the [43:36] very base of the Tree of Life and so [43:38] what you're gonna call us consciousness [43:40] I mean I have I have a very good [43:42] colleague who thinks that cautious a [43:44] consciousness appears when something she [43:46] calls you al unlimited associative [43:48] learning takes place and there are only [43:50] some organisms that are able to carry [43:51] out a kind of generic associative [43:54] learning and and some organisms cannot [43:55] and that's where she puts the line I'm [43:57] not sure I think it's as reasonable as [43:59] anything else but for sure it's a [44:01] continuum and whether it's consciousness [44:03] or just functional cognitive abilities [44:04] they go all the way to the beginning [44:06] thanks very much [44:07] I was amazed by your talk when you apply [44:14] these frequencies or this electronical [44:16] signal how do you do that do you put in [44:19] just like an electrode into something or [44:21] are you poking it with something and [44:22] what is the time frame that these [44:24] signals need to be applied right time [44:27] frame that's an important question [44:28] so under and we never apply any [44:31] electrical fields we have no electrodes [44:33] we apply no electrical fields and we [44:35] don't do anything with frequencies what [44:37] we try to do is to alter the endogenous [44:40] state of resting potential across cells [44:43] and the way we do that is we target the [44:46] ion channel proteins that are [44:48] endogenously expressed in these cells we [44:50] either close them we open them and we [44:52] thus affect the electrical state so this [44:55] is done either genetically or [44:56] pharmacologically in the case of our [44:58] regenerative medicine applications is [45:00] with drugs small molecule drugs but we [45:02] can in let's say in tadpoles with you [45:04] know a really nice way to do it as [45:05] optogenetics you put in something like [45:07] channelrhodopsin and then you shine a [45:08] light and and the pattern of the light [45:10] can tell it can real you can you can [45:12] really have pretty fine control over [45:13] where the electrical states go the time [45:16] scale is really critical because a lot [45:19] of the types of effects that I've shown [45:22] you are literally triggers to these what [45:24] we call these subroutine calls right of [45:26] patterning the time that you have to do [45:28] it is incredibly short so in the Frog [45:30] leg we treat the wound we treat the [45:33] amputation wound for 24 hours that gives [45:36] you nine months of leg growth so after [45:39] the first 24 hours you walk [45:41] you never touch it again and the whole [45:42] thing goes for 24 hours you could [45:45] intervene at different steps of the wave [45:46] instead of a frog leg you wanted [45:48] something else you could go in and try [45:49] to make it do something else and we're [45:51] certainly working on that as a way to [45:52] understand what's happening but the [45:54] treatments are often incredibly brief [45:56] because these one of the properties of [45:59] these electrical circuits is they have [46:00] these stable attractors that once you [46:03] knock them into that sort of [46:04] neighborhood of the state space you [46:06] don't need to hold them there they're [46:07] gonna they're gonna maintain state and [46:08] so our goal is to map out all of these [46:10] things to understand what is the minimum [46:12] we can do to get the maximum outcome [46:14] thank you hi could you talk more about [46:19] these new architectures these non neural [46:22] architectures and what would that get us [46:24] would we have more compositional systems [46:27] more flexible ones things that don't [46:30] just think but also feel or even look at [46:32] testbed for designing new organisms or [46:34] new like use that to design biology yeah [46:38] I think I think what it's going to get [46:40] us I don't know if it's going to get us [46:41] systems that feel and and how we would [46:44] know if they did that's a whole other [46:45] sort of kettle of fish but I what I [46:48] think it's going to get us is much more [46:50] robust adaptive systems one of the [46:53] things that biological systems are [46:54] really good at is adaptively adjusting [46:59] to very novel environments and so I [47:01] think one of the problems that we have [47:02] in this field and you guys can correct [47:04] me if I'm wrong but I think that one of [47:05] the problems we have is that systems [47:07] that are very good at one particular [47:09] task often are not that great when you [47:11] shift it to it to a different task [47:13] biology isn't like that these these [47:15] these cells are and these tissues are [47:17] incredibly good at dealing with with [47:19] very novel scenarios in an adaptive way [47:22] and well one of the differences between [47:25] biological structures and and the kinds [47:28] of artifacts that I've seen people make [47:29] is that biology as competency at every [47:31] level so the individual molecular [47:33] networks have memory and flexibility the [47:35] cells do the tissues do the organs do [47:36] every level has its own competency so [47:38] you're building things out of very [47:40] intelligent components and their various [47:42] other things that we can talk about it [47:44] you know at length but I think what it's [47:46] going to get us is systems that are [47:48] fundamentally good at pursuing goals in [47:51] the light of an unfriendly and [47:54] that's what that's what all of life is [47:55] right these biological systems have very [47:57] specific goals the environment throws [47:58] all kinds of crazy scenarios at them and [48:00] they are really good at navigating all [48:03] of that and I think we don't really know [48:05] how to do that so that's what we'll get [48:06] and one more do you have any sketch of [48:10] how we would build such a thing I have I [48:12] have many ideas on how we should try and [48:14] so I'm happy to talk to you afterwards [48:15] and talk details thank you hi I found [48:20] this is an absolutely incredible talk [48:22] and one thing that that came to mind [48:25] when I was thinking about this talk was [48:26] ethical concerns it seems like when [48:29] you're working with like small animals [48:30] doesn't really matter but when you start [48:32] getting into humans and start doing [48:34] things that could potentially have [48:35] really systemic effects in them that [48:38] they're that people may start objecting [48:39] to that and I'm wondering if you have [48:41] thoughts about how to head off those [48:42] ethical concerns sure yeah good question [48:45] I I can say a couple of things I think [48:47] that on the one hand I'll point out that [48:50] in terms of in terms of F Milk ethical [48:53] concerns the kinds of things that I've [48:54] been talking about here are way low on [48:57] the list of compared to competing [48:59] technologies that are already out there [49:00] so people that are doing genomic editing [49:02] people that are doing infectious disease [49:04] research you know creating novel viruses [49:07] and I mean the stuff that's that's being [49:09] done is like you know way beyond [49:11] anything we're talking about here and in [49:13] fact on that scale this is actually [49:14] fairly subtle because what I'm [49:16] specifically saying is I think we can [49:17] develop a strategy for regenerative [49:19] medicine that does not require [49:21] manipulating the genome it does not [49:23] require any kind of viral delivery [49:27] vectors are all this so so I think I [49:28] think I think there are way bigger [49:29] problems in in the field and then what [49:32] and what you know then what we're doing [49:34] in terms of in terms of ethics all right [49:36] well thanks very much hi so ml community [49:41] has done some work in mimicking signal [49:45] transduction networks particularly in [49:47] t-cells or immune cells in general and [49:50] so I'm curious if you can help me [49:53] understand what the relationship between [49:54] some of these ion channel based networks [49:57] and signal transduction networks are [50:00] and how that kind of ties into the AI [50:02] theme and secondly so the assistant the [50:06] computational assistants biology [50:08] community has kind of developed a lot of [50:11] computational models of this type of [50:12] thing so weight ranging from Eau de [50:14] based networks to partial differential [50:16] equations to like boolean networks time [50:18] to scale those up to a whole mob [50:19] wholesale models does that relate to [50:22] what you're talking about here in terms [50:23] this AI platform for non zero [50:25] architectures and relationships are you [50:28] doing any of that in your lab it's yeah [50:30] it's real a that that's actually a very [50:33] kind of lengthy question it sounds like [50:35] we should take that up okay yeah I think [50:39] we can talk about that at length that [50:41] that's that that'll take a pretty [50:43] lengthy a pretty lengthy answer but I'm [50:44] very happy to talk to you about it it's [50:47] high rated question it's there is a [50:53] longer story of reaction diffusion [50:55] models for patterning and so I was [50:58] wondering if it's a mechanism you and to [51:02] some extent it has been confirmed with [51:04] BMP signaling in name bands so I was [51:06] wondering what is your view of the [51:08] interaction of this electrical system [51:10] with a reaction diffusion system is it [51:12] upstream is it interfering with it is it [51:15] the reaction diffusion system itself [51:16] yeah so so two things we as far as we [51:20] can tell right now in the cases that [51:21] we've looked at it is upstream so it [51:23] helps control the downstream are the [51:26] systems for chemical morphogens and as [51:30] far as what was the other sorry what was [51:32] the second question [51:34] I was wondering if it could be the [51:36] reaction diffusion system by itself yeah [51:38] yeah actually that's true too so so one [51:40] of the really interesting things about [51:42] these kind of electrical networks is [51:44] that when they're spatialized and [51:46] distributed over a group of cells they [51:48] themselves purely at the electrical [51:49] level with no transcriptional changes [51:51] underneath Kendu spontaneous symmetry [51:53] breaking and touring like patterns so [51:55] that's that's absolutely true as well [51:57] okay thank you thank you thank you thank [52:01] you everyone for all the great questions [52:03] and thank Michael Evan