Sanofi Deploys AI Protein Models to Replace Physical Drug Screening

Faster, Better, Smarter Biologics: Going for Gold with AI With Sanofi — Ai4

Ai4

Adopting in silico foundation models and active learning has become an operational necessity for pharmaceutical survival as the complexity of multi-specific biologics outgrows traditional wet-lab capacities.

AI-assisted summary of Faster, Better, Smarter Biologics: Going for Gold with AI With Sanofi — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 26, 2025, this presentation features Maria Wendt, VP and Global Head of Preclinical Computational Innovation Strategy at Sanofi. According to the description, the session explores the intersection of artificial intelligence and biopharmaceutical innovation, detailing Sanofi's Digital Biologics Strategy, protein language models, and virtual screening methods designed to accelerate large molecule discovery and tackle previously undruggable targets.

## TL;DR Tracing the biological resolution revolution from Mendel's pea plants to live high-definition microscopy of cytotoxic immune cells, Maria Wendt outlines how modern medicine has reached a fourth wave centered on complex, multi-specific biologics. To manage the immense combinatorial challenge of designing multi-domain molecular machines, Sanofi is deploying protein language models, structural representations, and active learning pipelines that act as virtual surrogates for traditional wet-lab assays. This computational shift enables high-throughput in silico screening and de novo protein optimization, requiring pharma organizations to embrace tech-industry software engineering practices as a fundamental pillar for future survival.

## Atmosphere & style The presentation strikes a grounded, refreshing balance between tech-conference optimism and rigorous industrial pragmatism, opening with a lighthearted rejection of macro AI doom in favor of tangible medical impact. Speaking candidly as a computational mathematician who has spent 18 years leading wet-lab research without ever picking up a pipet, Wendt delivers an authentic, conversational walkthrough of modern drug discovery. Watching the session feels both intellectually clarifying and motivating, anchored by vivid cellular visualizations and a genuine enthusiasm for engineering precision molecular machines.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=4s) Preclinical drug discovery and healthcare represent one of the most immediate and impactful frontiers for AI to deliver tangible humanitarian benefits. - [02:01](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=121s) A modern biological 'resolution revolution' in high-definition microscopy and live cellular tracking generates the high-dimensional datasets required to train advanced AI models. - [06:23](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=383s) Biopharmaceutical organizations are increasingly appointing in silico and computational leaders to direct discovery strategy and modernize traditional wet-lab workflows. - [08:25](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=505s) The fourth wave of biopharmaceuticals focuses on smart, multi-specific antibodies that can bridge immune cells to tumors and trigger localized co-stimulatory signals. - [11:01](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=661s) Engineering modular biologics like tetravalent nanobodies requires simultaneous multi-parameter optimization across binding potency, expression yield, and biophysical stability. - [13:36](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=816s) Sanofi's Digital Biologics Strategy combines immediate machine learning augmentation of active pipelines with long-term transition to fully in silico de novo drug design. - [15:35](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=935s) Protein language models treat linear 20-amino-acid sequences like language tokens, learning evolutionary constraints and structural fitness via self-supervised pre-training. - [15:35](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=935s) In silico surrogate assay models embedded into agentic virtual screening systems can rapidly rank candidate molecules, replacing months of physical wet-lab testing. - [18:36](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1116s) Integrating 3D structural constraints with sequence models in an active learning framework allows researchers to test small, informative protein batches rather than screening tens of thousands of physical variants. - [19:20](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1160s) AI adoption in pharmaceutical discovery has evolved from exploratory pilot experiments into an industrialized, core operational necessity expanding across proteins, mRNA, and lipid nanoparticles.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=4s) The Big Promise of AI in Healthcare and Life Sciences

Setting a distinct tone from the tech-heavy presentations earlier in the day, the session opens with a colorful molecular visualization of an engineered antibody docking onto a cellular target. Reflecting on the morning keynote delivered by Jeff, the speaker notes how the opening warnings about "AI ending humanity was a little bit kind of a downer to start the day." However, the keynote concluded on an inspiring note regarding the transformative promise of artificial intelligence in healthcare.

Surveying the audience, she highlights how life sciences and healthcare applications represent some of the most meaningful frontiers for AI innovation. Despite an overall conference attendance of around 8,000 participants, presentations dedicated specifically to biopharmaceuticals and healthcare remain disproportionately rare—a dynamic that must urgently shift in the coming years as computational methods mature.

- The session focuses squarely on preclinical drug discovery and computational design of biologic therapeutics. - While macro AI discourse often dwells on existential risks, biopharmaceutical applications offer concrete, near-term humanitarian benefits.

---

[02:01](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=121s) The Biological Resolution Revolution: From Mendel's Peas to High-Definition Microscopy

To illustrate how dramatically biological research has evolved, the discussion contrasts modern imaging with the foundational experiments of Gregor Mendel 150 years ago. As the father of modern genetics, Mendel deduced the laws of inheritance by tallying discrete, macro-level physical metrics: the height of pea plants, the size of pea pods, and the count of individual peas within a pod.

```mermaid flowchart LR accTitle: Evolution of Biological Observation subgraph Classical["150 Years Ago: Classical Genetics"] Mendel["Gregor Mendel's Experiments"] --> MacroData["Macro Phenotypic Counts<br/>(Pea pod size, seed counts)"] end subgraph Modern["Modern Era: Resolution Revolution"] Microscopy["High-Definition Microscopy & Molecular Tagging"] --> CellularDynamics["Live Cellular Interrogation<br/>(Immune cell sniffing tumor cell)"] CellularDynamics --> BigBioData["Unprecedented Multimodal Data"] end MacroData -.->|"Resolution Revolution"| Microscopy ```

Today, the field is undergoing a "resolution revolution." Modern researchers observe cellular dynamics in real time using high-definition microscopy and sophisticated fluorescent molecular labels. A highlighted video demonstrates three immune cells—such as T-cells or macrophages—actively "sniffing" and probing a central cancer cell to evaluate whether it is healthy or malignant. Once classified as cancerous, the T-cell perforates the target's membrane and injects toxic granules (visible as red fluorescent markers) to destroy the tumor cell from within. This leap in observational capability generates massive, high-dimensional datasets that provide fertile ground for training sophisticated AI models.

- Gregor Mendel established the concept of inherited genes using basic phenotypic observations of pea plants. - Modern high-definition microscopy allows live tracking of immune effector mechanisms, producing unprecedented volumes of biological data.

---

[04:33](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=273s) Navigating the Drug Discovery Pipeline: From Disease Target to Clinical Translation

The biopharmaceutical research pipeline broadly progresses across three major stages prior to full clinical development:

1. **Disease Biology & Target Identification:** Investigating disease mechanisms to identify druggable biological targets. 2. **Drug Invention:** Designing, engineering, and optimizing the therapeutic molecule (the primary focus of this talk). 3. **Clinical Translation:** Validating that the engineered candidate molecule demonstrates safety, efficacy, and predictable pharmacokinetics in humans.

Wendt reflects on her early technical training as a chemical engineer operating bioreactors, running purification columns, and managing petrochemical processing plants. Driven by an affinity for applied mathematics, she pivoted early in her career toward artificial intelligence and bioinformatics during the early 2000s—a period when the field was predominantly defined by expert systems and single-layer perceptron neural networks.

---

[06:23](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=383s) In Silico Leadership in a Wet-Lab World

Following her transition to bioinformatics, Wendt spent 18 years working in Basel, Switzerland—a global epicenter of the European pharmaceutical industry. Her work focused entirely within in silico environments, developing software platforms and computational algorithms to analyze large-scale multi-omics data, including genomics, transcriptomics, proteomics, and metabolomics.

She was subsequently appointed Head of Biologics Research at Sanofi, tasked with directing an organization staffed predominantly by wet-lab scientists, including molecular biologists, protein chemists, and cellular biologists. Despite having "never held a pipet" throughout an 18-year computational career, her leadership role underscores the paradigm shift taking place across the industry: biopharma organizations are placing computational thinkers at the helm of discovery to modernize laboratory research and transition wet-lab workflows toward an in silico-first methodology.

- Basel, Switzerland served as the home base for 18 years of specialized in silico multi-omics research. - Modern biologics discovery requires bridging the traditional divide between wet-lab experimentalists and computational data scientists.

---

[08:25](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=505s) The Fourth Wave of Biopharmaceuticals: Engineering Smart Multi-Specific Antibodies

Biopharmaceuticals have entered the "fourth wave of biopharmaceutical innovation," moving beyond standard small-molecule drugs (such as aspirin) and basic monoclonal antibodies toward complex, "smart" engineered biologics. A featured milestone—highlighted on the cover of *Nature*—is a highly engineered tri-specific antibody developed at Sanofi designed to combat cancer.

```mermaid flowchart TD accTitle: Tri-Specific Antibody Mechanism of Action subgraph TriSpecific["Sanofi Tri-Specific Engineered Antibody"] Arm1["Arm 1: Tumor-Specific Binding"] Arm2["Arm 2: Immune Cell Engager (T-Cell / NK Cell)"] Arm3["Arm 3: Co-Stimulatory Signaling Domain"] end TumorCell["Malignant Tumor Cell"] --- Arm1 Arm2 --- ImmuneCell["Cytotoxic T-Cell / NK Cell"] Arm3 -.->|"Costimulatory Signal:<br/>'Fire until tumor is killed'"| ImmuneCell ```

Whereas native antibodies possess two identical antigen-binding arms, this tri-specific format incorporates three distinct functional arms: - **Arm 1 & Arm 2:** Function as a physical bridge, tethering a cancer cell directly to an immune effector cell, such as a Natural Killer (NK) or cytotoxic T-cell. - **Arm 3:** Delivers a localized co-stimulatory signal instructing the engaged immune cell: "don't get tired... fire fire until you kill the tumor cell."

Beyond immune cell engagement, fourth-wave therapeutics exhibit contextual responsiveness. Examples include engineered antibodies that actively transport therapeutic cargo across the blood-brain barrier by piggybacking on native receptors, or pH-sensitive biologics that remain inert while circulating in normal tissue and activate strictly within the acidic microenvironment of a solid tumor.

- Small-molecule therapeutics lack the physical scale and multi-valency required to assemble multi-protein cellular bridges. - Context-responsive biologics improve therapeutic windows by minimizing off-target toxicity in healthy peripheral tissues.

---

[11:01](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=661s) Assembling Molecular Machines: Complexity, Nanobodies, and Key Assays

Modern drug developers are expanding beyond single-protein architectures into complex formats, including mRNA therapies, genomic medicines, and hybrid conjugate molecules. Designing these multi-domain structures resembles the precision engineering of complex nanoscale machines.

To construct a modular biologic—such as a tetravalent nanobody—multiple individual functional units must be selected and assembled: - Two distinct binding domains targeted against separate antigens on a tumor cell. - A binding domain targeted to human serum albumin to extend circulatory half-life in the bloodstream. - An immune cell engagement domain (e.g., recruiting T-cells).

```mermaid flowchart LR accTitle: Modular Biologic Optimization Parameters subgraph MultiDomain["Tetravalent Nanobody Assembly"] D1["Target Antigen 1"] D2["Target Antigen 2"] D3["Serum Albumin (Half-life)"] D4["T-Cell Receptor Engager"] end MultiDomain --> Screening["Multi-Parameter Optimization"] subgraph Properties["Required Candidate Properties"] P1["Binding Potency"] P2["Expression / Manufacturability"] P3["Thermal & Serum Stability"] P4["Effector Functions"] end Screening --> Properties ```

Historically, discovering viable candidates required building and screening vast wet-lab physical display libraries to identify single "winners" for each module before combining them. The combined construct must then be rigorously optimized across several critical biophysical parameters: target binding potency, expression titers for manufacturing, biophysical stability to prevent aggregation in storage or in vivo degradation, and immune effector recruitment functions.

---

[13:36](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=816s) Sanofi's Digital Biologics Strategy: Augmenting Today, Transforming Tomorrow

To address the exponential combinatorial complexity of multi-domain biologics, Sanofi has deployed a two-pronged computational strategy:

1. **Immediate Pipeline Augmentation:** Integrating AI and machine learning tools into all active, running drug discovery programs to accelerate decision cycles. 2. **Transformational In Silico Discovery:** Building toward an end-to-end computational paradigm where therapeutic candidates are designed entirely in silico prior to experimental validation.

While fully autonomous, computer-generated de novo drug design might have been dismissed as speculative fantasy in past years, recent advances have made it "entirely probable." Large pharmaceutical enterprises with workforces exceeding 100,000 employees are rapidly expanding internal computational and machine learning units to match capabilities found in leading tech and financial institutions.

- Sanofi views computational design not merely as an efficiency tool, but as a replacement for traditional screening cascades. - Enterprise pharma organizations are building dedicated AI research teams rather than relying solely on external software providers.

---

[15:35](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=935s) Protein Language Models and In Silico Virtual Screening Pipelines

Proteins are fundamentally linear polymers composed of 20 natural amino acid letters, making their primary sequences naturally amenable to transformer-based foundational language model architectures. Just as large language models (LLMs) learn underlying syntax and semantics by performing next-token prediction on text, protein language models (pLMs) learn evolutionary constraints, biophysical rules, and structural fitness by predicting amino acid tokens.

```mermaid flowchart TD accTitle: Foundation Model to Agentic Virtual Screening Pipeline pLM["Self-Supervised Protein Language Model<br/>(Sequence & Surface Representations)"] --> Embeddings["Latent Embeddings"] subgraph InSilicoSurrogates["In Silico Surrogate Assay Models"] M1["Binding Affinity Predictor"] M2["Expression Yield Predictor"] M3["Thermal / Colloidal Stability Predictor"] end Embeddings --> InSilicoSurrogates InSilicoSurrogates --> AgenticSystem["Agentic Virtual Screening Engine<br/>(High-Throughput Candidate Ranking)"] ```

Sanofi leverages self-supervised foundation models that map both primary sequences and 3D molecular surface topologies into continuous latent embedding spaces. These representations power a suite of downstream predictive surrogate models designed to replace time-consuming wet-lab assays:

- In silico predictors evaluate critical candidate properties—such as binding affinity, expression yield, and stability—that traditionally consume months or years of laboratory testing. - Multiple surrogate models are threaded together into high-throughput, agentic virtual screening systems capable of searching vast combinatorial sequence spaces. - Predicting the next amino acid within a target sequence provides a principled, generative mechanism for protein engineering and optimization.

[18:36](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1116s) Active Learning and Structural Language Models for Complex Proteins

Rather than forcing wet-lab scientists to synthesize and express upwards of 10,000 individual protein variants, Sanofi employs active learning frameworks that require testing only a small, informative subset. To tackle the immense structural complexity of these large molecules, the team invested heavily in engineering advanced protein representations. This approach bridges sequence-based protein language models with structural context—infusing explicit 3D architectural constraints directly into the models.

This hybrid modeling strategy yields highly accurate predictive tools capable of both evaluating candidate libraries and steering de novo molecular design. By iteratively selecting the most promising variants through computational screening before wet-lab validation, researchers can pinpoint and optimize the most viable candidates for exceptionally complex biologic targets.

```mermaid flowchart LR accTitle: Active Learning and Hybrid Protein Modeling Loop subgraph DataGen ["Wet Lab Phase"] A["Express Small Protein Subset"] end subgraph AIModeling ["Computational Pipeline"] B["Protein Language Models<br/>+ Structural Data"] C["Active Learning Engine"] end subgraph Output ["Candidate Selection"] D["Optimized Design &<br/>Selection of Complex Biologics"] end

A --> B B --> C C --> D D -. "Iterative feedback".-> A ```

- **Experimental Efficiency:** Active learning dramatically lowers screening overhead by substituting mass laboratory expression (such as 10,000 proteins) with targeted, model-guided small batches. - **Multimodal Architectures:** Blending sequence-trained foundation models with structural representations enables precise virtual screening and generative optimization for difficult biologic modalities.

---

[19:20](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1160s) Industrializing AI in Pharma: From Exploratory Prototypes to Core Survival

Across the pharmaceutical sector, artificial intelligence has fundamentally transitioned away from isolated, exploratory pilot projects into a core prerequisite for long-term viability. Organizations like Sanofi are actively industrializing and productizing their machine learning pipelines, importing rigorous engineering methodologies, deployment practices, and operational standards directly from the software industry.

Furthermore, these computational approaches extend well beyond standard antibody and protein engineering. Sanofi is applying specialized language models across a diverse spectrum of modalities, including mRNA sequence optimization and lipid nanoparticle (LNP) delivery formulations. Realizing the full potential of these platforms requires bridging biopharma with top-tier tech talent, making the recruitment of computational and machine learning specialists a critical priority.

- **Strategic Realignment:** AI adoption in biopharma is no longer treated as an experiment, but as a critical operational necessity—viewed across the industry as "really survival in the future." - **Software Productization:** Enterprise biopharma is embedding software-engineering discipline to productize models into reliable, reusable discovery platforms. - **Expanding Modalities:** Foundation and language models are being actively deployed beyond proteins into mRNA design and lipid nanoparticle (LNP) optimization.

## Video description Maria Wendt, VP, Global Head of Preclinical Computational Innovation Strategy, Sanofi

Abstracts

What if we could revolutionize the way we design and develop biologics? In this talk, I will explore the cutting-edge intersection of artificial intelligence and biopharmaceutical innovation. I'll share how we're leveraging AI to accelerate large molecule research, reduce cycle times, and improve success rates in creating next-generation therapeutics. You'll learn about our comprehensive Digital Biologics Strategy, including sophisticated protein language models and virtual screening techniques for complex biologics. Join me to discover how AI is transforming the future of medicine, enabling us to create smarter, more functional biologics that address previously undruggable targets and complex diseases.

[Source](https://www.youtube.com/watch?v=OsTSYLbXRzM)

中文简报

赛诺菲部署AI蛋白质模型替代物理药物筛选 随着多特异性生物制剂的复杂性超出传统湿实验的能力极限，采用in silico基础模型和主动学习已成为制药企业生存的运营必然要求。 Maria Wendt梳理了从孟德尔豌豆实验到细胞毒性免疫细胞高清活体显微成像的生物学“分辨率革命”，指出当代医学已迎来以复杂多特异性生物制剂为核心的第四次浪潮。为应对设计多结构域分子机器面临的巨大组合挑战，赛诺菲正在部署蛋白质语言模型、结构表征及主动学习管线，以此作为传统湿实验检测的虚拟替代方案。这种向计算方法的转变实现了高通量计算机模拟筛选和从头蛋白质优化，同时也要求制药机构将科技行业的软件工程实践视为未来生存的基本支柱。 高分辨率显微成像与活细胞追踪技术的突破，正为训练先进 AI 模型提供高维度、动态的生物学数据集。 (2:01) 生物制药进入聚焦复杂多特异性抗体的第四波创新，赛诺菲展示了能够桥接免疫细胞与肿瘤并释放局部共刺激信号的三特异性抗体。 (8:25) 赛诺菲实施双轨数字生物制药战略，既利用机器学习工具加速现有研发管线，又推进端到端的全计算机从头药物设计。 (13:36) 蛋白质语言模型将氨基酸序列视作语言词元进行表征学习，并结合替代实验模型与主动学习循环，大幅削减传统湿实验筛选周期。 (15:35)

Reading summary

Sanofi is shifting its preclinical biologics discovery from traditional wet-lab screening toward an in silico-first engineering workflow driven by protein language models, active learning, and software-grade productization. The Rise of High-Dimensional Biological Data (2:01) Biopharmaceutical discovery has progressed from Gregor Mendel's macro-level phenotypic tallies 150 years ago to modern high-definition microscopy that captures dynamic cellular interactions in real time. Researchers can now observe immune cells probing tumor membranes and injecting cytotoxic granules, producing high-dimensional datasets that provide the empirical foundation required to train advanced machine learning architectures. Engineering Complex, Multi-Specific Molecular Architectures (8:25) Biopharmaceutical development has advanced into a "fourth wave" characterized by modular, multi-domain therapeutics rather than traditional small molecules or standard monoclonal antibodies. Formats such as tri-specific antibodies and tetravalent nanobodies require simultaneous multi-parameter optimization across binding affinities, expression titers, manufacturability, and biophysical stability. These complex formats can incorporate contextual responsiveness, such as crossing the blood-brain barrier via native receptors or activating only within the acidic microenvironment of solid tumors. A tri-specific cancer antibody highlighted by the speaker binds a tumor cell, recruits an effector cell like a cytotoxic T-cell or natural killer cell, and delivers a localized co-stimulatory signal instructing the immune cell to sustain its attack. Protein Language Models and Virtual Screening Pipelines (15:35) Because proteins are linear sequences of 20 natural amino acids, they are well suited for transformer architectures that learn evolutionary constraints and structural fitness through token prediction. Sanofi employs self-supervised foundation models that map primary sequences alongside 3D surface topologies into continuous latent embeddings to power predictive surrogate models for binding affinity, expression yield, and stability. Rather than expressing up to 10,000 physical protein variants in a laboratory, researchers run active learning loops that test only small, informative subsets experimentally. By coupling sequence-based language models with explicit structural representations, virtual screening platforms and agentic workflows can evaluate combinatorial sequence spaces and direct de novo molecular design before wet-lab validation. Industrializing Machine Learning Across Drug Modalities (19:20) Sanofi pursues a two-pronged computational strategy: augmenting active discovery pipelines to accelerate near-term decision-making, while building toward an end-to-end paradigm where molecules are designed computationally. This transition reflects an industry-wide pivot from treating artificial intelligence as exploratory prototypes to viewing it as a core operational requirement. Pharmaceutical organizations are adopting software engineering disciplines to productize their machine learning pipelines into reusable platforms. Beyond recombinant antibodies, these computational approaches and language models are being expanded into newer modalities, including mRNA sequence design and lipid nanoparticle formulations. While computational design has advanced from speculative concept to an active discovery strategy, realizing fully autonomous de novo drug development depends on successfully bridging computational data science with wet-lab experimental biology and recruiting specialized technical talent into enterprise biopharma.

精炼阅读摘要

赛诺菲（Sanofi）生物制剂研究团队指出，随着生物医药迈入多特异性与多域工程分子的“第四波创新”，传统的经验性湿实验筛选面临严重的组合爆炸瓶颈；大型药企必须依托蛋白质语言模型、结构多模态表征以及主动学习算法，推动药物研发由实验主导转向以计算模拟（in silico）为先的工业化范式。 显微成像革命与高维生物数据爆发 (2:01) 生物学观测手段在过去一个半世纪经历了深刻范式转变。150年前，现代遗传学奠基人孟德尔仅能通过测量豌豆植株高度、豆荚大小和豆粒数量等宏观物理表型，推导基本的遗传定律；而当今生命科学领域正处于一场“分辨率革命”之中。现代高分辨率显微镜结合荧光分子标记技术，使研究人员能够实时跟踪活细胞尺度的动态交互过程。 在具体的抗肿瘤研究中，活体成像已能清晰记录免疫细胞（如T细胞或巨噬细胞）主动探查癌细胞表面的过程；一旦识别出恶性表型，免疫细胞会迅速穿孔癌细胞膜并注入毒性颗粒。这类超高精细度动态观测不仅直观呈现了微观药理机制，更在单细胞层面产生了空前规模的多模态、高维度生物学数据，为下游深度学习模型的训练提供了可靠的基础数据资产。 第四代生物制药的分子工程化挑战 (8:25) 当今生物制药创新正处于第四次技术浪潮，研发焦点从小分子化药与基础单克隆抗体，大幅转向高度复杂、具备上下文感知能力的工程化智能生物制剂。赛诺菲曾发表于《Nature》封面的三特异性抗体展示了这一复杂性：分子拥有三个不同的功能臂，其中两个功能臂分别结合肿瘤抗原和免疫效应细胞（如T细胞或NK细胞）建立物理桥梁，第三个功能臂则递送局部共刺激信号，持续激活免疫细胞防止其发生免疫耗竭。 前沿管线还涵盖了能够搭载治疗分子跨越血脑屏障的工程化载体，以及利用肿瘤微环境酸性特征定向激活的pH敏感型分子。设计这类由多个功能模块组装而成的分子机器（如四价纳米抗体）时，研发人员必须在结合亲和力、工程表达滴度、胶体与热稳定性以及体内半衰期等多维参数之间寻求全局最优解。传统物理文库的线性筛选机制不仅耗时极长，而且难以在广阔的排列组合空间中定位兼具药理活性与成药性的候选分子。 蛋白质语言模型与虚拟替代筛选管线 (15:35) 为了突破多参数优化的计算瓶颈，赛诺菲采取了“即时增强在研管线”与“构建端到端全计算发现”并行的双轨战略。由于蛋白质本质上是由20种天然氨基酸排列构成的线性生物聚合物，基于Transformer架构的自监督语言模型能够直接迁移至蛋白质序列处理中。正如通用大语言模型通过预测下一个Token习得语法和语义，蛋白质语言模型（pLM）通过上下文氨基酸预测，能够自主捕捉蛋白质的进化保守性、生物物理规律以及构象适应度。 在此基础上，研发团队将蛋白质一级序列信息与三维分子表面拓扑几何共同映射至连续的隐空间嵌入向量中。这些自监督表征驱动了一系列下游的虚拟替代模型，用以计算预测结合亲和力、表达产量以及储存稳定性等关键理化属性。通过将多个代理模型整合成高通量智能体虚拟筛选引擎，算法能够在庞大的组合序列空间中快速排序高潜力突变体，在计算环境中直接替代传统流程中长达数月乃至数年的湿实验筛选环节。 主动学习与三维结构融入的闭环优化 (18:36) 面对大分子药物的巨大设计空间，穷举表达上万个蛋白质突变体会造成不可承受的实验成本。为此，赛诺菲引入了主动学习循环架构。该系统并不依赖盲目的大规模湿实验合成，而是通过算法自适应评估不确定性与潜在收益，定向挑选数量有限、信息密度极高的小规模蛋白质子集进行表达验证，并将实验数据迅速反馈至算法引擎以完成参数微调与迭代。 为了处理高度复杂的非天然工程构架，研发团队在单纯的一级序列语言模型基础上显式注入了三维空间结构约束。这种将序列基础模型与立体构象上下文紧密结合的多模态建模策略，不仅显著提高了对突变体理化特性的预测保真度，还使得算法具备了可靠的从头生成（de novo）能力，能够针对高难度生物靶点直接输出并优化满足多重物理化学约束的先导候选分子。 制药AI工业化与多元分子递送拓展 (19:20) 演讲者指出，人工智能在生物制药产业中的角色已经发生根本性位移，彻底从边缘的探索性概念验证项目演变为决定药企未来核心竞争力的战略基础设施。赛诺菲正在引入严格的软件工程规范，将机器学习研究代码重构成高可靠性、可复用且高度自动化的工业级药物设计平台。该范式转变在组织架构上同样得到印证：拥有近20年计算与多组学经验、未曾操作过移液枪的计算科学家已被任命执掌拥有庞大湿实验人员的生物制剂研发组织，全面推进向“计算模拟为主导”的工作流转型。 此外，基础模型与生成式管线的应用范围正迅速外溢至抗体工程以外的前沿技术领域。赛诺菲目前已将专用语言模型全面铺开至mRNA序列本身的二级结构与翻译效率优化，并进一步应用于脂质纳米颗粒（LNP）递送载体配方的理性设计，显示出计算范式对下一代核酸与偶联药物工程的全方位渗透。 赛诺菲研发团队总结认为，虽然完全由计算机自主驱动的端到端从头药物设计此前被视作理论构想，但随着底层算法与高分辨率生物数据的突破，这一路径在工程上已具备高度现实可行性。然而，该计算范式能否全面实现临床转化，仍有赖于制药工程与现代软件开发规范的深度结合，以及药企能否持续打破传统湿实验与数据科学之间的组织壁垒并吸引顶级跨学科技术人才。

Detailed notes — zh

## Background 该视频由 Ai4 官方频道于 2025 年 8 月 26 日发布在 YouTube 平台。演讲由赛诺菲（Sanofi）临床前计算创新策略全球负责人 Maria Wendt 主讲，主要介绍了赛诺菲如何通过 AI、蛋白质语言模型与虚拟筛选技术加速大分子生物药研发，推动下一代智能化疗法的创新。 ## TL;DR 演讲首先回顾了生物学从孟德尔豌豆实验迈向高分辨率活细胞成像的数据爆炸历程，并阐述了第四波生物制药浪潮中多特异性“智能药物”面临的工程化挑战。面对复杂的成药属性优化，赛诺菲制定了从加速现有管线到纯算法从头设计的双轨 AI 战略，力求将传统湿实验分析全面转化为硅基化预测模型。团队通过蛋白质语言模型将自然语言预测机制映射为蛋白质工程突变设计，并结合三维空间拓扑表征。最终，依托主动学习闭环与软件工业化思维，AI 已成为跨国药企推动大分子、mRNA 及 LNP 研发的核心底层驱动力。 ## Atmosphere & style 演讲整体基调专业干练且充满极客热情，Maria 借由自身从纯计算科学家到统领湿实验团队的跨界经历，生动拉近了前沿算法与生物医学的距离。现场演示的微观免疫杀伤影像与结构设计图极具视觉冲击力，叙事节奏紧凑有力。演讲既有对传统药企转型痛点的务实剖析，又洋溢着以计算科学重塑生命健康的开创性与使命感，给听众带来强烈的行业变革代入感。 ## Key takeaways - [02:01](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=121s) 现代生命科学正在经历“分辨率革命”，活细胞超高分辨率显微成像等技术让数据维度从宏观表型跃迁到微观分子动态，为 AI 的深度介入奠定了数据基石。 - [04:33](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=273s) 传统依赖试错的早期药物研发模式正面临深刻转型，赛诺菲等药企正积极拥抱由计算与 AI 驱动的现代化药物发现新范式。 - [08:26](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=506s) 生物医药行业正阔步迈入第四波创新浪潮，研发重点从被动分子转向设计具备跨越血脑屏障、肿瘤微环境特异性激活等复杂功能的主动型“智能药物”。 - [11:08](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=668s) 三特异性抗体和四价纳米抗体等复杂生物药需兼顾药效活性、可生产性、结构稳定性和免疫效应功能等多维成药性指标，传统湿实验筛选难度极大。 - [13:35](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=815s) 赛诺菲制定了双轨 AI 战略：短期全面赋能并加速现有管线，长期则致力于实现端到端由纯计算驱动的全新生成式分子创制模式。 - [14:51](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=891s) 通过自主研发蛋白质大语言模型与表征技术，药企正逐步实现湿实验分析的“全面硅基化”，构建高通量虚拟筛选及智能体流水线。 - [17:03](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1023s) 大语言模型的下一词预测机制在蛋白质工程中直接映射为氨基酸突变建议与分子生成，并可融合一维序列与三维表面形态特征。 - [18:18](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1098s) 引入“主动学习”闭环机制，只需测试算法精选的小样本即可动态迭代模型，取代了动辄上万个突变蛋白的盲目湿实验试错。 - [18:18](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1098s) 大型跨国药企正借鉴软件工业理念推动算法资产的工程化与产品化落地，并将 AI 能力进一步延伸至 mRNA 设计及 LNP 递送系统等前沿技术领域。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=4s) AI 赋能医疗与生命科学的广阔前景 演讲开场展示了一张色彩斑斓的幻灯片，画面中生动呈现了一个正在与细胞进行“对话”的真实抗体分子。赛诺菲（Sanofi）临床前计算创新策略全球负责人 Maria Wendt 借此引出本次分享的核心主题：利用人工智能进行药物研发与蛋白质分子设计。她首先与台下听众互动，询问有多少人听了当天上午 Jeff 的主题演讲。她打趣道，早上听到关于“AI 可能会终结人类”的开场论调确实让人有点沮丧，但令人欣慰的是，Jeff 在演讲尾声着重强调了 AI 彻底变革医疗健康的巨大潜力。 Maria 坦言，在这场汇聚了约 8,000 人的大型 AI 盛会上，涉及医疗健康和生命科学领域的专题演讲却屈指可数。她真诚地希望在未来几年内——而且是就在接下来的短短几年而非漫长的岁月里——这种局面能够迎来彻底改变。她希望通过赛诺菲在前沿领域的实际探索，向大家清晰揭示 AI 在生物医药领域所承载的巨大承诺与落地价值。 ### [02:01](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=121s) 从孟德尔豌豆实验到高分辨率活细胞成像：生物学的数据革命 回顾生物学的发展历程，150 年前被誉为现代遗传学之父的格雷戈尔·孟德尔（Gregor Mendel）通过开创性的豌豆杂交实验奠定了基因遗传的概念。在那个年代，生物学家所能记录和分析的数据形式极其原始，仅仅是豌豆植株的高度、豆荚的形态大小以及每个豆荚中豌豆的颗粒数量。 ```mermaid flowchart LR accTitle: 生物学研究演进与数据爆炸 subgraph Past["150 年前：孟德尔经典遗传学"] A["观察宏观表型<br/>(株高/豆荚大小/豌豆数量)"] end subgraph Present["现代生物学：微观动态与高维数据"] B["高分辨率活细胞显微成像"] --> C["实时追踪免疫细胞与肿瘤交互"] C["实时追踪免疫细胞与肿瘤交互"] --> D["多组学高通量海量数据爆发"] end Past --> Present ``` 快进到今天，生物学的研究范式发生了翻天覆地的跃迁。Maria 展示了一段极具震撼力的微观视频：免疫系统中的三个 T 细胞（或巨噬细胞）正在严密探查并“嗅探”中间的一个肿瘤细胞，动态判断它是健康细胞还是病变细胞，进而决定是放行还是予以诛灭。通过现代超高分辨率显微镜技术与分子标记手段，研究人员可以亲眼目睹这一过程——T 细胞一旦识别出癌细胞，便会精准打孔，将红色的细胞毒性分子直接注入其中，将其彻底炸毁瓦解。现代生物学家触手可及的数据维度已从宏观表象飞跃到微观分子动态，这场席卷生命科学的“分辨率革命（Resolution Revolution）”带来了前所未有的海量数据，也为 AI 的深度介入构筑了坚实的基石。 ### [04:33](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=273s) 早期药物研发全景与计算科学家的实验室转型 制药企业的早期研发通常涵盖三个关键阶段：首先是深入解析疾病生物学机制以发现并验证治疗靶点（Target Identification）；其次是根据靶点设计并创造出候选药物分子（Drug Invention）；最后则是临床转化（Clinical Translation），确保该分子在人体中兼具安全性和临床疗效。 ```mermaid flowchart TD accTitle: 药物早期研发全流程 subgraph R["早期研究阶段 (Research)"] S1["疾病生物学探索<br/>& 靶点发现"] --> S2["药物分子发明<br/>(Drug Invention)"] S2 --> S3["临床转化验证<br/>(安全与药效评估)"] end S3 --> D["后续临床开发阶段<br/>(Development)"] ``` 谈及药物发明的核心环节，Maria 分享了自己的职业成长经历。她早年接受的是化学工程专业训练，原本的职业路径是管理生物反应器、操作纯化色谱柱或在石化厂提炼柴油，但她很早就发现自己对数学与计算怀有极致的热爱。在 2000 年代初（她笑称这暴露了自身年龄），当时的 AI 主要是专家系统和单层感知机（即当年的“神经网络”）。凭借对数学的执着，她随后投身生物信息学，在欧洲制药重镇瑞士巴塞尔从事了 18 年的纯计算干实验（*in silico*）工作，专注于基因组学、转录组学、蛋白质组学及代谢组学软件与算法分析工具的开发。 最令人惊叹的转折发生在加入赛诺菲之后：一个“18 年来从未握过移液枪”的计算科学家，被任命为赛诺菲生物制品研究部门（Biologics Research）负责人，统领由分子生物学家、蛋白质化学家和细胞科学家组成的庞大物理湿实验团队。这一人事任命凸显了现代制药行业的深层变革——传统依赖试错的湿实验模式正迫切需要升级，全面拥抱以计算与 AI 驱动的现代化药物发现新范式。 ### [08:26](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=506s) 第四波生物制药创新浪潮：三特异性抗体与“智能药物”的诞生 Maria 展示了一项登上《自然》（Nature）杂志封面的突破性成果——由赛诺菲自主工程化设计的创新型三特异性抗体（Tri-specific Antibody）。与阿司匹林等传统小分子化学药相比，抗体等生物大分子拥有更大的结构体量，能够执行复杂的物理桥接任务： - **第一/第二结合臂**：精准拉近肿瘤细胞与免疫细胞（如自然杀伤细胞 NK 或 T 细胞）之间的物理距离，形成免疫突触； - **创新的第三结合臂**：赋予其激活和共刺激信号，持续向免疫细胞发出“不要疲惫、保持活力、持续开火直至消灭肿瘤细胞”的指令。 当前生物医药行业正阔步迈入“第四波生物制药创新浪潮”，即由被动分子转向设计主动执行复杂任务的“智能药物”（Smart Medicines）。这类下一代分子不再仅仅是无目的地在体内游弋、寄希望于随机结合来阻断某条信号通路，而是具备了前所未有的工程化智能特性： * **跨越生物屏障**：能够像“搭便车”一样穿透血脑屏障（BBB），将药效精准递送至大脑内部的深层靶点； * **微环境条件激活**：在全身血液循环中保持惰性，只有游弋到肿瘤微环境并感知到局部 pH 值等生理特征发生特异性变化时才会被触发激活，从而在最大化杀伤肿瘤的同时，大幅降低对正常组织的脱靶毒性，显著提高用药安全性。 Maria 强调，这种集多靶点识别、物理构象重组与环境响应于一体的复杂大分子设计，正是当下生物医药创新的最前沿，也是计算模拟与 AI 大显身手的绝佳舞台。 ### [11:08](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=668s) 复杂生物大分子的模块化组装与工程化挑战 生物治疗药物的形态与尺寸正在经历爆炸式演进，药物研发早已超越了单一的天然蛋白质范畴。现代生物药不仅涵盖蛋白质药物，还深度融合了 mRNA 疗法与基因药物，研发团队甚至将不同模态直接物理偶联，以实现针对人体病灶极其特异的作用机制。这种复杂性在多特异性分子中尤为明显，例如一种四价纳米抗体（tetravalent nanobody）：由于纳米抗体体积小巧，单个分子上可以集成四个不同的结合位点——其中两个位点用于锁定癌细胞表面的双重靶标，第三个位点结合人血清蛋白以大幅延长药物在体内的循环半衰期，第四个位点则用于招募并激活 T 细胞。 制造这类多功能分子犹如在微观尺度组装一台精密复杂的机械设备。在传统研发路径中，每一个构建模块（building block）都需要从庞大的分子库中进行旷日持久的湿实验筛选，期望在海量测试中为每个模块挑出一个“胜出者”，最后再将它们拼接成最终的治疗分子结构。这一过程极为艰难，且研发人员必须同时兼顾多项严苛属性： - **药效活性（Potency）**：确保分子能够精确触发预期的生物学反应。 - **可生产性与表达量（Manufacturability & Expression）**：保证工程菌或细胞株能够以可控、经济的产量表达该分子。 - **结构稳定性（Stability）**：分子不仅在冷链储存时不能降解聚集，进入人体循环后更要承受复杂生理环境，绝不能自行解体脱靶。 - **效应功能（Effector Functions）**：精准调动人体免疫系统的其他协同机制以增强杀伤效能。 ```mermaid flowchart TD accTitle: 复杂多特异性生物大分子的多模块设计与属性优化 subgraph S1["四价纳米抗体构建模块"] A1["肿瘤靶点结合位点 1"] A2["肿瘤靶点结合位点 2"] A3["人血清白蛋白结合位点\n(延长体内半衰期)"] A4["T 细胞招募位点\n(免疫激活)"] end subgraph S2["多目标成药性优化"] B["组装成完整多特异性生物药"] C1["超高药效活性\n(Potency)"] C2["高表达与易生产\n(Expression)"] C3["体内存贮双重稳定\n(Stability)"] C4["免疫效应功能\n(Effector Functions)"] end A1 --> B A2 --> B A3 --> B A4 --> B B --> C1 B --> C2 B --> C3 B --> C4 ``` --- ### [13:35](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=815s) 从流程补充到纯计算设计：赛诺菲的双轨 AI 战略 面对成药空间与分子复杂度的几何级数增长，计算与人工智能已经成为生物制药不可或缺的底层支撑。制药行业已经没有时间继续对 AI 保持观望，赛诺菲为此制定了兼顾当下与未来的双轨计算战略。第一步是对现有的全部在研项目进行 AI 赋能，利用算法全面加速与优化现有管线；但赛诺菲的目标绝不仅止步于“辅助现有流程”，而是要追求范式层面的彻底革新。 这种变革性的未来指向完全由纯计算驱动的分子创制模式。在过去，如果有人宣称“只要在计算机里设计好分子结构，转眼间（voila）一款新药就能诞生”，同行们几个月前大概还会对此一笑了之；但在当下，计算能力的跃升与生物模型的突破，已经让这种“端到端纯算法设计药物”的前景变得完全可信且触手可及。 - **双轨推进路径**：短期内对所有运行中的发现管线进行计算补充；长期则瞄准完全计算化、生成式的全新药物发现流程。 - **行业观念转变**：计算生物学从过去的辅助验证工具，演变为定义药物发现范式的核心驱动引擎。 --- ### [14:51](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=891s) 构建蛋白质表征模型与湿实验的“全面硅基化” 尽管科技、金融和传媒巨头自主研发 AI 模型已成常态，但拥有超过十万名员工、历来以生物学家为主导的传统跨国大型药企（Big Pharma），如今也正在迅速组建高水平的自研计算团队。在赛诺菲，团队自主研发了大语言模型等基础架构，因为充分发挥 AI 潜力的首要前提，是找到用数学语言表征生物大分子的最优解法。 通过有监督学习与自监督预训练，AI 能够学习蛋白质深层的数学向量嵌入（embeddings），并将这些表征用于下游各类成药性预测任务。传统湿实验中，科学家需要花费数天乃至数年时间在实验室中逐一测试候选分子的结合活性、表达量和热稳定性；而赛诺菲的核心目标，是为实验室中每一项湿实验分析（assay）都构建出对应的“硅基化（in silico）预测模型”。这些独立模型进而串联成统一的高通量虚拟筛选系统，并正在向更加自主协同的智能体系统（agentic systems）演进。 ```mermaid flowchart LR accTitle: 蛋白质基础表征与下游硅基化筛选流水线 subgraph G1["基础模型预训练"] A["蛋白质序列/结构数据"] --> B["自监督 / 有监督学习"] B --> C["高维向量表征\n(Embeddings)"] end subgraph G2["下游属性预测 (In Silico Assays)"] C --> D1["结合亲和力预测"] C --> D2["重组表达量预测"] C --> D3["结构稳定性预测"] end subgraph G3["智能筛选体系"] D1 & D2 & D3 --> E["虚拟筛选与设计体系\n(Agentic Virtual Screening)"] end ``` --- ### [17:03](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1023s) 蛋白质语言模型：将“预测下一个词”转化为蛋白质工程 蛋白质药物不仅在临床上彻底改写了众多重度难治性疾病的治疗格局，在商业上也构成了全球范围内营收规模最高的一类药物形式。这类分子的天然优势在于，它们在本质上均可被编码为由氨基酸字母构成的线性序列。因此，蛋白质天然契合大语言模型（LLM）的底层架构体系。 在大语言模型中，核心机制是根据上下文预测下一个单词；而在蛋白质语言模型（pLM）中，预测序列中的下一个氨基酸字母，实质上就构成了精准的“蛋白质工程（Protein Engineering）”。模型通过计算输出更有利于分子成药的氨基酸突变建议，从而直接生成效能更高、成药潜力更大的候选分子。除了单纯的一维序列，研发团队还将表征维度延伸至三维蛋白质表面形态（surface representations），使得算法能够同时感知分子的空间拓扑特征与化学理化环境。 - **语言模型与蛋白质工程的同构性**：词表预测（Next-token prediction）直接映射为蛋白质序列的最佳突变选择与序列生成。 - **多维度表征融合**：实现从一维一级序列字母向三维空间表面构型特征的跨越。 --- ### [18:18](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1098s) 主动学习迭代与制药工业的软件化工程革命 当面对极具挑战性、设计空间极其庞大的复杂多特异性蛋白质时，单纯依赖实验室穷举或单纯依靠湿实验工程专家的经验均难以应对。为了攻关这类复杂设计，赛诺菲引入了主动学习（Active Learning）机制：研发团队无需在实验室里盲目重组表达 10,000 个突变蛋白，而是仅表达并测试经过算法精选的一小部分样本，将真实的实验数据快速回流至模型中进行动态迭代更新。 结合语言模型对蛋白质序列的深度理解，以及引入结构特征后的强大表达能力，研发团队能够在极其复杂的蛋白质候选空间中精准设计或筛选出最优分子。在大型药企内部，AI 的应用性质已经彻底告别了早期的边缘探索阶段，转而被提升为关乎药企未来生死存亡的战略核心（"seen as really survival in the future"）。制药企业正全面借鉴软件工业的核心理念，将算法模型资产进行真正的工业化（industrialization）与产品化（productizing）落地。该项能力目前已横向拓展至 mRNA 设计以及脂质纳米颗粒（LNP）递送系统等前沿技术领域，亟需更多计算科学领域的顶尖人才共同推动生物制药的数字化转型。 - **降本增效的主动学习循环**：以算法小样本筛选替代万级蛋白湿实验表达测试，显著缩短优化周期。 - **制药业软件化转型**：将 AI 研发从试验性课题转变为高度工程化、产品化的药物研发底层基础设施。 - **跨模态技术拓展**：模型架构全面向 mRNA 序列优化与 LNP 载体设计等前沿生物技术迁移。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=4s) 早上听到关于‘AI 可能会终结人类’的开场论调确实让人有点沮丧，但令人欣慰的是，Jeff 在演讲尾声着重强调了 AI 彻底变革医疗健康的巨大潜力。 — Maria Wendt > [02:01](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=121s) 这场席卷生命科学的‘分辨率革命（Resolution Revolution）’带来了前所未有的海量数据，也为 AI 的深度介入构筑了坚实的基石。 — Maria Wendt > [04:33](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=273s) 18 年来从未握过移液枪 — Maria Wendt > [08:26](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=506s) 不要疲惫、保持活力、持续开火直至消灭肿瘤细胞 — Maria Wendt > [08:26](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=506s) 第四波生物制药创新浪潮 — Maria Wendt > [13:35](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=815s) 只要在计算机里设计好分子结构，转眼间（voila）一款新药就能诞生 — Maria Wendt > [17:03](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1023s) 在大语言模型中，核心机制是根据上下文预测下一个单词；而在蛋白质语言模型（pLM）中，预测序列中的下一个氨基酸字母，实质上就构成了精准的‘蛋白质工程’。 — Maria Wendt > [18:18](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1098s) seen as really survival in the future — Maria Wendt > [18:18](https://www.youtube.com/watch?v=OsTSYLbXRzM&t=1098s) 在大型药企内部，AI 的应用性质已经彻底告别了早期的边缘探索阶段，转而被提升为关乎药企未来生死存亡的战略核心。 — Maria Wendt ## Mentioned Maria Wendt, 赛诺菲 (Sanofi), Jeff, 格雷戈尔·孟德尔 (Gregor Mendel), 《自然》(Nature), 三特异性抗体 (Tri-specific Antibody), 四价纳米抗体 (Tetravalent Nanobody), 蛋白质语言模型 (pLM), mRNA, 脂质纳米颗粒 (LNP), 血脑屏障 (BBB), 瑞士巴塞尔 ## Video description Maria Wendt, VP, Global Head of Preclinical Computational Innovation Strategy, Sanofi Abstracts What if we could revolutionize the way we design and develop biologics? In this talk, I will explore the cutting-edge intersection of artificial intelligence and biopharmaceutical innovation. I'll share how we're leveraging AI to accelerate large molecule research, reduce cycle times, and improve success rates in creating next-generation therapeutics. You'll learn about our comprehensive Digital Biologics Strategy, including sophisticated protein language models and virtual screening techniques for complex biologics. Join me to discover how AI is transforming the future of medicine, enabling us to create smarter, more functional biologics that address previously undruggable targets and complex diseases. [Source](https://www.youtube.com/watch?v=OsTSYLbXRzM)

Source quotes

[0:04] Maria Wendt: 早上听到关于‘AI 可能会终结人类’的开场论调确实让人有点沮丧，但令人欣慰的是，Jeff 在演讲尾声着重强调了 AI 彻底变革医疗健康的巨大潜力。 [2:01] Maria Wendt: 这场席卷生命科学的‘分辨率革命（Resolution Revolution）’带来了前所未有的海量数据，也为 AI 的深度介入构筑了坚实的基石。 [4:33] Maria Wendt: 18 年来从未握过移液枪 [8:26] Maria Wendt: 不要疲惫、保持活力、持续开火直至消灭肿瘤细胞 [8:26] Maria Wendt: 第四波生物制药创新浪潮 [13:35] Maria Wendt: 只要在计算机里设计好分子结构，转眼间（voila）一款新药就能诞生 [17:03] Maria Wendt: 在大语言模型中，核心机制是根据上下文预测下一个单词；而在蛋白质语言模型（pLM）中，预测序列中的下一个氨基酸字母，实质上就构成了精准的‘蛋白质工程’。 [18:18] Maria Wendt: seen as really survival in the future [18:18] Maria Wendt: 在大型药企内部，AI 的应用性质已经彻底告别了早期的边缘探索阶段，转而被提升为关乎药企未来生死存亡的战略核心。

Original transcript

[0:04] I think this is uh we're going to be [0:05] taking a different track from the [0:07] previous talks today because as you can [0:09] tell there's something maybe very [0:11] colorful on my slide, right? Um and [0:14] that's an actual antibbody [0:17] that's um well trying to talk to a cell. [0:21] So this is really going to be about um [0:23] drug discovery. It's about using AI for [0:26] designing drugs. Um, which reminds me, [0:29] um, so how many of you attended the [0:31] keynote this morning by Jeff? [0:34] Great, great, great. So happy to see [0:36] this. And how many of you, um, are in [0:39] healthcare or the life sciences? Wow, I [0:43] love this. Um, of course I wish that [0:46] there were a couple of hundred of us [0:47] here right now, you know, because, you [0:49] know, this this morning, you know, that [0:51] first message about, you know, AI ending [0:53] humanity was a little bit uh kind of a [0:56] downer to start the day, but I was so [0:59] happy that he ended, you know, that [1:01] session talking about the promise really [1:03] of AI for changing healthcare. Um, and [1:07] that's really a very very meaningful [1:09] application of AI and that is what I'm [1:11] going to talk about today. [1:13] Um and I'm so happy that there are in [1:16] fact uh people here who are into it [1:19] because with the breadth of this [1:21] conference you know there's like what [1:23] about 8,000 people it's quite surprising [1:26] how few presentations there are you know [1:30] touching healthcare and also touching [1:31] the life sciences. So I really really [1:34] hope this changes you know in the next [1:36] few years and really few years not many [1:38] years. Um and this is why um I really [1:42] love to share some of these things with [1:44] you. Um for those of you are also [1:46] thinking or wondering about what is this [1:49] talk about the big big promise of AI. Um [1:53] I hope that I can illuminate some of [1:55] those with you. All right. So let's go [1:59] to the first slide. So you're probably [2:01] wondering uh why am I showing you a [2:03] picture of a pee and a peep pod? Um it's [2:06] really to just share you know just [2:09] looking back 150 years ago um this was [2:13] how the study of biology was. Um this is [2:16] Gregor Mandel he's considered the father [2:19] of modern biology. Um if you recall he [2:22] was the guy that uh found out about the [2:25] concept of a gene you know that [2:27] something can be inherited and passed on [2:29] from one being to another. And this was [2:32] from 150 years ago. A very very seinal [2:34] experiment. And guess what? His data [2:37] type was the size of a pea plant, the [2:40] size of a peep pad, the number of peas [2:42] in a peep pod. Fast forward today. How [2:46] do we study biology today? Um, and this [2:49] is one example. I absolutely love this [2:52] image. It's actually three different T- [2:54] cells, you know, some cells in our [2:56] immune system that's basically sniffing [2:58] out a cancer cell. And that's the one in [3:00] the middle. And what you're seeing up [3:02] there, it's an actual video um of a T-C [3:05] cell or I think it's a macroofage. I'm [3:07] not sure which one. Um actively sniffing [3:10] a cancer cell trying to determine [3:13] whether he is this a healthy cell or a [3:16] sick cell. What should I do with this [3:18] afterward? Should I kill it or should I [3:19] let it live? And this is what's [3:22] happening in our body over and over. And [3:25] it's quite amazing that today, you know, [3:27] we can almost, well, not almost, well, [3:29] almost, we can almost see it, right? [3:32] Because to see is to believe. Um, and [3:35] what we're seeing here, what this means [3:37] is, you know, we now have all of these [3:39] beautiful highde microscopy, [3:42] all sorts of ways of labeling molecules. [3:45] Uh, like for example, those red dots are [3:47] actually toxic agents. And at some [3:49] point, what that T- cell will do is to [3:51] basically decide this is a a cancer [3:53] cell. I need to kill it. I'll perforate [3:56] it and then I'll inject those red [3:58] molecules into it to literally blast off [4:01] the cancer cell. And that is what can be [4:04] seen under a microscope. [4:06] But you can imagine the type of data [4:08] that it is collecting and the type of [4:11] data that biologists now have at their [4:13] fingertips. So this is really the major [4:16] major driver and our hope you know for [4:21] AI that there's really a big resolution [4:24] revolution happening in the study of [4:26] biology. There's an unprecedented amount [4:28] of data and therefore also an [4:30] unprecedented opportunities for AI. [4:33] Um and for I guess u most of well all of [4:38] the pharmaceutical uh companies uh they [4:41] would typically undergo these steps in [4:43] research. Um and this is only research. [4:45] This is not yet the development parts [4:47] you know because it's usually R&D. Um [4:49] but you know we don't have so much time. [4:51] I think I only have a few more minutes. [4:53] Um so you I think most of you know [4:57] especially those who are um working in [4:59] healthcare there's the first initial [5:01] work of really trying to understand that [5:03] disease right trying to understand it [5:06] biology trying to identify what could be [5:08] the good targets. Um and then [5:10] afterwards, hey, now let's create the [5:13] medicine for that. And that's what is [5:16] that middle part, the drug invention. [5:18] And that's what I'll be um spending most [5:21] of my time right now in the next few [5:23] minutes. Um but after that magical [5:26] molecule, that magical drug is created, [5:29] it does need to now go into the human [5:32] and that's what we call here as clinical [5:34] translation. How do we ensure um that [5:37] that particular molecule would actually [5:40] translate into the clinic would be [5:42] actually applicable i.e. efficacious and [5:45] safe for human use. [5:49] So as I mentioned I'm going to talk [5:50] about um the middle part um because in [5:53] fact um just to let you know so I was [5:56] actually trained as a chemical engineer [5:59] um but I decided you know chemical [6:01] engineers these are the people who run [6:02] bioreactors you know run purification [6:05] columns you know who run also what um [6:09] chemical prochemical plants you know for [6:11] creating uh diesel fuels etc etc um I [6:17] realized early on. I really really loved [6:20] the math side of things. So, in fact, [6:23] and I'm going to be kind of giving you a [6:25] clue of my age. Um, this was in the what [6:28] early uh 2000s, you know, that was when [6:30] there was artificial intelligence, but [6:32] actually that was really only expert [6:34] systems. Sorry, quote unquote only. So, [6:37] expert systems and also I would say a [6:39] single layer perceptron and that's what [6:42] they would call in neural net back then. [6:44] Um but uh I then moved into [6:47] bioinformatics actually afterwards [6:49] because of my love for math and wanted [6:50] to go to the life sciences. Um I worked [6:53] in Basel and Switzerland which is you [6:55] know a hotbed of uh some pharmaceutical [6:57] companies in Europe. I was there for 18 [7:00] years but I was working in a very very [7:01] incilico company. Um and this is a [7:04] company that develops software and also [7:06] analysis tools to study genomics. and [7:09] transcrytoics, proteomics, metabonomics, [7:11] data. And after 18 years there, [7:15] I actually became the head of biologics [7:17] research um at Sanopy, which is to [7:19] oversee a very very physical [7:22] organization [7:24] um an organization, molecular [7:25] biologists, um protein chemists, you [7:28] know, cell scientists. Um and it to me [7:31] it was quite astounding that someone who [7:33] has never held a pipet and had an [7:35] 18-year incilico career um was given the [7:39] task of managing such um physical [7:42] organization. Um and that is really [7:45] because of the great need today to [7:47] change the way we do drug discovery to [7:50] really lift it up now and bring it into [7:53] the modern more incilico way of doing [7:56] things. So I just wanted to share that [7:58] with you because um I think uh this is I [8:02] think a strong message for all of us who [8:03] are in AI or in compute um that this is [8:07] really now a new era of looking at [8:09] science of looking at industries. It is [8:12] very very flexible [8:14] um AI and computational people are very [8:17] very very welcome to be in the labs [8:20] today. In fact it's a very very big need [8:23] today. So what I'm showing you here um [8:26] is in fact one of the latest types of [8:28] medicines um and this appeared in the [8:31] cover of nature. Um and this is one of [8:34] our molecules at SATHI um which is a [8:36] highly engineered antibbody and that's [8:38] what you're seeing there. This molecule [8:40] that's bridging a tumor cell with a [8:42] cancer cell. Um and this particular um [8:44] antibbody it's a protein molecule. So [8:47] it's quite a a relatively big well it's [8:49] a much much bigger molecule than let's [8:51] say the molecule behind the aspirin. Um [8:54] but being big that means it can also do [8:56] some physical things that really [8:58] connecting a tumor cell um with the [9:00] immune system like a natural killer tea [9:03] cell. And the other magical thing, well [9:06] the magical thing basically that we did [9:08] at Sanhi is that um we added a third arm [9:12] to this antibbody because normally [9:14] antibodies would only have two arms to [9:16] to um attach to two different drug [9:18] targets but we added a third arm that um [9:22] after the T- cell and the cancer cell [9:25] are brought together we wanted to give [9:27] it a signal that hey don't get tired [9:30] get get alive fire fire until you kill [9:34] the tumor cell. So these are just um one [9:37] of those really to me miraculous types [9:40] of medicines that are happening that are [9:42] getting created today. Um really [9:44] changing the landscape of uh medicine [9:47] for cancer patients in particular. You [9:49] know Jeff in fact specifically [9:51] highlighted oncology during his talk. Um [9:54] and this is really in fact because today [9:56] we really are in what we call the fourth [9:58] wave of biioharmaceutical innovation. [10:00] This is now the creation of smart [10:02] medicines. Sorry, I won't have time to [10:04] really uh talk you through that, but [10:06] please feel free to take a look um at [10:07] some of the articles around it. But what [10:10] you're seeing is that these are [10:11] medicines who are not just, you know, [10:13] happily flying around in your body and [10:16] hoping to latch on to something that [10:18] would stop a particular pathway. No, [10:20] these are now medicines that are doing [10:22] such smart things. like I said, bridging [10:24] the immune system to your sick cell or [10:26] for example um uh allowing you know to [10:31] piggy back from outside the brain to get [10:34] into the brain to get to the drug target [10:37] within the brain. So these are really [10:39] much much smarter molecules. Um and [10:42] there's also the cap capability that [10:44] let's say these um medicines they're [10:46] swimming in your body and only when it's [10:48] next to a tumor cell it senses that [10:51] there's a change in the pH then all of a [10:53] sudden it gets activated. Therefore, [10:56] this is going to become a much much [10:57] safer medicine. [11:00] And [11:02] I showed you, you know, what that we [11:04] call a tri-specific antibbody. There's [11:06] so much excitement with these new types [11:08] of mechanisms as possible that there's [11:10] actually so many other different shapes [11:13] and and sizes of these new molecules [11:15] today. So, you can just imagine that's [11:17] really an increasing increasing amount [11:19] of complexity. And we don't just stop at [11:22] protein types of medicines. We also have [11:25] mRNA based medicines and also genomic [11:28] medicines and sometimes we mix them all [11:31] together because sometimes we even [11:32] attach them with each other um to get [11:35] the type of mechanism the very very [11:36] specific type of mechanism we want to [11:38] get into a human patient. Um and so now [11:41] how do we make this and how does AI get [11:44] into this? Um so normally because [11:46] they're really complex things they would [11:48] usually be built from different building [11:51] blocks and these are for example uh an [11:54] example of a um tetravalent um [11:57] antibbody. So it's we call it a [11:59] nanobbody because it is much smaller um [12:02] and in this particular case you know [12:04] it's attaching to two things on the [12:06] cancer cell. It's attaching to a human [12:08] serum you know so that then it can last [12:11] longer circulating in the body. um and [12:14] it also attaches to another target. [12:16] Maybe it's going towards a T- cell. So [12:18] that's why you have four different [12:19] attachment points to it. So each of [12:22] these building blocks would then have to [12:25] either be created and tested in the lab [12:29] um which is very very very difficult. [12:32] You would normally have libraries and [12:35] libraries of molecules that you have to [12:36] test to hope that you can find one [12:38] winner for each one of those building [12:40] blocks. Um and afterwards once you have [12:42] the winners you need to put them all [12:44] together to form the final therapeutic [12:47] format like what you see here. So it can [12:50] really be very very uh complex. So this [12:52] is another example of a complex um [12:54] antibbody. It's really like creating a [12:56] very very complex machine. Um and there [13:00] are so many things that we have to uh um [13:03] make sure you know that number one that [13:06] it's going to be potent. That's what [13:07] you're seeing there. Also that you can [13:09] manufacture them. um uh not necessarily [13:13] easily but at least manageably. So [13:15] that's what you're seeing there is [13:16] expression and also that that they are [13:19] stable enough so that they can last when [13:21] they're stored but also that when it's [13:22] put into a human body it will just not [13:25] pull apart. Um and you also have what we [13:28] call functions. It's this um ability to [13:30] call the immune system uh for add for [13:33] added power. [13:35] So what's our strategy? So this is [13:40] basically um the many many ways uh where [13:43] we can incorporate computation AI today. [13:47] Um again this might be uh too technical [13:50] but for those of you who are into drug [13:51] discovery I'm sure you are recognizing a [13:53] few things here. Um but my main message [13:56] here actually too is that we cannot wait [13:59] any longer um not to use computation and [14:03] AI and that's why we said that hey for [14:06] all our running discovery processes [14:08] today let's complement them now with AI [14:12] but we are not happy with just [14:15] complementing our processes today we do [14:17] need to shoot for a completely [14:19] completely new way of creating these [14:22] medicines a very transformational way [14:25] and that transformational way is very [14:27] likely purely computational. Um and [14:30] think you know our imaginations of you [14:32] know some molecule just getting designed [14:34] by a computer and then voila we have a [14:37] medicine. So we probably would have [14:40] laughed at that um you know a few months [14:42] ago um but today this is to us entirely [14:47] entirely probable. [14:51] Um and so at Sanhi so now to share with [14:54] you exactly uh what we have developed um [14:57] in AI. So of course a lot of tech [15:00] companies, finance companies, what uh [15:02] media companies uh they we hear that a [15:05] lot. So they're developing a lot of [15:07] their own um AI methods. Um and so [15:10] you're wondering what about pharma [15:11] companies would they you know because [15:13] these are big humongous corporations [15:14] like Santa 100,000 people. Humongous [15:17] corporations have what? biologists, [15:19] let's say, not really computing uh [15:22] people. Uh but to let you know, yes, [15:25] today pharma companies now have really [15:28] really growing themes of computing um [15:31] capabilities. Um and in our case at Sani [15:35] we are able to create our own language [15:37] models for example because we realized [15:41] that um to maximize uh the use of AI we [15:46] need to find the very very best way of [15:48] representing our molecules [15:51] mathematically so that we can operate on [15:53] them more effectively and this is of [15:56] course the representations right um and [15:58] creating foundational models um and in [16:01] fact uh what you're seeing There would [16:03] be the typical way for those of you who [16:05] are not uh too much into drug discovery. [16:07] Um but basically to find um a good [16:10] representation for the proteins. Um and [16:12] this could be via some um supervised or [16:15] self-supervised learning. Um and finally [16:18] uh whatever would be the result of that. [16:20] Um the embeddings could then be used for [16:23] downstream tasks um of predicting some [16:27] of these properties that we would [16:28] normally have to test in the lab. you [16:30] know like for example does it bind does [16:32] it express is it stable each one of [16:34] those properties you know a scientist [16:36] would spend days testing in the lab [16:38] sometimes years um but what we're trying [16:41] to do is to create an incilico [16:43] equivalent for every single assay that [16:46] every single scientist would do in the [16:48] lab okay so you can imagine we would [16:51] have this whole collection of models [16:53] that would be threaded together into a v [16:55] a virtual screening system that could [16:58] now become really um composed of agentic [17:01] systems. But here um because now I have [17:03] to fast forward um are just some of [17:06] those uh uh really important concept of [17:11] um just like uh what we do with language [17:13] models to predict the next word. the [17:15] beauty of proteins. Um, and really these [17:18] proteins are right now, just to let you [17:19] know, at the highest um, revenue type of [17:23] medicines um, in the world because [17:26] they're effective in patients because [17:28] they're really changing the landscape of [17:30] the most difficult diseases. These [17:32] proteins happen to be expressable as [17:35] amino acid sequences that are real [17:37] letters um that can then be amunable [17:41] right to large language um model uh type [17:45] of architectures. Um and to us basically [17:49] predicting the next letter um of an [17:52] amino acid sequence you know within a [17:54] protein um is basically protein [17:57] engineering is trying to give you what [17:58] would be the next letter so that you can [18:00] have a more successful protein and [18:02] therefore potentially a more successful [18:04] medicine. It's really quite astounding [18:06] um this bridging um and for us we also [18:10] go beyond um the sequence but als but [18:13] really to go into uh surface [18:15] representations as well. So uh fast [18:18] forward um going back to the most uh [18:22] difficult formats. So what we are also [18:24] doing here because um really the [18:26] landscape of all the designs explode at [18:30] this point we need all the weapons that [18:32] we can get not only you know the [18:34] beautiful protein engineering expertise [18:36] of our scientists in the lab but truly [18:38] AI um so this is also what we do um and [18:42] that is instead of um having for example [18:44] to express like 10,000 different [18:47] proteins we would just express a small [18:49] subset of that um and do active learning [18:52] and of Of course, these proteins as I [18:55] showed you there, it's really really [18:56] very complex because we invested on [18:59] finding out the best way to represent [19:02] them using language models but also some [19:04] other magic um of getting in structure [19:07] into the language models. Um then we are [19:09] able to create uh beautiful models that [19:11] allow us to um design um and also if not [19:15] design pick the very very best uh for [19:17] these very very complex proteins. So [19:20] maybe um to close my talk, this is also [19:24] one important um aspect I wish to share [19:26] with you that in the pharma companies [19:28] it's no longer just exploratory use. Um [19:32] this is now seen as really survival you [19:36] know in the future um that pharma [19:38] companies are now just like at Santa Fe [19:40] embarking on real industrialization [19:44] um of these different models and really [19:46] productizing also these models. So the [19:49] concepts coming from the software [19:50] industry is permeating as well. And for [19:53] those of you who might be um interested [19:55] also in all the other things that we do [19:57] with language models for mRNA uh for um [20:01] uh lipid nano particles and all that um [20:04] it's really quite amazing um uh what [20:07] we're able to do today and we need more [20:10] um computation um experts to help us [20:14] this. So thanks a lot.