Building a High Performing Organization that Integrates AI to Unlock Transformation With Montai

Building a High Performing Organization that Integrates AI to Unlock Transformation With Montai — Ai4

Ai4

AI-assisted summary of Building a High Performing Organization that Integrates AI to Unlock Transformation With Montai — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 27, 2025, this presentation features Margo Georgiadis, Co-Founder & CEO of Montai Therapeutics. The description highlights her journey as a multi-time technology executive crossing into life sciences, sharing actionable insights on designing high-performing, agile organizations that integrate AI with deep domain expertise to accelerate breakthrough innovation.

## TL;DR Drawing on executive leadership experience across Google and Flagship Pioneering, Margo Georgiadis outlines how Montai Therapeutics leverages AI to develop oral small-molecule replacements for biologic therapies targeting chronic diseases. She explains how the company decodes a vast universe of human-grounded chemistry using multimodal AI foundation models, replacing traditional sequential development with integrated upfront profiling. Crucially, Georgiadis details how Montai bridges the cultural divide between machine learning generalists and specialized life scientists through in-person co-location, shared technical glossaries, unified cross-functional OKRs, and an 'expedition team' culture.

## Atmosphere & style The presentation carries the polished, authoritative rhythm of a seasoned executive keynote, opening with self-deprecating humor about standing between the audience and their evening cocktails before transitioning into a focused operational masterclass. Georgiadis speaks with clarity and conviction, demystifying high-tech buzzwords in favor of grounded organizational design principles. The atmosphere feels pragmatic, candid, and inspiring, conveying genuine respect for both computational and biological crafts while championing intentional human collaboration as the true engine of transformation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=4s) Organizational design, change management, and leadership are just as critical, if not more important, than the underlying technology when executing large-scale transformation. - [01:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=114s) Embracing a '10x vs. 10%' mindset bypasses incremental defensive skepticism and forces teams to fundamentally rethink assumptions to solve previously impossible challenges. - [01:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=114s) Life sciences faces unique structural headwinds for innovation, including extreme domain specialization, data silos, and a high-stakes development model averaging $1 billion per program with a 10% clinical success rate. - [04:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=268s) The convergence of cost-effective high-fidelity biological data, population-scale clinical datasets, and advanced AI foundation models defines the current era as the 'biology decade'. - [04:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=268s) Despite biologics dominating $65 billion of drug spending in immunology and inflammation, high costs and delivery constraints mean fewer than 10% of diagnosed patients gain access to them. - [06:31](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=391s) Montai leverages multimodal AI and human-grounded chemistry—expanding natural bioactive sources into a 1B+ drug-like molecule library—to discover oral small molecules targeting undruggable biologic pathways. - [09:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=594s) Replacing traditional sequential drug discovery with integrated Target Candidate Profiles (TCP) and live AI scenario modeling enables screening libraries down to 5,000–10,000 high-priority candidates for wet-lab validation. - [11:06](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=666s) Bridging the divide between exploratory machine learning generalists and risk-mitigating life sciences specialists requires physical co-location and positioning AI strictly as a force multiplier for drug developers. - [12:43](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=763s) Developing a shared technical glossary and holding mandatory cross-disciplinary Lunch and Learns breaks down domain silos and normalizes asking foundational questions. - [14:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=868s) High-performing cross-functional organizations should normalize productive friction between computational and laboratory scientists as the primary engine for innovation rather than avoiding tension. - [15:56](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=956s) Establishing unified, joint OKRs across machine learning and biology functions prevents siloed success metrics and aligns the entire team under shared accountability. - [17:53](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1073s) Navigating biological ambiguity requires cultivating an 'expedition team' culture, supported by structured communication training (H.E.A.R.) and peer recognition mechanisms.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=4s) Leadership, Transformation, and Crossing into Life Sciences

Opening with a humorous nod to the packed room—noting that she stands directly between the audience and their evening cocktails—Margo Georgiadis emphasizes that while technology conferences naturally center on technical breakthroughs, three decades of experience at the intersection of technology and transformation have taught her that success hinges just as heavily, if not more, on people, leadership, and managing organizational change. She draws directly on her prior executive career, including years at Google when it was primarily just Google Search. During that era, the team had to figure out how to scale novel technologies while simultaneously creating brand-new consumer and enterprise markets from scratch across platforms like YouTube, Google Shopping, and Google Chrome. Nearly every time, industry observers claimed those transformations were impossible, yet the teams systematically cracked the model.

After serving as CEO and board advisor across multiple tech and consumer companies, Georgiadis crossed over into life sciences to co-found Montai Therapeutics alongside Flagship Pioneering roughly five years ago. Stepping into biotech from the outside requires immense humility; she credits both her team at Montai and Flagship Pioneering with helping her grow into a translational executive capable of bridging the gap between high-velocity technology cultures and deep biological science. Across every industry she has touched, she observes four universal requirements for technology-driven transformation: establishing a principled, ambitious view of the future; identifying the right learning vector to maximize "knowledge return"; resetting operational rules rather than accepting legacy workflows; and deliberately building an integrated, collaborative culture.

- **Foundational background**: Executive roles spanning Google (Search, YouTube, Chrome, Shopping), multiple CEO positions, and co-founding Montai Therapeutics in venture creation partnership with Flagship Pioneering. - **The leadership thesis**: Organizational design and change leadership are equally or more heavily weighted than underlying technology in determining whether an enterprise succeeds at transformation.

---

[01:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=114s) Four Principles of Transformation and the Realities of Biotech

To drive genuine transformation rather than incrementalism, organizations must embrace a "10x versus 10%" mindset. Georgiadis explains that when leaders propose a mere 10% efficiency improvement, teams immediately fixate on all the procedural obstacles and reasons why change will fail; conversely, framing a challenge around achieving a 10x leap forces people to imagine doing things that were previously impossible. She illustrates the second principle—defining a clear learning vector—using YouTube's early evolution. YouTube fundamentally overturned traditional media distribution by putting the consumer in control, learning over time that video consumption was not confined to traditional "lean-back" living-room moments, but stretched across everyday micro-occasions like standing in a lunch line, thereby unleashing an entirely new creator economy.

Applying these principles to life sciences, however, presents unique organizational hurdles that make biotech far harder to transform than early-adopting tech sectors. In early tech industries, abundant data and existential urgency drove rapid adaptation. In contrast, life sciences contends with deeply entrenched data silos, regulatory box-checking mentalities, and extreme functional specialization—where a researcher might dedicate 20 years to studying a single cell type, making multi-factorial systems thinking uncomfortable. Furthermore, drug development decisions are unusually high-stakes and capital-intensive: bringing a single drug forward typically demands roughly $1 billion with only a 10% probability of clinical success. These economics naturally breed risk aversion, requiring deliberate cultural adaptation to enable AI integration.

```mermaid flowchart TD accTitle: Core Transformation Principles vs Life Sciences Obstacles subgraph Principles["Universal Transformation Principles"] P1["10x Mindset<br/>(Reject 10% incrementalism)"] P2["Learning Vectors<br/>(Maximize knowledge return)"] P3["Resetting Rules<br/>(Overhaul legacy workflows)"] P4["Integrated Culture<br/>(Cross-functional alignment)"] end

subgraph Challenges["Life Sciences Headwinds"] C1["Data Silos & Fragmented Access"] C2["Extreme Skills Specialization"] C3["Regulatory Box-Checking Habits"] C4["High Stakes: ~$1B Cost & 10% Odds"] end

Principles -->|Overcomes| Challenges ```

- **10x vs. 10% philosophy**: Incremental targets trigger defensive skepticism; transformational targets liberate teams to rethink foundational assumptions. - **Knowledge return in practice**: YouTube proved how discovering non-traditional consumer use cases (e.g., the lunch line) unlocked massive scale and democratized content creation. - **Biopharma development headwinds**: $1 billion average capital deployment per program against a ~10% clinical success rate creates structural resistance to altering established research methods.

---

[04:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=268s) The Biology Decade and the Chronic Disease Tsunami

Despite these systemic frictions, Georgiadis underscores that the current era is definitively "the biology decade" due to the convergence of three powerful technological macro-trends. First, scaled, high-fidelity biological data has become dramatically more cost-effective through advances in genomics, single-cell sequencing, continuous monitoring wearables, and ultra-high-throughput screening. Second, massive population-scale clinical datasets now enable researchers to map real-world clinical outcomes, identify novel disease biomarkers, and accurately predict disease progression to pinpoint exactly when and how to intervene. Third, modern cloud computing, specialized hardware, and AI foundation models provide the computational horsepower necessary to process these petabytes of multimodal data and uncover patterns that were previously invisible.

Montai Therapeutics was built to direct this technological convergence toward what Georgiadis terms the "chronic disease tsunami." Chronic conditions represent a multi-trillion-dollar societal and economic burden, with six in ten Americans suffering from at least one chronic illness and four in ten living with two or more. Within the immunology and inflammation (I&I) space—spanning conditions such as rheumatoid arthritis, multiple sclerosis (MS), ulcerative colitis, inflammatory bowel disease (IBD), psoriasis, and atopic dermatitis (AD)—the top ten therapeutic drugs are all biologics, consuming $65 billion of the $175 billion spent on medications in the category. Yet despite their clinical potency in validating disease-modifying biological pathways, these biologic drugs reach less than 50% of the most severe patient populations and fewer than 10% of all diagnosed patients, leaving the vast majority without access to disease-modifying therapies.

- **Key epidemiological metrics**: 60% of US adults live with at least one chronic illness; 40% have multimorbidity (two or more conditions). - **Economic & access disparity**: In I&I, top biologics account for $65 billion out of $175 billion in total drug spending, yet fewer than 10% of diagnosed patients ever gain access to them due to cost, delivery constraints, or access barriers.

---

[06:31](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=391s) Montai's 10x Approach: Decoding Human-Grounded Chemistry with AI

Faced with these severe access limitations, Montai set out on a 10x mission: discovering oral, drug-like small molecules that can target the exact same high-value biological pathways validated by biologics. While prior AI biopharma initiatives over the past decade concentrated heavily on accelerating known structural biology pipelines, biologic target pathways frequently involve proteins that lack known, druggable binding pockets. Rather than attempting incremental structural fixes, Montai re-examined natural products—a chemical realm explored in the 1970s and 1980s but largely abandoned by the pharmaceutical industry due to issues with polypharmacology, chemical complexity, and manufacturing bottlenecks.

Montai developed a discovery platform centered on a human-grounded chemical universe: small molecules sourced from food, herbs, and traditional medicines that have safely inhabited human bodies for thousands of years. The company computationally expanded this foundation into a curated library of over one billion drug-like starting points by preserving bioactive natural cores while derivatizing their chemical structures to ensure high drug-likeness and synthesis within three or fewer retrosynthetic steps. By collaborating closely with NVIDIA to deploy multimodal AI foundation models, Montai created a scaffold-agnostic decoding engine that precision-ranks the top 1% of this billion-molecule space across diverse therapeutic modalities—including covalent binders, molecular glues, and macrocycles—to target historically "undruggable" disease pathways.

```mermaid flowchart LR accTitle: Montai Therapeutics Discovery and Decoding Architecture subgraph Sourcing["Human-Grounded Universe"] A["Food, Herbs &<br/>Traditional Medicine"] --> B["1B+ Derivatized Starting Points<br/>(≤3 Retrosynthetic Steps)"] end

subgraph AI_Decoding["Multimodal AI Engine"] B --> C["NVIDIA-Powered<br/>Foundation Models"] C --> D["Scaffold-Agnostic<br/>Decoding Engine"] end

subgraph Modalities["Precision Multi-Modality Targeting"] D --> E1["Covalent Binders"] D --> E2["Molecular Glues"] D --> E3["Macrocycles"] E1 & E2 & E3 --> F["Top 1% Precision Ranked<br/>for Uncorrelated Pathways"] end ```

- **Targeting biologic pathways**: Developing small-molecule therapies against proven biologic targets that lack classical, well-defined binding pockets. - **Curated chemical space**: Over 1 billion human-grounded, derivatized molecules engineered for synthesis in three or fewer retrosynthetic steps. - **Modality-agnostic search**: Universal search architecture decoding nature's chemistry across molecular glues, macrocycles, and covalent binders rather than siloing around a single chemical modality. - **NVIDIA partnership**: Utilizing multimodal foundation models to map intricate structure-function relationships and precision-rank the top 1% of drug candidates for biological activity, safety, and drug-likeness.

[09:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=594s) Transforming Drug Discovery from Sequential Stages to Integrated AI Modeling

Rather than following the traditional, highly sequential steps of pharmaceutical development, Montai Therapeutics re-engineered the process into a multi-factorial, integrated problem-solving framework. The workflow begins upfront with a comprehensive Target Candidate Profile (TCP), aligning the entire team on exactly "what good looks like" from day one. By evaluating downstream liabilities and safety profiles before wet-lab chemistry begins, the team avoids the pitfalls of discovering disqualifying flaws years into the pipeline. Supporting this approach required engineering extensive AI tools capable of running real-time scenarios, allowing team members to evaluate compound selection live during collaborative sessions.

Testing every potential molecule in a chemical library is economically and practically impossible. To make screening feasible, Montai utilizes machine learning to triage vast candidate spaces down to the top 5,000 to 10,000 highest-potential molecules for wet-lab synthesis and testing. To validate that this platform is repeatable rather than a one-off success, the company tackled 10 of the most challenging, well-studied biological targets to develop biologic replacements, generating compounding "network learning" across the system before broadening into partnered oncology programs.

```mermaid flowchart TD accTitle: Integrated Drug Discovery Workflow accDescription: Flowchart illustrating the transition from massive molecule libraries to wet lab synthesis via AI prioritization and Target Candidate Profiles.

subgraph Evaluation ["1. Integrated Evaluation"] TCP["Define Target Candidate Profile<br/>& Upfront Liabilities"] Sim["Run Live AI Scenario Models"] TCP --> Sim end

subgraph Selection ["2. AI Chemotype Prioritization"] Pool["Billions of Candidate Molecules"] Filter["AI Triage & Ranking Engine"] Top["Select Top 5,000–10,000 Molecules"] Pool --> Filter Sim --> Filter Filter --> Top end

subgraph Execution ["3. Validation & Scaling"] WetLab["Wet Lab Synthesis & Testing"] Targets["10 High-Difficulty Targets<br/>(Biologic Replacements)"] Top --> WetLab WetLab --> Targets end ```

- Synthesizing and testing billions of molecules would cost billions of dollars and take years; AI winnows the pool down to 5,000–10,000 priority molecules. - Montai demonstrated platform repeatability across 10 difficult biologic replacement targets, capturing shared network learning across projects.

---

[11:06](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=666s) Bridging the Divide: Machine Learning Generalists vs. Life Science Specialists

Aligning machine learning practitioners with life sciences researchers proved to be one of the most formidable operational hurdles. Machine learning scientists function largely as generalists—roughly 60% to 70% can be cross-trained and deployed across varied technical domains—approaching challenges as multi-factorial, directional, and exploratory. In contrast, biologists and medicinal chemists are deeply specialized domain experts accustomed to linear, sequential, and highly precise risk-mitigation methodologies. Furthermore, biologists and chemists often do not even share the same technical vocabulary.

Overcoming these cultural and operational differences required abandoning early assumptions about distributed work. Acknowledging that having machine learning staff on the West Coast while wet-lab scientists worked elsewhere was "never going to work," the company mandated that teams sit face-to-face in the same physical office. Leaders framed AI strictly as a force multiplier for drug developers rather than an end in itself: "The AI does not matter at all. What matters is how do we help drug developers overcome their obstacles."

| Characteristic | Machine Learning Scientists | Life Sciences Specialists | |:--- |:--- |:--- | | **Skill Breadth** | Generalists (60–70% cross-trainable across domains) | Deeply specialized, narrow domain experts | | **Problem-Solving Style** | Multi-factorial, integrated, directional | Sequential, stepwise, highly precise | | **Operational Mindset** | Exploratory, rapid scenario testing | Risk-mitigating, empirical verification |

---

[12:43](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=763s) Establishing Shared Vocabulary, Expedition Teams, and Continuous Learning

To build alignment between disparate disciplines, Montai instituted what it calls "great expedition behaviors," based on the principle of a mountaineering team where members possess distinct roles and specialized gear, but commit to leaving no one behind. Recognizing that cross-functional meetings often suffered from silent misunderstandings, the company dedicated an entire month to compiling an exhaustive internal glossary. This reference document defines every scientific and technical term used across the company, outlining its exact context and operational application to ensure total clarity and accelerate employee onboarding.

To foster psychological safety and break down domain silos, the organization instituted mandatory cross-functional "Lunch and Learns." Covering every major biological concept and AI technology deployed within the company, these sessions are recorded and maintain 100% cross-functional attendance. The initiative removed the stigma around asking foundational questions, normalizing the practice of raising one's hand to say "I don't know" and establishing mutual respect for each group's craft.

- The comprehensive company glossary took one month to build and resolved post-meeting misalignment across ML, biology, and chemistry. - Weekly cross-functional Lunch and Learns maintain 100% attendance across the company to build shared technical literacy.

---

[14:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=868s) Montai’s Core Values and Engineering Productive Friction

The name "Montai" reflects the company's core mission: climbing mountains with AI to overcome chronic disease. Across every wall and conference room, the company reinforces four foundational organizational behaviors designed to harness the creative friction between computational scientists and drug hunters:

* **Human-Driven Pioneering:** Emphasizing that transformative breakthroughs stem from people and collaboration rather than technology alone. * **Relentless Hunger for Impact:** Injecting urgency and rapid iteration into a traditional drug discovery mindset where "fast" is historically unfamiliar. * **Interdependently Ambitious:** Recognizing that neither computational models nor laboratory assays can succeed without equal, synchronized contribution from the other. * **Inclusive Courage:** Empowering every employee to challenge assumptions, speak up when a plan is flawed, and rigorously consult stakeholders inside and outside the company.

Rather than trying to eliminate tension between data scientists and wet-lab researchers, the culture seeks to "normalize the friction." The back-and-forth dialogue—such as a drug developer demanding specific biological parameters and an ML scientist countering with alternative molecular representations—is treated as the primary catalyst for genuine innovation.

---

[15:56](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=956s) Journey Mapping and Unifying Cross-Functional OKRs

To operationalize collaboration, Montai engineered an end-to-end "journey map" detailing every task in the discovery process, explicitly identifying decision-makers and performance metrics at each stage. While biological researchers initially pushed back against rapid quarterly Objectives and Key Results (OKRs)—arguing that "science is a journey" that cannot be neatly scheduled—the framework forced proactive problem-solving. When a wet-lab experiment stalled, machine learning teams immediately mobilized to identify alternative chemical series or target profiles.

A pivotal early mistake occurred when the organization set separate OKRs for the machine learning team and the biology team. This division caused friction, leading to scenarios where the computational team met its goals while the biology team fell behind, fostering mutual resentment. Montai eradicated this divide by enforcing single, unified OKRs that require joint accountability, ensuring that computational and biological teams succeed or fail together as one expedition.

- Discovery workflows are documented in a comprehensive Journey Map establishing task owners, decision rights, and baseline metrics. - Discarding siloed team goals in favor of unified, shared OKRs eliminated friction where machine learning succeeded independently of wet-lab outcomes.

---

[17:53](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1073s) Living in the Gray: HEAR Training and Monthly Compass Awards

Because life sciences operate in an environment filled with ambiguity, imperfect assays, and probabilistic answers, the organization trains its employees to comfortably "live in the gray." To navigate high-stakes technical disagreements without personal conflict, Montai implemented formal "H.E.A.R." training. This communication framework equips team members with structured methods to conduct respectful, productive debates when viewpoints diverge, preventing employees from retreating into defensive disciplinary corners.

To reinforce these collaborative values, the company hosts a monthly all-hands recognition ceremony featuring the "Great Expedition Behavior Award." Team members submit cross-functional peer nominations for colleagues who embody the company's shared principles. Winners are presented with physical brass compasses and have their portraits mounted on the company wall, serving as a continuous cultural touchpoint for collaborative excellence across AI and life sciences.

[19:08](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1148s) Fostering Cross-Functional Collaboration and Concluding as an Expedition Team

To actively reinforce peer collaboration across disparate fields, the company established a dedicated Slack channel called "You're awesome." Whenever a colleague from another function steps in to assist upstream, downstream, or anywhere along the task chain, team members post a shout-out to acknowledge their contribution. This simple mechanism provides an immediate "dopamine hit" that motivates employees to continuously support one another. Emphasizing organizational psychology, Margo Georgiadis points out that when distinct professions—such as AI technologists and life sciences experts—lack natural overlap, leaders must deliberately design systems and incentives that bring them together.

Reflecting on the company's journey, Georgiadis draws on the metaphor of scaling mountains, asserting that Montai could never have succeeded without deliberately designing the company around an "expedition team" mindset. Cultivating shared trust, cross-functional alignment, and mutual respect allowed the organization to bridge deep industry divides and drive breakthrough innovation. Expressing immense pride in her team's daily achievements, she concludes her presentation with gratitude to the audience.

- **"You're awesome" Slack Channel**: Serves as a continuous recognition loop for cross-functional support across all stages of the workflow. - **Organizational Psychology**: Intentional cultural practices are necessary to bridge distinct domains that historically have had little reason to collaborate closely. - **The Expedition Team Model**: Framing the workforce as an interconnected expedition team proved vital for navigating the complex intersection of AI and biotechnology.

## Notable quotes > [11:06](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=666s) The AI does not matter at all. What matters is how do we help drug developers overcome their obstacles. — Margo Georgiadis

> [01:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=114s) 10x versus 10% — Margo Georgiadis

> [04:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=268s) the biology decade — Margo Georgiadis

> [04:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=268s) chronic disease tsunami — Margo Georgiadis

> [09:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=594s) what good looks like — Margo Georgiadis

> [12:43](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=763s) great expedition behaviors — Margo Georgiadis

> [12:43](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=763s) I don't know — Margo Georgiadis

> [14:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=868s) normalize the friction — Margo Georgiadis

> [15:56](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=956s) science is a journey — Margo Georgiadis

> [17:53](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1073s) live in the gray — Margo Georgiadis

> [19:08](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1148s) You're awesome — Margo Georgiadis

## Mentioned Ai4, Flagship Pioneering, Google, Google Chrome, Google Search, Google Shopping, Margo Georgiadis, Montai Therapeutics, NVIDIA, Slack, YouTube

## Video description Margo Georgiadis, Co-Founder & CEO, Montai Therapeutics

Abstracts

AI enables us to make the impossible possible across industries. To fully harness its potential, we have to think differently about how we design and build organizations that integrate the best of AI and deep industry expertise to drive innovation. As a multi-time CEO in technology who crossed over into innovating in biotech with AI, Margo Georgiadis will share insights on what it takes to create a high performing organization at the intersection of AI and life sciences. By pioneering a collaborative model that harmonizes diverse skills, cultivates adaptive culture, and fosters strategic partnerships, the presentation details the journey of creating an agile organization where AI can accelerate breakthrough innovation.

[Source](https://www.youtube.com/watch?v=NlPm7PEZygQ)

中文简报

携手Montai：融合AI打造高效组织，释放转型潜力 结合在Google和Flagship Pioneering的高管领导经验，Margo Georgiadis概述了Montai Therapeutics如何利用AI开发针对慢性疾病的口服小分子替代药物，以取代生物疗法。她解释了该公司如何运用多模态AI基础模型解码庞大的基于人体的化学分子空间，以前期一体化特征分析替代传统的串行研发流程。至关重要的是，Georgiadis详细介绍了Montai如何通过线下同地办公、共享技术术语表、统一跨职能OKR以及“远征队”文化，弥合机器学习通才与专业生命科学家之间的文化隔阂。 在推动大规模技术变革时，组织架构设计、变革管理与领导力的重要性不亚于底层技术本身；确立“10x 而非 10%”的跃升思维能够打破增量防守心态，迫使团队重构假设以解决既往难题。 (0:04) 生命科学面临着深层数据壁垒、极端专业化分工，以及单个项目约 10 亿美元且临床成功率仅 10% 的高风险研发模式，但低成本高保真生物数据、人群规模临床数据集与 AI 大模型的结合，正开启崭新的“生物学十年”。 (1:54) 针对免疫与炎症领域生物药耗资 650 亿美元却仅惠及不足 10% 确诊患者的现状，Montai 结合多模态 AI 与天然生物活性来源，构建出包含超 10 亿药物级分子的化学库，致力于寻找靶向生物通路的口服小分子药物。 (4:28) Montai 摒弃传统线性串联流程，采用全流程靶标候选分子画像（TCP）与实时 AI 模拟，将数十亿分子初筛至 5000 至 10000 个高潜力分子进行湿实验验证，并在 10 个高难度靶点上验证了平台的可重复性。 (9:54) 为弥合机器学习通才与生命科学专家的文化隔阂，公司通过同地协同办公、统一跨职能 OKR、建立专业词汇表及“探险队”文化机制，将 AI 明确定义为药物研发人员的赋能杠杆而非目的本身。 (11:06)

Reading summary

Drawing on her background across tech and life sciences, Montai Therapeutics co-founder Margo Georgiadis argues that organizational design and cultural alignment are just as critical as technical capabilities when deploying AI to solve high-stakes biopharma challenges. Principles of 10x Transformation in Biotech (1:54) Georgiadis outlines four universal principles for technology-driven transformation: setting an ambitious vision, identifying a clear learning vector, rewriting operational rules, and fostering an integrated culture. In contrast to incremental 10% efficiency goals that trigger skepticism, a 10x mindset forces teams to rethink foundational assumptions. While early tech industries adapted quickly due to abundant data and urgency, biotechnology faces deeply entrenched structural hurdles, including siloed data, extreme functional specialization, and regulatory box-checking. Biopharma's risk aversion is reinforced by development economics, where advancing a single drug typically requires approximately $1 billion against a 10% clinical success rate. Overcoming this inertia requires shifting how scientific organizations evaluate failure, establish learning vectors, and manage knowledge return. Targeting Chronic Diseases with Natural Chemistry (4:28) Montai Therapeutics was founded to address what Georgiadis terms a "chronic disease tsunami," noting that 60% of US adults live with at least one chronic illness and 40% suffer from multimorbidity. In immunology and inflammation, the top 10 drugs are biologics, accounting for $65 billion of $175 billion in medication spending. However, because of cost, administration constraints, and access barriers, these biologics reach less than 50% of severe patients and fewer than 10% of all diagnosed individuals. To bridge this gap, Montai targets identical, validated biologic pathways using oral small molecules. The platform computationally screens a human-grounded chemical universe derived from food, herbs, and traditional medicine, which has been derivatized into over one billion starting points synthesizable within three or fewer retrosynthetic steps. Collaborating with NVIDIA, Montai uses multimodal AI models to precision-rank the top 1% of candidate molecules across diverse modalities—such as molecular glues, macrocycles, and covalent binders—to engage targets lacking traditional druggable binding pockets. Integrating Workflows and Prioritizing Compounds (9:54) Rather than following the pharma industry's traditional sequential discovery steps, Montai uses an integrated model centered on an upfront Target Candidate Profile (TCP) to align on standards before wet-lab chemistry begins. Evaluating downstream safety and liabilities early avoids uncovering fatal flaws years into development, supported by live scenario-modeling software. Because synthesizing every molecule is cost-prohibitive, machine learning models triage the billion-compound space down to the top 5,000 to 10,000 priority candidates for wet-lab synthesis and testing. Montai validated this repeatable framework across 10 difficult biologic replacement targets, generating network learning across programs before expanding into partnered oncology research. Bridging Cultures Between Technologists and Biologists (11:06) Integrating computational scientists with life sciences researchers presented substantial operational challenges. While roughly 60% to 70% of machine learning practitioners operate as cross-trainable generalists who favor rapid, directional exploration, biologists and medicinal chemists are narrow domain specialists trained in sequential, risk-mitigating methodologies. To bridge this divide, Montai abandoned distributed work to co-locate computational and wet-lab teams in the same office, framing AI purely as a force multiplier for drug developers. To align technical vocabularies, the company spent a month creating a comprehensive internal glossary and launched mandatory cross-functional "Lunch and Learns" with 100% attendance. To prevent cross-functional misalignment, Montai transitioned from separate functional targets to unified cross-functional OKRs, documented workflows through end-to-end journey maps, and instituted "H.E.A.R." communication training to normalize productive debate and help teams navigate probabilistic outcomes without conflict. Georgiadis concludes that deploying AI in biotechnology relies on deliberate organizational psychology and an "expedition team" mindset. Sustained collaboration mechanisms—such as unified incentives, peer recognition channels, and monthly awards—are essential to successfully merge computational and biological disciplines.

精炼阅读摘要

Margo Georgiadis 指出，在生命科学领域整合人工智能的核心挑战不仅在于算法突破，更在于组织设计与跨学科协同；唯有确立十倍跃升思维、重构线性流程并建立共同承担的考核机制，才能将计算能力转化为攻克慢性病的研发生产力。 以十倍思维突破生物制药研发阻力 (1:54) Georgiadis 强调，组织在推进变革时若仅提出 10% 的渐进式改良目标，团队容易将精力消耗在对既有流程限制的防御性解释中；相反，确立“10 倍跃升（10x）”目标能够迫使团队跳脱常规框架，探索全新可能。在互联网科技早期，海量数据与生存紧迫感驱动了组织的极速迭代，但生命科学领域存在截然不同的组织惯性：长期形成的数据孤岛、合规审查思维，以及极度的学科垂直分工——研究人员可能专注于单一细胞类型长达二十年，天然对多因素系统化思维产生抵触。 更为关键的是，生物医药研发面临着高风险与重资本投入的结构性压力。根据行业常规数据，推进单个候选药物的平均成本约为 10 亿美元，而临床成功率仅约 10%。这种极度不对称的经济风险加剧了研发团队对规避风险的执念，使得打破旧有研究范式变得格外艰难，必须自上而下重置运营规则。 解码天然化学空间以应对慢性病海啸 (4:28) 当前高保真生物数据成本骤降、大规模真实世界临床数据集积累，以及 AI 基础模型算力飞跃，共同构成了“生物学十年”的基础。Montai Therapeutics 致力于将这三股趋势导向应对“慢性病海啸”。流行病学数据显示，60% 的美国成年人患有至少一种慢性疾病，40% 的人群面临两种及以上共病。在免疫与炎症（I&I）领域，前十大药物均为生物制剂，占据了 1750 亿美元总用药支出中的 650 亿美元；然而，由于生产成本高昂和给药途径受限，这些生物药仅覆盖了不足 50% 的重症患者，在全部确诊患者中的覆盖率甚至低于 10%。 针对大量经过生物学验证却缺乏传统结合口袋的靶点，Montai 避开了对现有靶点结构的微调，重新审视数千年来在人体内具备安全性验证的天然产物（食物、草药与传统药物）。公司通过计算将这一化学宇宙衍生为一个包含超过 10 亿个药物样起始分子的数据库，并限制其逆合成路线在三步以内。借助与 NVIDIA 联合部署的多模态 AI 基础模型，该平台能够无视特定骨架限制，跨越分子胶、大环化合物与共价结合剂等多种形态，精准筛选出前 1% 的候选分子，以开发出靶向生物药通路的口服小分子替代疗法。 重构研发流程：前置评估与端到端筛选 (9:54) 传统的药物研发往往遵循前后割裂的线性阶段，Montai 则将其重组为集成的多因素问题求解框架。研发工作从最初即制定严谨的“目标候选物画像（Target Candidate Profile, TCP）”，在进入湿实验室之前全面评估下游开发风险与安全性指标，并利用 AI 工具进行实时情景模拟，避免数年后才暴露致命缺陷。 由于在湿实验中合成与测试数十亿个分子在成本和周期上并不现实，团队利用机器学习算法将庞大的化学库初筛至最具潜力的 5,000 到 10,000 个分子进入湿实验验证。为证明该计算筛选体系具备可复现性而非单一偶然成果，Montai 首先在 10 个极具挑战性且已被行业深入研究的高难度靶点上进行生物药替代分子的开发，通过沉淀跨项目的网络化经验，为后续拓展至肿瘤等合作管线奠定基础。 弥合算法人才与生命科学专家的认知断层 (11:06) 让机器学习从业者与生命科学研究人员协同工作是落地转型中最严峻的管理考验。机器学习人才普遍具有较强的通用性（约 60% 至 70% 的技术人员可跨领域迁移），倾向于采用多因素、探索性和方向性思维；而生物学家与药物化学家则是高度垂直的领域专家，习惯于线性、循序渐进且严格控制风险的严谨实验，两者之间甚至缺乏互通的技术语言。 为了消除文化与沟通壁垒，Montai 放弃了算法团队异地办公的设想，强制要求两类团队在同一物理办公室内面对面办公。管理层反复强调，AI 本身并非目的，其核心定位是帮助药物研发人员克服障碍的倍增器。为此，公司耗时整整一个月编写了一份详尽的内部术语词典，明确所有生物与技术术语的具体语境；同时设立全员 100% 参会的每周跨领域知识分享会（Lunch and Learns），营造包容且平等的学术交流环境，消除提出基础概念问题时的心理顾虑。 登山队机制：统一绩效考核与良性冲突 (15:56) Montai 建立了以“登山探险队（Expedition Team）”为核心的企业文化，强调团队成员分工不同但绝不抛弃任何同伴。公司在早期曾将机器学习团队与生物团队的季度目标（OKR）分开设定，结果出现了算法指标顺利达成而生物实验受阻、双方互相抱怨的局面；随后公司果断废除独立考核，推行跨职能的统一 OKR，迫使计算与实验团队休戚与共。 为将协作制度化，公司制定了涵盖研发全流程的任务与决策权旅程地图（Journey Map），并推行 H.E.A.R. 沟通框架培训，指导员工在面对模糊与不确定的实验数据时展开理性的专业辩论，实现“良性冲突的常态化”。此外，每月跨部门提名颁发的实体黄铜罗盘奖以及内部 Slack 频道的即时致谢机制，为跨领域日常协作提供了持续的正向反馈回路。 Georgiadis 总结认为，AI 在生物科技领域的成功转型本质上是组织心理学与管理制度的重塑；即便拥有前沿算法和海量化学数据，企业若不能在物理空间、考核利益及文化心理上实现算法通用人才与垂直领域专家的深度融合，依然无法跨越药物发现的复杂险阻。

Detailed notes — zh

## Background 本内容由 YouTube 频道 Ai4 于 2025 年 8 月 27 日发布，主讲人为 Montai Therapeutics 联合创始人兼 CEO Margo Georgiadis。视频主要探讨如何打破科技与生命科学的跨界壁垒，通过组织重塑、文化协同与 AI 技术的深度融合，打造高效敏捷的跨学科团队以加速突破性药物研发。 ## TL;DR Margo Georgiadis 结合其在 Google 等科技巨头及生物医药领域的跨界领导经验，分享了将 AI 深度融入生命科学组织的系统方法论。Montai Therapeutics 聚焦慢性病难题，利用多模态 AI 解码数十亿级天然化学分子库，颠覆了传统的线性研发模式。面对 AI 科学家与生物学家之间的认知断层与思维冲突，公司推行“伟大远征”文化，通过统一词汇表、共享 OKR、HEAR 沟通培训以及即时正向激励机制，成功将跨学科摩擦转化为驱动创新的核心动力。 ## Atmosphere & style 演讲风格干练自信、幽默且富有感染力，Margo 带着硅谷科技高管的敏锐视角与对生命科学的谦逊敬畏，娓娓道来跨界创业的实战洞见。从开场调侃“挡在大家与鸡尾酒会之间”，到剖析团队磨合踩过的“管理误区”，整体叙事真诚而接地气，既有宏大的产业愿景，又充满极具温度的组织心理学细节，让听众深刻体会到带领异构团队攀登科学险峰时的激情与坚韧。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=4s) 跨界转型的组织领导力四大法则：以第一性原理树立 10x 愿景、明确知识回报的学习向量、重塑执行规则，以及构建深度融合的跨学科文化。 - [03:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=184s) “生物学十年”三大交汇浪潮正在打破传统生命科学的数据孤岛与高壁垒：高保真生物数据规模化、人群临床数据深化，以及云计算与 AI 算力的跨越式突破。 - [05:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=328s) 攻克“慢性病海啸”的可及性困境：现有生物大分子药物成本高昂且覆盖不到 10% 确诊患者，Montai 致力于利用 AI 研发可普惠的突破性口服小分子替代疗法。 - [06:31](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=391s) 重塑天然产物化学空间：基于数千年人体耐受经验的天然分子，利用 AI 逆合成设计构建超 10 亿级易合成类药分子库，实现全模态通用搜索。 - [06:31](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=391s) 自研多模态 AI 解码引擎：联合 Nvidia 训练骨架无关的基础模型，直接解析化学底层语言，从数十亿分子中精准筛选排名前 1% 的高成药性候选物。 - [09:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=594s) 从线性研发到前置整合决策：全学科在研发起点共同制定目标候选物画像（Target Candidate Profile），用 AI 模拟推演将测试范围收敛至 5,000-10,000 个分子进行实测。 - [11:06](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=666s) 弥合机器学习通才与生物学专才的认知断层：明确 AI 的核心使命是消除药物研发限速瓶颈，坚持让不同背景专家物理共处、协同解题。 - [12:44](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=764s) 建立跨学科统一语境：编制《全公司词汇表》消除术语隔阂，并常态化开展跨职能午餐研讨会，营造敢于承认“我不知道”的心理安全氛围。 - [14:29](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=869s) 将跨学科摩擦常态化（Normalizing Friction）：践行以人为本、持续渴望影响力、相互依存与包容勇气四大准则，在推拉博弈中激发算法与生物学的创新火花。 - [15:57](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=957s) 打破部门孤岛的统一 OKR 机制：摒弃给 ML 和生物团队分别设标的做法，推行全流程旅程地图与全员共享 OKR，实验受阻时算法团队可迅速介入重构方案。 - [17:53](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1073s) 应对生命科学的“灰色地带”：引入 HEAR 沟通培训化解根本性分歧，设立月度全员“指南针罗盘大奖”持续强化登山远征队的文化认同。 - [19:08](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1148s) 运用组织心理学与即时正向激励：设立 Slack “You're Awesome” 频道即时肯定跨部门协作，以休戚与共的远征队精神支撑高风险交叉学科的深度变革。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=4s) 从科技巨头到生命科学：跨界转型的组织领导力与四大法则 在拥挤的会场中，演讲者 Margo Georgiadis 以幽默的开场调侃自己正“挡在大家与鸡尾酒会之间”。回顾自己长达 30 年跨越科技与业务转型的职业生涯，她强调技术固然是行业热议的焦点，但组织的构建方式以及领导团队应对变革的能力，对最终成功起到了同样重要、甚至权重更高的决定性作用。Margo 曾长期供职于 Google，经历了公司从单一的 Google Search 扩张并开创 YouTube、Google Shopping、Chrome 等全球级平台的黄金时代。在每一次开拓新市场时，外界几乎总是断言“那绝不可能实现”，但团队最终都找到了破局之道。在历任多家公司的 CEO、早期董事及顾问后，她跨界迈入了一个全新的领域——生命科学，并与 Flagship Pioneering 共同创立了 Montai Therapeutics。面对生命科学极高的行业壁垒，她带着极大的谦逊与敬意，转型成为连接科技与生物医药的跨界高管（translational executive）。 基于数十年在科技与转型交叉领域的实践，Margo 总结出四条放之四海而皆准的核心法则： 1. **树立基于第一性原理的未来愿景（Principled View of the Future）**：正如 Google 倡导的“10 倍（10x）而非 10% 的思考方式”。如果仅仅追求 10% 的微小改进，人们往往会找出无数个行不通的借口；唯有思考如何通过彻底变革去实现前所未有的突破，才能真正打开思路。 2. **明确知识回报的学习向量（Learning Vector for Knowledge Return）**：以 YouTube 为例，平台从根本上重新定义了内容的概念并赋权消费者。团队逐步洞察到，用户的观看场景远不止传统的客厅“后仰（leanback）时刻”，还包括午餐排队等海量碎片化时间，进而开创了全新的内容创作生态与消费模式。 3. **重塑行事规则（Reset the Rules for Execution）**：颠覆性的变革绝不可能沿用旧有的运作路径。 4. **构建深度融合的组织文化（Integrated Culture）**：将跨学科能力与敏捷文化有机结合，支撑颠覆性创新的落地。 --- ### [03:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=184s) 行业鸿沟：高壁垒生命科学与“生物学十年”的融合浪潮 谈及行业转型的难度，Margo 对比了数字化转型的“早期采纳行业”（如互联网与零售）与深水区的生命科学。前者天然拥有海量数据，且普遍怀有“不转型便灭亡”的强烈生存危机感；而生命科学领域则充斥着严峻的结构性挑战： - **数据孤岛与严重的数据访问限制**。 - **高度碎片化的专业分工**：许多科学家可能耗费 20 年时间深耕单一细胞类型，这种高度垂直的功能化视角让他们很难跳出来想象多因素（multifactorial）协同的全局世界。 - **严格的行业合规与监管心理**：从业者往往拘泥于机械式的“按部就班打钩（check certain boxes）”，从而在体制上抑制了求变。 - **高风险、多方参与的决策流程结合极高的资本密集度**：在生物医药领域，研发一款新药往往需要耗资 10 亿美元，而成功率仅有 10%。在每个关口都是高额赌注的情况下，人们天然抗拒变革。 尽管困难重重，Margo 坚信当前正是真正的“生物学十年（Biology Decade）”。三大关键趋势的交汇正在让曾经的“不可能”变为“可能”： ```mermaid flowchart TD accTitle: 生命科学三大交汇趋势 subgraph S1["1. 高保真生物数据规模化与低成本化"] A["基因组学 / 单细胞技术"] B["高通量筛选 (HTS)"] C["可穿戴设备与持续监测"] end subgraph S2["2. 人群全景数据深化"] D["临床结局追踪"] E["疾病生物标志物挖掘"] F["疾病进展细分与预测"] end subgraph S3["3. 云计算与算力跃升"] G["大规模 AI 基础模型"] H["从海量数据中破解未知模式"] end S1 --> I["驱动下一代突破性药物研发"] S2 --> I S3 --> I ``` - **高保真生物数据的规模化与成本效益大幅提升**：单细胞测序、基因组学、规模化高通量筛选以及可穿戴持续监测设备，捕获了前所未有的大尺度数据。 - **人群与临床数据的深度理解**：能够精准解析临床结局与疾病生物标志物，预测疾病进展阶段，从而明确最佳干预时机与方式。 - **云计算与 AI 算力突破**：使从海量复杂生物数据中解析从未被人类发现的底层模式成为现实。 --- ### [05:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=328s) 慢性病海啸与现有生物大分子药物的可及性困境 Montai Therapeutics 致力于攻克医疗领域最庞大的系统性挑战之一——“慢性病海啸（Chronic Disease Tsunami）”。慢性病已演变为波及全美乃至全球的数万亿美元级沉重负担： - **全美有 60% 的成年人（十分之六）** 患有至少一种慢性疾病。 - **40% 的成年人（十分之四）** 饱受两种或两种以上慢性病的折磨。 以炎症与自身免疫性疾病（I&I，包含类风湿性关节炎、多发性硬化症 MS、溃疡性结肠炎、炎症性肠病 IBD、银屑病、特应性皮炎 AD 等）为例，尽管生物科技在该领域取得了一定进展，但仍存在严重的结构性矛盾： | 行业指标 | 统计数据 / 现状 | |:--- |:--- | | **I&I 药物市场支出总额** | 约 1750 亿美元 | | **排名前 10 的生物大分子药物（Biologics）耗资** | 吞噬了高达 650 亿美元的医疗支出 | | **重症患者渗透率** | 仅覆盖不到 50% 最严重的重症患者 | | **确诊患者总体可及率** | **不到 10%** 的确诊患者能够获得这些前沿疗法 | “这是不可接受的（That's not okay）。” Margo 强调，科学已经明确证实这些生物学通路能够有效改变疾病进程，但现有生物大分子药物高昂的成本与给药限制将其锁死在极小的人群中。如何打破常规，寻找能够靶向并调节这些关键通路的全新口服小分子药物，成为 Montai 成立的核心使命。 --- ### [06:31](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=391s) Montai 的 10x 破局：从人类经验自然化学到多模态 AI 解码引擎 过去十年间，AI 在生物医药领域的应用主要集中在已知结构生物学方法的效率优化上。然而，许多已被生物大分子药物验证的关键疾病靶点，在结构上往往缺乏已知的结合口袋（binding pockets），传统工具根本无法“下手”。面对这一行业死结，Montai 确立了其 10x 愿景：能否攻克每一座看似无法逾越的高山，让靶向这些已验证通路的疗法实现普惠？ 当 Montai 决定重返天然产物（Natural Products）化学空间时，传统制药界的同仁几乎都认为他们“疯了”。在 1970 至 1980 年代，制药界曾探索过天然产物，但因其多药理学效应过于复杂（polypharmacology）且难以人工化学合成而基本放弃。Montai 采取了截然不同的策略：聚焦于数千年来已在人体内经历过生物考验的天然化学子集（源自食品、草本及传统药物），将其作为起始分子，并利用 AI 赋予其现代药物属性。 ```mermaid flowchart LR accTitle: Montai 药物发现与 AI 解码引擎流程 subgraph Prep["1. 底层化学空间构建"] A["人类经验天然分子<br/>(食品/草药/传统药物)"] --> B["衍生化与可合成性设计<br/>(≤3步逆合成路线)"] B --> C["10亿+ 类药起始分子库"] end subgraph AI["2. 基础模型与多模态解码"] C --> D["Nvidia 驱动的多模态 AI 引擎"] D --> E["骨架无关 (Scaffold-Agnostic)<br/>结构与功能关联解码"] end subgraph Screen["3. 精准预测与排序"] E --> F["精准筛选前 1% 优质分子<br/>(生物活性/类药性多维排序)"] F --> G["共价结合剂 / 分子胶 / 大环等<br/>多模态非相关骨架候选药"] end ``` - **构建超 10 亿级类药起始分子库**：从具备人体耐受历史的天然核心结构出发，通过精准衍生化设计，确保分子具备优异的类药性质，并且必须能在 3 步或更少的逆合成反应（retrosynthetic steps）内完成化学制造。 - **全模态通用搜索（Universal Search）**：不同于专攻单一模态（如共价结合剂、分子胶或大环化合物）的传统 Biotech 公司，Montai 认为大自然已囊括所有化学模态，平台可在全模态空间中跨越式寻找最佳通路匹配。 - **骨架无关（Scaffold-Agnostic）的全球顶尖解码引擎**：分子结构与功能之间的关联存在上百万种复杂模式。Montai 联合英伟达（Nvidia），利用多模态生物化学数据训练自研基础模型，直接解码大自然写在化学中的底层语言。 - **从 10 亿分子中精准筛选排名前 1% 的候选物**：系统同时对生物活性、成药性等多维度属性进行高精度排序，输出互不相关的多样化骨架分子，从源头极大提升药物研发的成功概率。 ### [09:54](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=594s) 颠覆传统药物研发：从线性链条到 AI 赋能的前置整合决策 传统的药物研发长期受制于线性的、按部就班的串联步骤，而 Montai Therapeutics 彻底打破了这一惯例。团队从项目伊始就直接确立目标候选物画像（Target Candidate Profile），让所有学科的专家在最前端共同界定究竟什么才算真正达标的“优秀成果（what good looks like）”。当团队挑选分子送入湿实验验证时，所有人已经预先将后续可能面临的所有潜在成药缺陷与风险（liabilities）考量在内。这种向“高度整合的问题解决流程”的蜕变，彻底颠覆了既往各自为战的研发范式。 为了支撑这一工作流，团队搭建了海量的 AI 工具，用于高通量模拟推演。研究人员能够在系统内实时提出各种假设性问题：“如果我们采用这种方案会怎样？如果要采购或合成这一批分子而不是另一批，又会产生什么影响？”毕竟面对天文数字般的化学空间，没有人能够真正合成并测试十亿量级的分子——那将耗资数十亿美元且耗费数十年光阴。AI 的核心价值在于精准锁定最具潜力的 5,000 到 10,000 个分子用于实体合成与检测。随后，团队再对其进行优先级排序，在化学型多样性（chemotype breadth）与分子特性之间做出权衡。为了证明该方法的可重复性，Montai 直接挑战了 10 个难度极高的靶点，这些靶点全部是旨在替代生物制剂（biologic replacements）的慢性病治疗靶点，由于都属于已被广泛研究的网络靶点（worked targets），这让团队积累了极其丰富的网络化学习数据，目前他们正在将这一能力向肿瘤学等更广泛的合作伙伴领域拓展。 ```mermaid flowchart TD accTitle: Montai Therapeutics 整合型 AI 药物研发工作流 accDescription: 展示从全学科制定目标画像、AI大规模场景模拟与分子筛选，再到湿实验验证与网络化学习的闭环流程 subgraph FrontEnd ["前置对齐与整合定义"] A["明确目标候选物画像 (Target Candidate Profile)"] B["全学科共同界定达标标准与潜在风险 (Liabilities)"] A --> B end subgraph AIScreening ["AI 推演与大规模分子初筛"] C["AI 场景模拟工具 (实时多维问答推演)"] D["从数十亿候选空间中锁定 5,000-10,000 个最优分子"] E["化学型多样性排序与实验决策"] B --> C C --> D D --> E end subgraph WetLabLoop ["湿实验验证与网络化迭代"] F["湿实验精准合成与功能检测"] G["积累网络化学习经验 (扩展至慢性病与肿瘤学)"] E --> F F --> G G -. 闭环反馈迭代.-> C end ``` - **十亿分子与万级精选**：彻底规避合成测试十亿分子的天价成本与漫长时间，AI 的介入将范围收敛至 5,000 至 10,000 个分子进行实测。 - **高难度靶点验证**：连续攻克 10 个极具挑战性的生物替代靶点，并通过跨项目的网络化学习（network learning）提升研发飞轮效率。 --- ### [11:06](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=666s) 跨学科的思维碰撞：机器学习专家与生物学家的认知断层 规划宏伟蓝图是一回事，把两拨底层思维完全不同的人揉捏成一支高效团队则是完全不同的另一回事。正如 Margo Georgiadis 直言：“搭建团队并接受摩擦究竟发生在何处，这绝非易事。”在团队构成上，机器学习科学家多属于通才（generalists），团队里有 60% 至 70% 的 AI 人员可以进行交叉培训并在任何环节工作；而另一侧的生命科学家则是高度专业化、钻研极深的纵向专精人才。在思维方式上，AI 科学家倾向于将问题看作多元且高度整合的系统（multifactorial and integrated），偏好方向性探索与容忍试错；生命科学专家则习惯于线性的、按部就班的推导，追求精确性并极力规避风险。 Margo 自嘲在创立公司之初曾抱有天真的幻想，一度打算将机器学习团队留在西海岸，将生物学家部署在东海岸。但她很快意识到“这绝不可能成功”——他们必须坐在同一个屋檐下，面对面注视着彼此的眼睛来共同解题。作为 AI 领袖，首要认知是认清技术的本质定位：“AI 本身根本无关紧要，真正重要的是我们如何帮助药物研发专家跨越阻碍他们前进的限速步骤。”当双方坐到一起时，现实甚至更为残酷：甚至只要同时找来四个生物学家，或者把生物学家与化学家放在一起，他们内部所使用的专业术语和行业语言都无法互通。学会如何“一起讲人类能懂的语言（learn to speak human together）”，成为了团队生存的第一要务。 --- ### [12:44](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=764s) 打造伟大远征行为：公司词汇表、午餐研讨会与团队根基 为了弥合鸿沟，Montai 从最底层的价值观与沟通习惯着手，确立了被称为“伟大远征行为准则（Great Expedition Behaviors）”的团队规范。攀登科学的高峰就像组织一支登山远征队：队员们各司其职、技能各异，但在向山顶冲刺的过程中，团队绝不会抛下任何一名队友，谁也不能在雪山上掉队甚至“丧命”。这一理念深深植根于企业文化之中——“Montai”一词的本意正是“运用 AI 攀登险峰，攻克慢性病难题”。 在实际执行中，消除沟通障碍的第一步是建立统一的语言体系。Margo 发现最初开会时，大家各说各话，开完会五分钟后就发现彼此根本没听懂对方的意思，毫无协同共识可言。于是她果断叫停，耗时整整一个月梳理并编写了一份详尽的《全公司词汇表（Company Glossary）》，明确定义了每一个专业术语在 Montai 内部的确切含义、应用场景及工作原理，这成为新员工入职培训的核心神器。与此同时，公司设立了常态化的跨职能“午餐学习会（Lunch and Learns）”，由各领域专家拆解讲解内部的每一项关键技术或生物学机制。这类研讨会全程录音录像，跨职能员工的出勤率达到了 100%。它为所有人创造了一个可以安心举手承认“我不知道”的安全氛围，彻底清除了因彼此知识壁垒而滋生的隐性隔阂。 --- ### [14:29](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=869s) 将文化刻在墙上：常态化摩擦与四大核心行为准则 在 Montai 的每一间会议室和走廊墙壁上，都张贴着醒目的远征主题海报，每一项理念下方都明确细化出约四条具体的落地行为规范。这些准则并非空洞的口号，而是旨在把两个学科交汇时必然产生的摩擦“常态化（normalizing the friction）”。因为在 Margo 看来，创新真正的灵感火花恰恰迸发于这种不可避免的摩擦之中——当药物研发人员提出“但我需要这样的特定效果”，而机器学习专家回应“如果从这个角度看呢？我可以给你提供这个，但作为交换必须那样做”。这种在推拉博弈中碰撞出的交锋，正是技术突破的温床。 ```mermaid flowchart LR accTitle: Montai 远征行为四大核心价值观 accDescription: 展示以人为本的开拓精神、对影响力的持续渴望、相互依存的雄心壮志与包容的勇气四大支柱 subgraph ExpeditionCore ["伟大远征四大核心准则"] HDP["以人为本的开拓精神<br/>(Human-driven Pioneering)"] RHI["对影响力的持续渴望<br/>(Relentless Hunger for Impact)"] IA["相互依存的雄心壮志<br/>(Interdependently Ambitious)"] IC["包容的勇气<br/>(Inclusive Courage)"] end ExpeditionCore --> Magic["将跨学科摩擦常态化<br/>激发生物与算法的创新交锋"] ``` - **以人为本的开拓精神（Human-driven Pioneering）**：解决世界级难题的核心依靠的是人，技术与算法只是赋能者的工具。 - **对影响力的持续渴望（Relentless Hunger for Impact）**：药物研发关乎重症病患，每一天都至关重要。传统生物学家往往无法理解“快（fast）”，但通过对紧迫感与社会影响力的重新定义，研发节奏被大幅激活。 - **相互依存的雄心壮志（Interdependently Ambitious）**：从第一天起全体成员就达成共识：没有任何单一学科能够在脱离对方的情况下独立取得成功。 - **包容的勇气（Inclusive Courage）**：鼓励任何人当发现流程不妥时敢于挺身举手质疑，同时保持对内外部意见的极度敏锐，确保不遗漏任何关键细节。 --- ### [15:57](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=957s) 流程全景与对齐目标：从全流程旅程地图到打破部门孤岛的统一 OKR 为了将远征计划落实为清晰的行动纲领，团队绘制了一份详尽的“全流程旅程地图（Journey Map）”。地图精细拆解了研发体系中的每一项任务，自上而下明确了决策权归属以及对应的量化评估指标。通过按季度运行 OKR（目标与关键结果法），整套流程每周、每月、每季度都在持续迭代精进。尽管生物医药界起初对此充满抵触——“科学是一场未知的探索，药物化学天然缺乏可预测性，实验随时可能失败”——但 OKR 的真正作用在于强制团队展开对话：一旦某项实验遇阻或耗时超预期，机器学习团队便能立刻介入，寻找更优的替代化学空间，从而把科研人员从传统思维定势中解放出来。 在这个过程中，公司曾踩过一个极大的管理误区：最初管理层分别给机器学习团队和生物学团队制定了独立的 OKR。这迅速导致了一种危险的对立情绪——当生物学进展卡壳时，机器学习团队觉得自己已经完成了指标，是在单方面向对方做出妥协让步。Margo 敏锐地意识到这一机制极为荒谬：“我们本来就是同一支登山远征队！”公司随即推倒重来，要求所有人围坐在一起，制定全公司唯一一套完全共享的 OKR，所有分支工作均向同一个终极目标对齐。这一看似微小的机制调整，彻底扭转了团队的对话氛围，消除了部门壁垒滋生的低效与隔阂。 --- ### [17:53](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1073s) 在灰色地带中共事：HEAR 沟通培训与象征方向的远征罗盘 生命科学的本质决定了研究者必须长期身处“灰色地带（living in the gray）”。生物学领域没有绝对完美的测试方案，也极少存在唯一确凿的标准答案。为了帮助不同背景的成员学会共同应对这种模糊性与不确定性，Montai 引入了专门的沟通方法论——“HEAR 沟通培训（HEAR Training）”。这项专门指导团队如何在持有根本性分歧时展开高质量对话的技术，教会员工如何在尊重对方的前提下探讨异见，通过富有建设性的辩论促成共识，避免因观点对立而各自退回孤立的阵营之中。 此外，为了让远征文化深入人心，公司在每月全员大会（Town Hall）上设立了“伟大远征行为大奖（Great Expedition Behavior Award）”。获奖的个人或团队完全由跨职能的跨部门同事自发提名产生，获胜者将获得一个货真价实的物理“指南针罗盘（Compass）”，其照片将被隆重张贴在公司的荣誉墙上。这种看似有些“老派甚至是鸡汤（cheesy）”的仪式感，却在全公司范围内持续释放出极具分量的信号：每一个月都在向全员实实在在地重申并强化——身处一支伟大的远征团队究竟意味着什么，以及同行伙伴对彼此卓越付出的由衷认可。 ### [19:08](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1148s) 运用组织心理学与正向激励打造跨领域远征团队 在团队协作机制中，微小的即时正向激励往往能产生巨大的推动力。Montai Therapeutics 在内部专门建立了一个名为“你太棒了”（You're Awesome）的 Slack 频道。每当有跨职能同事在任务链路的前端、后端或横向协作中提供帮助时，员工都会在频道里公开致谢发帖。这种公开认可不仅提供了即时的多巴胺奖励（"dopamine hit"），更能激发团队成员持续互助的意愿，让大家发自内心地觉得“我想多做一些这样的事”。管理者往往容易忽视“组织心理学”（psychology of organizations）的力量，但在生命科学与人工智能这种底层逻辑截然不同、本没有任何自然交集的专业领域之间，如果想要促成真正的深度协作，就必须主动构建并悉心经营这种彼此联结的土壤。 面对生物医药与 AI 交叉融合的险峻高山，整个公司就像一支共同攀登高峰的“远征团队”（expedition team）。Margo Georgiadis 坚信，如果当初没有采用这种打破界限、重构协作流程的组织建设方式，没有学会像探险队一样休戚与共、紧密协同，公司绝不可能在如此艰难且复杂的行业中取得突破。她在演讲的最后表达了对团队每一天辛勤付出的自豪，并对全场听众致以谢意。 - **Slack “You're Awesome” 频道机制**：在任务链的前沿、后方或侧向协作节点上，只要跨部门同事伸出援手即可公开点赞，通过即时多巴胺回馈建立正向行为循环。 - **组织心理学驱动融合**：跨学科人员在传统认知中缺乏天然协同纽带，企业必须有意识地通过制度设计与心理激励手段打破部门壁垒。 - **远征队式（Expedition Team）文化**：将企业跨越学科鸿沟的攻坚历程比作攀登险峰，依靠高度协同与信任共渡难关。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=4s) 10 倍（10x）而非 10% 的思考方式 — Margo Georgiadis > [03:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=184s) 按部就班打钩（check certain boxes） — Margo Georgiadis > [03:04](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=184s) 生物学十年（Biology Decade） — Margo Georgiadis > [05:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=328s) 慢性病海啸（Chronic Disease Tsunami） — Margo Georgiadis > [05:28](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=328s) 这是不可接受的（That's not okay） — Margo Georgiadis > [11:06](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=666s) 搭建团队并接受摩擦究竟发生在何处，这绝非易事 — Margo Georgiadis > [11:06](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=666s) AI 本身根本无关紧要，真正重要的是我们如何帮助药物研发专家跨越阻碍他们前进的限速步骤 — Margo Georgiadis > [11:06](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=666s) 一起讲人类能懂的语言（learn to speak human together） — Margo Georgiadis > [14:29](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=869s) 常态化摩擦（normalizing the friction） — Margo Georgiadis > [15:57](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=957s) 我们本来就是同一支登山远征队！ — Margo Georgiadis > [17:53](https://www.youtube.com/watch?v=NlPm7PEZygQ&t=1073s) 身处灰色地带（living in the gray） — Margo Georgiadis ## Mentioned Margo Georgiadis, Montai Therapeutics, Flagship Pioneering, Google, YouTube, Google Shopping, Chrome, Nvidia, Slack, Ai4 ## Video description Margo Georgiadis, Co-Founder & CEO, Montai Therapeutics Abstracts AI enables us to make the impossible possible across industries. To fully harness its potential, we have to think differently about how we design and build organizations that integrate the best of AI and deep industry expertise to drive innovation. As a multi-time CEO in technology who crossed over into innovating in biotech with AI, Margo Georgiadis will share insights on what it takes to create a high performing organization at the intersection of AI and life sciences. By pioneering a collaborative model that harmonizes diverse skills, cultivates adaptive culture, and fosters strategic partnerships, the presentation details the journey of creating an agile organization where AI can accelerate breakthrough innovation. [Source](https://www.youtube.com/watch?v=NlPm7PEZygQ)

Source quotes

[0:04] Margo Georgiadis: 10 倍（10x）而非 10% 的思考方式 [3:04] Margo Georgiadis: 按部就班打钩（check certain boxes） [3:04] Margo Georgiadis: 生物学十年（Biology Decade） [5:28] Margo Georgiadis: 慢性病海啸（Chronic Disease Tsunami） [5:28] Margo Georgiadis: 这是不可接受的（That's not okay） [11:06] Margo Georgiadis: 搭建团队并接受摩擦究竟发生在何处，这绝非易事 [11:06] Margo Georgiadis: AI 本身根本无关紧要，真正重要的是我们如何帮助药物研发专家跨越阻碍他们前进的限速步骤 [11:06] Margo Georgiadis: 一起讲人类能懂的语言（learn to speak human together） [14:29] Margo Georgiadis: 常态化摩擦（normalizing the friction） [15:57] Margo Georgiadis: 我们本来就是同一支登山远征队！ [17:53] Margo Georgiadis: 身处灰色地带（living in the gray）

Source quotes

[11:06] Margo Georgiadis: The AI does not matter at all. What matters is how do we help drug developers overcome their obstacles. [1:54] Margo Georgiadis: 10x versus 10% [4:28] Margo Georgiadis: the biology decade [4:28] Margo Georgiadis: chronic disease tsunami [9:54] Margo Georgiadis: what good looks like [12:43] Margo Georgiadis: great expedition behaviors [12:43] Margo Georgiadis: I don't know [14:28] Margo Georgiadis: normalize the friction [15:56] Margo Georgiadis: science is a journey [17:53] Margo Georgiadis: live in the gray [19:08] Margo Georgiadis: You're awesome

Original transcript

[0:04] Well, I'll endeavor to make this [0:06] interesting. It's a crowded room and I'm [0:08] standing between you and cocktails. So, [0:11] um, but I think, you know, we all come [0:14] here to talk about our technology, but [0:17] my experience after a 30-year career at [0:19] this intersection of technology and [0:21] transformation has taught me a lot about [0:23] the fact that the organizations and how [0:25] we lead them can have a big impact on [0:28] our success. In fact, I believe it's [0:30] equally if not higher weighted on the [0:32] people and how we lead each other [0:34] through change than just the technology. [0:37] And so my work that I'll talk about here [0:40] is the company I helped found um about [0:43] five years ago in life sciences, but it [0:46] draws on a lot of my own experiences [0:49] from a lot of other uh companies. So [0:52] after um some previous career I spent a [0:55] long time at Google when we were just [0:57] Google search and we created a lot of [0:59] different platforms YouTube shopping [1:02] chrome right all sorts of different [1:04] platforms that I'm sure you guys all use [1:06] and that taught me a lot about what does [1:08] it take in each one of the instances not [1:10] just to scale the technology but to [1:12] really create a new market and I would [1:14] say almost every time people said that [1:16] would be impossible and we somehow [1:18] figured it out and then I went on as a [1:20] CEO [1:21] to run a bunch of different companies or [1:24] was a an advisor an early board member [1:26] and a whole bunch of others before I got [1:28] to a whole new industry called life [1:31] sciences. How many people in this room [1:32] are actually in life sciences? [1:35] Right? So if you know that industry, you [1:38] know it's not an easy one to come from [1:39] outside in. And I'd had a lot of [1:42] experience, but I with great humility [1:44] really respect my team at Monzai and [1:47] Flagship Pioneering for helping me [1:49] become that translational executive and [1:52] build this company. And what I've [1:54] learned over the years is there's four [1:55] things that in every company I've ever [1:58] worked in at this intersection of [2:00] technology and transformation are true. [2:03] You have to create a principled view of [2:05] the future. We used to talk about it as [2:07] Google as thinking 10x versus 10%. So [2:10] for anyone integrating AI, if you're [2:12] just thinking about a 10% improvement, [2:13] people think about all the reasons why [2:15] it's not going to work. But if you think [2:16] about why it's going to transform and [2:18] let you do things you could never do [2:19] before, you will think differently. And [2:22] then you have to figure out what is that [2:23] learning vector? Where can I get [2:25] knowledge return? You know, think about [2:27] a business like YouTube where we really [2:30] fundamentally redefined what did content [2:32] mean and how do you put a consumer in [2:34] charge. And over time we learned that [2:36] people have many, many more occasions [2:38] and the leanback moment. They have the [2:40] lunch line. They have so many different [2:41] ways. And now we have so many more ways [2:44] of creating content and people that are [2:46] doing that. And so we've unleashed a [2:48] whole new way of putting people ahead. [2:51] You also then have to reset the rules [2:53] for how you get things done. Almost [2:54] never if you're really going to do [2:56] something transformational are you doing [2:57] it the way it was done. And then you [3:00] have to build an integrated culture that [3:01] lets you make that possible. [3:04] So I I was glad I started on the left [3:06] side of this page [3:08] which is all the early adopters because [3:11] these are industries where fundamentally [3:12] we had a lot of data and we had a sense [3:15] of urgency that if we didn't transform [3:17] right we would not exist and we know a [3:20] lot of of companies that were there and [3:22] didn't embrace the future. The right [3:24] hand side is where I work now and I'm [3:27] glad I worked in the left because it [3:28] created the humility if that was hard. [3:31] When you enter this arena where we have [3:33] a whole bunch of challenges, data access [3:36] and silos, high skills fragmentation, [3:39] right? When you work in an industry [3:40] where somebody might have studied one [3:42] cell type for 20 years and they're very [3:45] functionally [3:47] oriented, it's much harder for them to [3:49] imagine a world that's multiffactorial [3:51] or where you have regulation. People [3:54] believe they have to check certain boxes [3:55] in a certain way. And so that embeds [3:58] right us to not make change or where [4:02] decision is complex. They're high stakes [4:04] decisions with a lot of people involved. [4:06] And you couple that with capital [4:07] intensity when you look in biotech you [4:10] spend a billion dollars with a 10% odds [4:12] of success of a drug working. So these [4:15] are really high stakes decisions at each [4:17] moment. And so people don't necessarily [4:20] want to change and so we really have to [4:22] think about that adaptation process and [4:24] how we create an organization that can [4:26] make it happen. But the opportunity is [4:28] unbelievable in this sector. I truly [4:31] agree with everything that's written [4:33] that this is the biology decade. [4:36] When we integrate the three major trends [4:38] that are coming in life sciences, we can [4:41] make the impossible possible over the [4:43] next decade. When we have scaled [4:45] highfidelity biology data that's more [4:48] cost effective genomics, single cell, [4:51] right? Scaled high throughput screening, [4:54] wearables and continuous monitoring. We [4:55] couldn't have imagined the scale of data [4:57] we could capture at the same economics [5:00] population data. Now we can understand [5:03] clinical outcomes. We can understand [5:05] biomarkers of disease segment and [5:08] predict the progression of disease. So [5:10] we know better when to intervene and how [5:13] and cloud and comput and those tools are [5:16] essential for us to be able to take all [5:18] this data and make sense of it and find [5:21] new patterns that we never knew existed. [5:23] And so that's what makes it exciting to [5:25] be in this industry. Where we're focused [5:28] at Montai is on a big huge problem. We [5:30] call it the chronic disease tsunami. Six [5:32] in 10 Americans suffer from one chronic [5:35] disease. Four in 10 have two or more. [5:38] And it's a multi-t trillion dollar [5:40] problem. We've done tons of innovation [5:42] in life sciences. This is just data from [5:45] INI, inflammation, autoimmune diseases, [5:47] which is one of the biggest areas. [5:49] Probably know somebody with rheumatoid [5:50] arthritis, MS, ulcerative colitis, IBD, [5:54] psoriasis, AD, many other diseases. So [5:57] if you look at the top 10 drugs, they're [5:59] biologics. They eat 65 billion of the [6:03] 175 billion we spend on drugs. The tough [6:06] part is they get to less than 50% of the [6:08] most severe patients and less than 10% [6:10] of all patients diagnosed ever get [6:13] access to these therapies. [6:15] That's not okay. We know that these [6:17] pathways work, that they can modify the [6:20] course of disease. We have to find a way [6:22] which is the challenge we started with [6:25] of drugging those pathways. How could we [6:27] think differently? So at Monte we [6:31] started with a simple premise. Most of [6:33] the AI that was introduced into this [6:36] industry over the last decade was really [6:38] focusing on sort of these known [6:40] structural biology methods. How could we [6:42] make those much more efficient? But for [6:45] those biologic targets that we've [6:46] proven, those targets actually typically [6:49] don't have any known binding pockets. [6:52] You don't know anywhere to latch on to [6:54] that. So we had to think differently [6:56] about where we started and how we solve [6:58] the problem. And so that was our 10x [7:01] view of the future. What if we could [7:02] climb every single one of those [7:04] mountains so we could democratize access [7:07] to therapies for these known pathways? [7:10] And how could we figure that out? And so [7:12] we had to do something really different [7:14] from what the industry was doing. And I [7:16] would venture similar to my Google [7:18] experience. Everyone in the drug [7:19] industry looked at me and thought we [7:21] were crazy when we said, "What if we go [7:23] back to this pool of chemistry?" A lot [7:25] of people in the 70s and the 80s looked [7:28] at natural products, but they were like, [7:30] "Oh, these things are too hard to figure [7:31] out. They're polyfarm pharmacy. They're [7:34] can't manufacture them." Well, what if [7:36] we could actually for the first time [7:38] take that entire green space, all the [7:39] language nature wrote, and we could [7:41] focus on the subset that had already [7:42] been in human bodies for thousands of [7:44] years, and we could make that a useful [7:47] starting point for drugs. [7:50] And then we could actually start with [7:52] the biology and we could use machine [7:54] learning to precision rank it. So out of [7:57] that billion that we curated, we could [7:59] precision rank the best 1% on all its [8:02] drug-like properties, bi bioactivity and [8:04] all those other properties. And then we [8:06] could actually look at multiple [8:08] modalities and come up with completely [8:10] uncorrelated scaffolds to improve [8:11] probability of success. If we could do [8:13] that, we would be able to climb those [8:16] mountains. That was our belief. And it's [8:19] really exciting that it's actually [8:21] working. We've managed to curate over a [8:24] billion drug-like starting points. These [8:27] are all core molecules that are human [8:29] grounded. They come from food, herbs, [8:31] and traditional medicines. So, they're [8:32] core, but then we derivatized them using [8:37] uh derivation and and derivation. We [8:40] kept the core and then we divert we [8:42] iterated them so they would become [8:44] drug-like. So, they would have better [8:45] drug-like properties and be [8:46] manufacturable in three or less [8:48] retroynthetic steps. And we actually [8:51] look at the broadest range of [8:53] modalities. So whole biotech companies [8:55] are just focusing on coalent binders or [8:58] molecular glues or macro cycles. Well, [9:00] we said nature wrote all of those. Why [9:02] can't we have universal search and find [9:04] the best possible match for a pathway? [9:06] We then built the world's leading [9:08] decoding engine which is not scaffold [9:10] dependent. It's scaffold agnostic. So [9:12] we're able to decode that complex. So [9:14] there's a million ways in any molecule [9:16] structure and function are linked. And [9:18] now we can use giant decoding engines. [9:20] We have a big partnership with Nvidia [9:22] that enables us to use multimodal data [9:24] so that we can decode what's written [9:27] into that chemistry, that language, and [9:29] we can precision match it to a pathway. [9:31] So we built our foundation models. We [9:34] then understand and build a really [9:36] efficient set of label data that [9:38] represents that diverse 1 billion [9:40] chemical space. And then we're able to [9:42] predict and rank molecules from that [9:45] with precision for a pathway. And then [9:48] we had to think differently about the [9:50] drug development process because you're [9:51] already ranking on all these. It's not [9:54] the historic sequential steps of drug [9:56] development. You actually start with the [9:57] target candidate profile from the [9:59] beginning and everybody weighs in on [10:01] exactly what good looks like and then as [10:04] we choose the chemistry that we put [10:05] through the wet lab, we are actually [10:07] thinking up front about all the [10:09] liabilities we're taking on. So we [10:11] completely changed it to a much more [10:13] integrated problem solving process. And [10:15] to do that we had to create enormous [10:17] amounts of AI tools so we could run [10:19] scenarios so that everybody could ask [10:21] their questions live and we would be [10:23] able to say well if you want to look it [10:24] like that that if you wanted to buy [10:26] these molecules or buy those molecules [10:28] because it you can't make and test all [10:30] billion. It would cost you many billions [10:32] of dollars and take years. You have to [10:35] identify the best 5 to 10,000 to make [10:37] and test and that's what makes it [10:39] efficient. But then you have to rank [10:41] those and choose what you're willing to [10:43] take on the chemotype, breath, the types [10:45] of molecules. And then we've had to [10:48] prove that it's repeatable because it's [10:50] great if you can do it once, but we took [10:52] on 10 of the hardest targets. Um, all [10:54] biologic replacements that could be [10:57] transformative in the care. And of [10:58] course, these are allworked targets. So [11:00] we got a lot of network learning. And [11:02] then we're going broader with partners [11:03] in oncology. So all that sounds great. [11:06] just a preamble [11:08] to say, "Boy, getting this done was not [11:10] easy. Getting these two disciplines to [11:13] start working totally differently [11:14] together. So you can chart the vision. [11:18] You can imagine what's possible and then [11:20] you actually have to build the team and [11:22] accept where is the friction for us. [11:26] Machine learning scientists are [11:28] generalists, right? 60 70% of my people [11:31] can be crossrained and work anywhere. On [11:34] the right, I got narrow and deep, highly [11:36] specialized experts. ML scientists look [11:40] at problems as multiffactorial and [11:41] integrated, whereas life sciences are [11:44] right sequential and stepwise, [11:46] directional versus precise, explore [11:48] versus mitigate risk, right? How do you [11:50] get these people to actually work in a [11:52] team? In fact, I had the naivity when I [11:54] first started the company to think I'd [11:55] put machine learning people with me on [11:57] the West Coast and biologists and and I [11:59] was like, that's never going to work. [12:00] They have to sit in the same office, [12:02] look at each other face to face and [12:04] figure it out. And we have to think of [12:07] ourselves as AI leaders as how do we [12:10] actually overcome the rate limiting [12:12] steps and how do we supercharge drug [12:15] drug developers. The AI does not matter [12:17] at all. What matters is how do we help [12:20] drug developers overcome their [12:22] obstacles. And once we got through that, [12:25] we started to get better and better [12:28] because this was part was no fun, right? [12:31] And I could make the same on the other [12:32] side on the biology and even worse in [12:34] biology. You ask four biologists, the [12:36] biologist and chemist, they don't even [12:37] use the same language. So, how do we [12:39] figure this out and learn to speak human [12:42] together? Well, the first thing, it [12:44] sounds pretty obvious, but it's actually [12:46] a lot of work. I recommend everyone do [12:48] this in any organization. We started out [12:50] with if we're going to solve this huge [12:52] problem, what are the shared values that [12:54] uniquely we have to be good at as a [12:58] team? If we're going to do that, we [13:00] called it great expedition behaviors [13:03] because we're going to climb those [13:04] mountains. We need to be a great [13:05] expedition team. What is a great [13:07] expedition team about? You don't leave [13:08] anybody behind. Everybody plays [13:11] different roles and has different [13:12] skills, but you all work together as a [13:14] team and you stick together. You do not [13:16] leave anybody, right? You don't want [13:18] anyone to die in that mountain going up. [13:20] And so how do we create that? We also [13:23] had to create literally a shared [13:24] vocabulary. By the way, this really [13:26] helps when new people start in the [13:27] company. We have a company glossery for [13:31] every single term, how we use it, what [13:33] it's used for, how it works, so that [13:35] everybody knows when they're in a [13:37] meeting because I found in the beginning [13:39] people would say things, they walk out [13:40] of the meeting and five minutes later I [13:41] realized nobody understood what the [13:43] other person said and there was [13:44] absolutely no alignment whatsoever. So [13:46] after that I said forget that we're not [13:48] doing that anymore. We're all going to [13:49] have a shared vocabulary. It took us [13:51] like a month to do it. Best thing we [13:53] ever did. Super helps to onboard people. [13:56] And then we created shared forums to [13:57] learn. [13:59] We have to teach each other each other's [14:00] craft. We have to make it okay to raise [14:02] your hand and say I don't know. So we [14:04] did lunch and learns on every major [14:06] piece of technology or biology that [14:08] existed in the company got a lunch and [14:09] learn and we do them. They're recorded [14:12] and they get 100% attendance [14:13] crossunctionally because if we don't [14:16] commit to teaching each other how each [14:17] other thinks, then nobody understands [14:19] how to ask questions and you just create [14:21] this uncomfortable problem that we don't [14:24] get through and we can't build great [14:26] companies when we do that. So these [14:29] things really exist. We I didn't make [14:31] this stuff up about mountains. That's [14:33] what Montai means. Climbing mountains [14:35] with AI to solve this chronic disease [14:37] challenge. And these posters exist in [14:39] every conference room and every wall you [14:41] get it. What does that mean? And under [14:44] each underneath each of these are like [14:46] four behaviors. Human-driven pioneering [14:49] because we're trying to solve a big [14:50] problem, but it's only about the people. [14:52] It's not about the technology. [14:53] Relentless hunger for impact because [14:55] this is super hard and it's urgent. [14:57] Every day matters with us not actually [15:00] going faster. Fast is not a concept that [15:02] biologists and drug developers can [15:04] imagine. But if you reframe it in a [15:07] different kind of way, interdependently [15:10] ambitious, we accepted from the [15:11] beginning that none nobody could do [15:13] something without the other and a whole [15:15] set of behaviors that go with that. And [15:17] inclusive courage was everyone has to be [15:19] willing to sit up and raise their hand [15:20] if they ever think we're not doing [15:21] something right or you have to be [15:22] obsessed about talking about everybody [15:24] inside and outside the building to [15:26] ensure that you didn't miss anything. [15:28] And we kind of made that a core value. [15:30] It sounds kind of obvious, but what [15:32] we're doing is normalizing [15:35] the friction because it's the friction [15:36] between those two sides which is where [15:38] the magic happens. It's those hard [15:40] questions. It's the drug developer [15:41] saying, "But I need it like this." And [15:43] the ML person saying, "Well, what about [15:45] that? I could give you this, but that," [15:46] and you know, those give and takes and [15:48] the backs and force actually are what [15:49] creates the inspiration for innovation. [15:53] How do you engineer that into your [15:55] culture? The other thing a great [15:57] expedition team does is they have a [15:59] plan. [16:00] And what we found early on, and these [16:03] things again sound obvious, but I highly [16:05] recommend. They're not as easy as they [16:06] sound. We literally created a journey [16:08] map. It literally has every single task [16:12] of how we do it. It says who makes those [16:15] decisions top and bottom [16:17] and it hasn't a metric. So again, what [16:20] we've normalized is just agree. This is [16:22] what we agreed for now, but then every [16:25] quarter we have OKRs and that thing gets [16:27] better. every week, every month, every [16:30] quarter, which in the beginning [16:31] biologists don't like, okay, it's like [16:33] science is a journey. Anyone in here [16:35] knows about med that's not predictable, [16:38] right? Things go wrong. But people [16:39] quickly realize that. But what that does [16:42] is it forces discussion. If that's [16:44] taking longer or that wasn't exactly [16:46] what you want, machine learning can go [16:47] find you something better. They can go [16:49] and so releasing people from the [16:52] traditional ways of thinking and [16:53] constraints was very powerful in OKRs as [16:56] well. They faced a lot of resistance, [16:59] especially quarterly. People quickly [17:01] realized they became the foundation for [17:04] innovation collectively. And again, I I [17:07] really recommend if you're in one of [17:09] these tougher to change industries that [17:11] you use them in that way. You use them [17:13] to create. And we made a big mistake in [17:15] the beginning where we had OKRs that [17:18] were for the machine learning and OKRs [17:20] for the biologists. And we quickly [17:22] realized what that made people feel is [17:23] that machine learning conceded when the [17:25] biologist didn't. And I thought, wait, [17:27] that's crazy. We're on one expedition [17:29] team. Just the art of sitting in a room [17:32] and agreeing only OKRs that everybody [17:34] shared and everything fed into that [17:37] completely changed the dialogue. And I [17:39] see my head nods. When you're in a [17:41] company where everyone doesn't have the [17:43] same shared goal and then ladder into [17:45] it, that creates its own friction. I [17:47] realized that we were not efficient [17:50] literally for something that simple. [17:53] How do you create safe space for tough [17:55] conversations? When you work in biology, [17:57] you live in the gray. All those [17:58] industries on the right, there's a lot [18:00] of gray zones. There's no perfect test. [18:04] There is no perfect answer. [18:07] How do we help people live in the gray [18:09] together? We literally do something [18:10] called a here training. H E A R. No [18:13] kidding. You can look it up online. It [18:15] really helps how you actually have a [18:18] conversation when you fundamentally [18:20] think you don't agree. How do you do [18:21] that in a respectful [18:23] and valuable way so you both can move [18:26] forward together versus people going to [18:28] corners? This really helped us that was [18:30] slowing us down and we committed to [18:32] that. And then we do something. We have [18:34] a great expedition behavior award every [18:36] month at the town hall. These are crossf [18:38] functional nominations of individuals or [18:40] teams and we give them compasses [18:43] and it's about an action that they took [18:45] that was above and beyond that [18:47] represents those values and the fact [18:50] that everybody wants to be on that wall [18:52] with their picture with that compass. I [18:55] know it's again maybe sounds cheesy to [18:57] some of you but it but it really sends a [18:59] message when every single month you are [19:01] refor reinforcing what does it mean to [19:03] be on a great expedition team like [19:05] actually and that your peers nominate we [19:08] also have a slack channel which is [19:10] called you're awesome so every time a [19:12] crossf functional colleague does [19:14] something to help you front back and [19:16] side of that task chain you post [19:18] something on that what does that do [19:20] that's a dopamine hit and people say oh [19:22] I want to do more of that and again I [19:24] think we Don't think about the [19:25] psychology of organizations. If we want [19:27] people to collaborate when fundamentally [19:29] there is no reason that their [19:31] professions make sense together, we have [19:34] to create that together. So as a [19:37] company, we're scaling these mountains. [19:39] And I truly believe that if we hadn't [19:42] built that organization the way we did [19:46] and really learn to live together as an [19:48] expedition team, we could have never [19:50] succeeded in this industry. And I'm [19:52] really proud of this team every day. And [19:54] thank you for listening.