Synthesize Bio Models RNA Shifts to Slash 90% Clinical Failure Rate

How to Build a Virtual Human for Drug Discovery With Synthesize Bio — Ai4

Ai4

Simulating post-treatment RNA dynamics with generative latent diffusion more than doubles patient response prediction accuracy over standard clinical biomarkers.

AI-assisted summary of How to Build a Virtual Human for Drug Discovery With Synthesize Bio — original source. Check the original for full context.

Source published: 2026-08-19

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 19, 2026, this presentation features Gregory Koytiger, Vice President and Head of Artificial Intelligence at Synthesize Bio. The talk outlines Synthesize Bio's computational approach to overcoming drug discovery bottlenecks by constructing virtual humans that model biological systems and predict patient responses to novel therapeutics.

## TL;DR Gregory Koytiger argues that clinical drug trial failures stem primarily from unpredictable human biology rather than poor molecular design. To solve this, Synthesize Bio built a biological 'world model' that uses transcriptomics as a dynamic coordinate system, pairing a Variational Autoencoder with a metadata-conditioned Diffusion Transformer trained on over 15,000 standardized biological interventions. The system operates in both synthetic population generation and personalized reference conditioning modes to forecast gene expression shifts and downstream clinical responses. Benchmarked against empirical wet-lab reproducibility and validated on Phase 3 oncology trial data like IMbrave150, the platform enables rapid in silico clinical trials and indication expansion.

## Atmosphere & style The presentation maintains a sharp, authoritative, and methodical tone characteristic of high-level biotechnology keynotes. Koytiger speaks with measured confidence, transitioning smoothly from sobering clinical trial failure statistics to elegant architectural solutions and concrete empirical benchmarks. The atmosphere feels pragmatic and forward-looking, conveying a sense of engineering realism rather than speculative hype as complex biological simulations are demonstrated to match real-world wet-lab reproducibility.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) The primary bottleneck in modern drug discovery is human biology rather than molecular engineering, as roughly 90% of clinical candidates fail and more than one-third of Phase 3 failures stem from lack of therapeutic efficacy despite confirmed target binding. - [01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) Synthesize Bio addresses this challenge by building a biological 'world model' that simulates how therapeutic actions transition a biological system from an initial pre-treatment state to a future post-treatment state. - [01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) RNA abundance (transcriptomics) was selected as the core state representation because it dynamically captures granular, gene-level molecular shifts across tissues, diseases, and modalities better than static DNA or high-level electronic health records. - [03:51](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=231s) An agentic LLM-driven data harmonization engine resolves experimental fragmentation across public studies, structuring a unified training corpus covering over 15,000 biological interventions. - [03:51](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=231s) The model architecture couples a Variational Autoencoder (VAE) to compress high-dimensional transcriptomics away from batch effects with a metadata-conditioned Diffusion Transformer that generates post-treatment profiles. - [05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) The platform functions in two distinct regimes: population generation to synthesize diverse, demographically accurate virtual patient cohorts, and reference conditioning to predict personalized outcomes from an individual patient's baseline profile. - [05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) Treatments are parameterized by biophysical mechanisms—including target identity, binding affinity, dose, and exposure duration—allowing small molecules, antibodies, and genetic therapies to share a continuous modeling framework. - [07:54](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=474s) Evaluated on a strict post-July 2024 temporal holdout dataset, the model's correlation scores for differentially expressed genes fall within the empirical bounds of wet-lab experimental reproducibility. - [10:27](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=627s) In a validation study using Genentech's Phase 3 IMbrave150 trial data, simulated post-treatment molecular shifts provided more than double the patient stratification accuracy of standard PD-L1 immunohistochemistry biomarkers. - [12:02](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=722s) Virtual clinical trials simulating 100-patient cohorts across multiple oncology indications ran in minutes, accurately prioritizing squamous cell carcinoma over adenocarcinoma for combination therapy expansion. - [12:02](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=722s) The broader objective of biological world modeling is to invert the pharmaceutical paradigm, moving from a 90% clinical failure rate to a 90% success rate.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) The Core Bottleneck in Drug Discovery: Biology Over Molecules

Bringing a novel therapeutic from early research to regulatory approval remains one of the most expensive and inefficient endeavors in modern science. Currently, roughly 90% of all drug candidates that enter human clinical trials fail, resulting in an average cost exceeding $1 billion per approved drug. The leading cause of this attrition—responsible for more than one in three Phase 3 failures—is a fundamental lack of clinical efficacy. In many of these cases, the candidate molecule performs its intended biochemical task with absolute precision: it binds to its target and inhibits its designated pathway, yet this on-target modulation fails to alter the trajectory of the disease or meaningfully benefit the patient. As Gregory Koytiger stresses, "The bottleneck of drug discovery is not the molecule it's the biology."

To resolve this bottleneck, computational biology must move beyond standalone structural predictions and tackle the operational questions that dictate clinical success: identifying which specific patient populations will respond to a therapy, forecasting whether a clinical trial will meet its primary endpoints, and determining which subsequent indications should be pursued if a trial succeeds.

- **Clinical trial failure rate:** Approximately 90% of candidates fail upon entering human trials. - **Cost per approval:** ~$1 billion per successful drug. - **Phase 3 efficacy failures:** More than 33% of Phase 3 failures stem directly from lack of therapeutic efficacy despite verified molecular target engagement.

---

[01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) Defining a World Model: Why RNA Abundance Is the Ideal State Representation

Addressing these biological bottlenecks requires constructing a true "world model for human biology" rather than another protein folding algorithm. While the term "world model" has become an industry buzzword, its foundational definition is straightforward: it takes a numerical state representing a system at a given point in time, applies an action, and simulates the resulting state at a future time. In drug discovery, the "actions" correspond directly to therapeutic interventions—ranging from approved medicines currently in the clinic to hypothetical compounds existing only in a medicinal chemist's design pipeline.

```mermaid flowchart LR accTitle: World Model of Biology Architecture subgraph Input["System Input"] S0["Biological State S(0) (Pre-treatment RNA Abundance)"] A["Therapeutic Action (Target, Affinity, Dose, Modality)"] end subgraph Model["World Model"] WM["Diffusion Transformer & Latent VAE"] end subgraph Output["System Output"] S1["Biological State S(t) (Post-treatment RNA Profile)"] P["Phenotypic Readout (Response vs Non-Response)"] end

S0 --> WM A --> WM WM --> S1 S1 --> P ```

Defining the "state" of human biology requires a coordinate system that is both dynamic and universally comparable across conditions. Synthesize Bio selected RNA abundance (transcriptomics) as the core state representation. Unlike DNA, which remains largely static across treatments, gene expression dynamically mirrors immediate and downstream therapeutic responses. Furthermore, unlike high-level electronic health records (EHR), RNA quantification operates at the granular resolution of individual genes. This gene-level coordinate system captures subtle molecular differences between patients with identical clinical diagnoses, distinguishing potential responders from non-responders across all tissues, diseases, and therapeutic modalities.

---

[03:51](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=231s) Architecture: Data Harmonization and Latent Diffusion Transformers

Constructing a universal predictive model requires overcoming extreme data fragmentation. Training data is distributed across thousands of independent studies, each utilizing idiosyncratic naming conventions, varied sample preparation protocols, and disparate mRNA quantification pipelines. To overcome this, Synthesize Bio developed an agentic, LLM-driven data harmonization engine designed to map unstructured experimental studies into a standardized, controlled vocabulary enriched with external biomedical ontologies. This pipeline generated a unified dataset covering more than 15,000 distinct biological interventions across all major human tissues and disease areas.

To model this massive, heterogeneous corpus, the team adopted a two-stage latent diffusion architecture adapted from modern generative image models:

1. **Latent Compressor (Variational Autoencoder / VAE):** Compresses high-dimensional, noisy transcriptome measurements into a lower-dimensional latent space. This prevents downstream models from overfitting to batch effects and experimental noise, forcing the architecture to learn core semantic biology. 2. **Generative Core (Metadata-Conditioned Diffusion Transformer):** Operates on the latent space, progressively denoising random latents or reference states conditioned on rich therapeutic metadata to generate accurate post-treatment transcriptomic profiles.

- **Intervention scale:** 15,000+ distinct interventions across major tissue groups and diseases. - **Two-stage generative pipeline:** High-dimensional VAE compression coupled with a metadata-conditioned Diffusion Transformer.

---

[05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) Operating Modes: Population Generation vs. Personalized Reference Conditioning

The generative diffusion architecture operates in two complementary regimes:

* **Population Generation Mode:** The model takes pure noise alongside conditional metadata prompts (such as disease type and demographic parameters) and synthesizes realistic, multimodal virtual patient cohorts. When tasked with simulating breast cancer, for example, the model autonomously generates appropriate distributions across all clinical subtypes while strictly preserving biological realities, such as female demographic skew and relevant age distributions. * **Reference Conditioning Mode:** Rather than synthesizing an entire population from scratch, the model accepts a specific patient’s measured baseline pre-treatment gene expression profile. It evaluates subtle pre-existing molecular features to predict that individual's unique post-treatment gene expression state, revealing the precise biological mechanisms driving therapeutic response or resistance.

To prevent the system from degrading into isolated single-drug lookup models, therapies are parameterized by their underlying biophysical mechanisms: molecular target identity, target binding affinity, dosage concentration, and exposure duration. This mechanistic conditioning enables a unified framework capable of handling small molecules, monoclonal antibodies, and genetic interventions within a single continuous biological model.

---

[07:54](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=474s) Temporal Validation and Matching Experimental Reproducibility

To verify that the platform extrapolates to novel biology rather than simply interpolating within its training distribution, Synthesize Bio instituted a strict temporal evaluation split. The architecture was trained exclusively on data published up to July 2024 and evaluated against subsequent experimental datasets gathered after that cutoff. Because post-cutoff studies represent genuine experimental efforts to answer unresolved biological questions, they serve as a rigorous out-of-distribution benchmark for novel drug ideas.

```mermaid flowchart TD accTitle: Model Accuracy Benchmarking Range subgraph RealWorld["Empirical Wet-Lab Reproducibility"] CrossLab["Cross-Lab Reproducibility (Pearson r ~0.3 - 0.6)"] IntraLab["Intra-Lab Reproducibility (Pearson r ~0.6 - 0.7)"] end subgraph InSilico["Synthesize Bio World Model"] Pred["Model Predictions on Post-July 2024 Holdout (Matches Experimental Reproducibility Range)"] end

CrossLab --- Pred IntraLab --- Pred ```

Evaluating gene-level predictions via Pearson correlation on differentially expressed genes requires appropriate contextual framing, as wet-lab assays inherently exhibit experimental variation. Replicating the exact same biological experiment across different laboratories yields empirical correlation scores ranging from 0.3 to 0.6, while intra-lab replicates reach approximately 0.6 to 0.7. Synthesize Bio’s model achieves prediction scores directly within this experimental reproducibility ceiling across both small molecules and antibody modalities. Model performance remains robust across clinical and primary human tissue samples (avoiding immortalized cell line artifacts), with slight performance dips occurring primarily in data-sparse domains such as rare genetic syndromes.

---

[10:27](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=627s) Translating Molecular Shifts into Clinical Response: The IMbrave150 Case Study

To deliver clinical value, predicted gene expression trajectories must translate directly into patient stratification metrics. By feeding simulated post-treatment molecular shifts into downstream classification models, the platform forecasts overall treatment response to determine optimal clinical trial enrollment criteria.

Synthesize Bio validated this approach using Genentech's Phase 3 IMbrave150 clinical trial, which evaluated the combination of atezolizumab (anti-PD-L1) and bevacizumab (anti-VEGF/anti-angiogenesis) in patients with unresectable hepatocellular carcinoma (HCC):

* **Standard Clinical Biomarker (PD-L1 IHC):** Stratification based solely on target protein expression via immunohistochemistry (IHC) demonstrated poor discriminative capacity in separating responders from non-responders. * **Pre-treatment Transcriptomics:** Baseline unconditioned gene expression profiles provided modest predictive utility. * **Synthesize Bio Virtual Human Model:** Leveraging simulated post-treatment response shifts achieved superior predictive power, delivering more than twice the classification accuracy of standard methods at a 10% false positive rate threshold.

---

[12:02](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=722s) Simulating Virtual Clinical Trials for Indication Expansion

The predictive framework established in hepatocellular carcinoma was scaled across oncology to simulate in silico clinical trials for broad indication expansion. Synthesize Bio generated virtual cohorts of 100 diverse patients for every major cancer indication, simulating their response to the anti-PD-L1 and anti-VEGF combination regimen in a matter of minutes rather than the years required for multi-center physical trials.

```mermaid flowchart LR accTitle: In Silico Indication Expansion Workflow subgraph Cohorts["Virtual Cohort Generation"] C1["100 Virtual Patients (Adenocarcinoma)"] C2["100 Virtual Patients (Squamous Cell Carcinoma)"] Cn["100 Virtual Patients (Other Major Cancers)"] end subgraph Simulation["Rapid In Silico Trial (Minutes)"] Sim["Simulate Combination Therapy (Anti-PD-L1 + Anti-VEGF)"] end subgraph Prioritization["Indication Prioritization"] Rank["Ranked Therapeutic Response (Prioritizes Squamous Cell over Adenocarcinoma)"] end

C1 --> Sim C2 --> Sim Cn --> Sim Sim --> Rank ```

Where real-world clinical data was available for comparison—specifically in adenocarcinoma and squamous cell carcinoma—the simulation correctly predicted overall response rates and accurately prioritized squamous cell carcinoma over adenocarcinoma as the superior candidate for label expansion.

Gregory Koytiger concludes with the overarching vision of Synthesize Bio: transforming the biopharmaceutical paradigm within our lifetime from one where 90% of clinical candidates fail to one where 90% succeed, powered by models that make human biology predictable.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) The bottleneck of drug discovery is not the molecule it's the biology. — Gregory Koytiger

> [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) Roughly 90% of all drug candidates that enter human clinical trials fail, resulting in an average cost exceeding $1 billion per approved drug. — Gregory Koytiger

> [01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) Addressing these biological bottlenecks requires constructing a true "world model for human biology" rather than another protein folding algorithm. — Gregory Koytiger

> [01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) Unlike DNA, which remains largely static across treatments, gene expression dynamically mirrors immediate and downstream therapeutic responses.

> [03:51](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=231s) Compresses high-dimensional, noisy transcriptome measurements into a lower-dimensional latent space. This prevents downstream models from overfitting to batch effects and experimental noise, forcing the architecture to learn core semantic biology.

> [05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) The model takes pure noise alongside conditional metadata prompts (such as disease type and demographic parameters) and synthesizes realistic, multimodal virtual patient cohorts.

> [05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) It evaluates subtle pre-existing molecular features to predict that individual's unique post-treatment gene expression state, revealing the precise biological mechanisms driving therapeutic response or resistance.

> [07:54](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=474s) Replicating the exact same biological experiment across different laboratories yields empirical correlation scores ranging from 0.3 to 0.6, while intra-lab replicates reach approximately 0.6 to 0.7.

> [07:54](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=474s) Synthesize Bio’s model achieves prediction scores directly within this experimental reproducibility ceiling across both small molecules and antibody modalities.

> [10:27](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=627s) Leveraging simulated post-treatment response shifts achieved superior predictive power, delivering more than twice the classification accuracy of standard methods at a 10% false positive rate threshold.

> [12:02](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=722s) Transforming the biopharmaceutical paradigm within our lifetime from one where 90% of clinical candidates fail to one where 90% succeed, powered by models that make human biology predictable. — Gregory Koytiger

## Mentioned Gregory Koytiger, Synthesize Bio, Ai4, Genentech, IMbrave150, Atezolizumab, Bevacizumab, PD-L1, VEGF, Diffusion Transformer, Variational Autoencoder (VAE)

## Video description Gregory Koytiger - Vice President, Head of Artificial Intelligence, Synthesize Bio

Can we predict how a patient will respond to a therapy - including novel ones not yet in the clinic? Doing so would improve personalized treatment and accelerate the discovery of new medicines but requires overcoming significant challenges. In this talk, I will cover Synthesize Bio's approach to the problem and how we are building virtual humans to model therapeutic outcomes

[Source](https://www.youtube.com/watch?v=aj5hkhuF3eg)

中文简报

Synthesize Bio建模RNA变化以削减90%的临床失败率利用生成式潜在扩散模型模拟治疗后RNA动态，相较于标准临床生物标志物，患者反应预测准确率提高了一倍以上。 Gregory Koytiger认为，临床药物试验失败主要源于难以预测的人体生物学特性，而非糟糕的分子设计。为此，Synthesize Bio构建了一个生物学“世界模型”，将转录组学作为动态坐标系，把变分自编码器与在超过15,000项标准化生物干预上训练的元数据条件扩散Transformer结合。该系统在合成人群生成和个性化参考条件化两种模式下运行，用以预测基因表达变化及下游临床反应。该平台以湿实验经验复现性为基准，并在IMbrave150等3期肿瘤学临床试验数据上得到验证，能够实现快速的计算机模拟临床试验和适应症拓展。现代药物研发的主要瓶颈在于人体生物学而非分子工程，约90%的临床候选药物面临失败，且在3期临床试验失败中，超过三分之一即便确认靶点结合也仍缺乏治疗疗效。 (0:04) Synthesize Bio通过构建生物学“世界模型”应对该挑战，该模型可模拟治疗手段如何将生物系统从治疗前的初始状态转变为治疗后的未来状态。 (1:21) RNA丰度（转录组学）被选为核心状态表征，因其相比静态DNA或宏观电子健康记录，能够更好地动态捕捉跨组织、疾病和干预模态的基因级细微分子变化。 (1:21) 基于智能体LLM驱动的数据协调引擎解决了各公开研究之间的实验碎片化问题，构建了涵盖15,000多个生物干预的统一训练语料库。 (3:51) 该模型架构将用于压缩高维转录组数据并消除批次效应的变分自编码器（VAE），与用于生成治疗后特征谱的元数据条件扩散Transformer相结合。 (3:51)

Reading summary

Synthesize Bio proposes addressing the pharmaceutical industry's 90% clinical failure rate by shifting computational focus from molecular structure prediction to simulating dynamic human biology. Founder Gregory Koytiger contends that simulating post-treatment biological states can resolve Phase 3 efficacy bottlenecks, identify responder populations, and prioritize clinical trial indications. Overcoming the Clinical Efficacy Bottleneck (0:04) Roughly 90% of drug candidates entering human clinical trials fail, contributing to an average development cost exceeding $1 billion per approved therapy. Crucially, more than one in three Phase 3 trial failures stems directly from a lack of therapeutic efficacy, even when a candidate molecule binds to its intended target and modulates its pathway as designed. Koytiger emphasizes that the primary bottleneck in modern drug development is not the molecule itself, but the complex biological system in which it operates. To alter this trajectory, computational biology must address clinical decision points rather than isolated structural challenges. Effective platforms must forecast whether candidates will meet trial endpoints, identify the precise patient subpopulations most likely to benefit, and determine which follow-on disease indications justify pursuit after initial success. Modeling Biological States via Transcriptomics (1:21) Synthesize Bio defines a biological "world model" as an engine that takes an initial state S(0), applies a therapeutic action, and predicts the subsequent state S(t). The platform adopts RNA abundance (transcriptomics) as its universal state representation. Unlike static DNA, RNA dynamically reflects therapeutic responses, and unlike high-level electronic health records, it provides granular, gene-level resolution capable of differentiating responders from non-responders across tissues and disease types. To overcome disparate experimental pipelines, the company built an LLM-driven data harmonization pipeline that unified records across more than 15,000 distinct biological interventions. The core predictive architecture relies on a two-stage latent diffusion model adapted from generative vision: a Variational Autoencoder (VAE) compresses high-dimensional transcriptomic measurements to strip out experimental batch effects, while a metadata-conditioned Diffusion Transformer generates post-treatment transcriptomic profiles. Mechanistic Simulation and Experimental Benchmarks (5:49) The diffusion architecture operates in two distinct modes. In population generation mode, the model synthesizes diverse virtual cohorts from noise and demographic metadata, autonomously mirroring realistic disease subtype proportions and demographic skews. In reference conditioning mode, it takes a specific patient's pre-treatment gene expression profile to predict their unique molecular trajectory following therapeutic exposure. Interventions are parameterized by biophysical properties—such as target identity, binding affinity, dose, and duration—unifying small molecules, antibodies, and genetic therapies under one model. To ensure true generalization rather than memorization, the model was trained on studies published before July 2024 and evaluated against subsequent experimental datasets. On differentially expressed genes, the model's Pearson correlation scores matched empirical wet-lab reproducibility ranges (cross-lab r ~0.3–0.6; intra-lab r ~0.6–0.7) across both small molecules and biologics, though performance dipped slightly in data-sparse domains like rare genetic syndromes. Patient Stratification and In Silico Trials (10:27) Translating molecular forecasts into clinical utility was demonstrated using Genentech's Phase 3 IMbrave150 trial evaluating atezolizumab and bevacizumab in unresectable hepatocellular carcinoma. While standard PD-L1 immunohistochemistry showed poor discriminative power and baseline transcriptomics offered only modest predictive utility, Synthesize Bio's simulated post-treatment shifts achieved more than twice the classification accuracy of standard methods at a 10% false positive rate. Scaling this approach, the team simulated in silico clinical trials for indication expansion by generating 100-patient virtual cohorts across major cancers in minutes. Where clinical data was available, the simulation matched real-world response rates and successfully prioritized squamous cell carcinoma over adenocarcinoma for label expansion under the combination regimen. Synthesize Bio aims to invert biopharma's historical paradigm from a 90% failure rate to a 90% success rate by making human biology predictable through simulation. However, the platform's predictive performance experiences modest dips in data-sparse domains like rare genetic syndromes, and expanding virtual clinical trials across all oncology indications will depend on ongoing validation against real-world trial readouts.

精炼阅读摘要

计算生物学公司Synthesize Bio提出，新药研发的核心瓶颈在于生物学机制而非分子合成；通过构建以转录组（RNA丰度）为核心状态的生物学“世界模型”，能够在计算机中模拟药物干预后的分子与表型变化，从而应对临床试验高失败率的挑战。临床研发瓶颈与RNA核心状态表征 (0:04) 目前约90%进入人体临床试验的候选药物最终失败，平均每款成功获批药物的研发成本超过10亿美元。超过三分之一的3期临床试验失败直接源于缺乏临床疗效。在许多失败案例中，候选分子能精确结合预设靶点并抑制指定通路，但这种靶向干预未能改变疾病进程或使患者受益。Synthesize Bio首席执行官Gregory Koytiger强调，药物研发的根本瓶颈在于生物学本身，计算生物学必须从孤立的蛋白质结构预测转向解决患者群体响应分层、试验终点预测以及适应症拓展等实际运营问题。为构建能够推演生物动态的世界模型，Synthesize Bio选用RNA丰度（转录组学）作为底层状态坐标系统。与在治疗中相对恒定的DNA相比，基因表达水平能够动态反映治疗后的即刻与下游生物学反应；与粗粒度的电子健康记录（EHR）相比，RNA定量能深入到单基因分辨率，从而在各类组织、疾病和疗法中捕捉细微分子差异，区分临床表征相同但反应不同的患者。跨研究数据协调与潜变量扩散架构 (3:51) 构建通用预测模型需要解决训练数据高度分散的难题。各独立研究往往使用不同的命名规范、样本制备方案及mRNA定量流程。Synthesize Bio开发了由大语言模型（LLM）驱动的数据协调引擎，将非结构化实验数据映射至标准化受控词表及生物医学本体库，整合出涵盖所有主要人类组织与疾病领域、包含超过15,000种生物学干预的统一数据集。在建模策略上，团队借鉴现代生成式图像模型，构建了双阶段潜变量扩散架构。首先，利用变分自编码器（VAE）将高维且含噪声的转录组测量值压缩至低维潜空间，防止模型过度拟合批次效应和实验噪声，促使其学习核心生物语义；随后，以元数据为条件的扩散变换器（Diffusion Transformer）在潜空间内对随机潜变量或参考状态进行逐步去噪，从而生成干预后的基因表达谱。机制化参数调节与湿实验可重复性验证 (5:49) 该扩散架构具备两种操作模式：在群体生成模式下，模型结合纯噪声与疾病类型、人口学参数等条件提示，合成符合生物学规律的虚拟患者队列（例如合成乳腺癌队列时能够自主保留亚型分布、女性性别偏向与年龄结构）；在参考条件化模式下，模型输入具体患者基线治疗前的基因表达谱，预测其特异性的治疗后表达状态，以解析响应或耐药机制。该系统将疗法细化为分子靶点、结合亲和力、剂量浓度与暴露时间等生物物理机制参数，使单一模型能同时处理小分子、单克隆抗体及遗传干预。为检验模型的泛化能力而非单纯内插记忆，团队设立了严格的时间切片验证，仅使用2024年7月之前发表的数据进行训练，并使用该截止日期后发表的未知实验数据进行评估。在差异表达基因的皮尔逊相关系数（Pearson r）评估中，湿实验跨实验室重复的经验得分通常为0.3至0.6，实验室内重复约为0.6至0.7。Synthesize Bio的模型在小分子和抗体疗法上的预测表现落在此类实验可重复性区间内，且在临床和原代人体组织样本中表现稳健，仅在罕见遗传综合征等数据稀疏领域出现轻微性能下降。临床响应分层与适应症拓展模拟 (10:27) 为了将模拟的分子变化转化为临床决策依据，Synthesize Bio使用罗氏旗下基因泰克（Genentech）的IMbrave150 3期临床试验数据进行了回顾性验证。该试验评估了阿替利珠单抗（PD-L1单抗）联合贝伐珠单抗（抗VEGF/抗血管生成单抗）治疗不可切除肝细胞癌（HCC）的效果。验证表明，仅依据PD-L1蛋白表达（IHC）或治疗前基线转录组对患者进行分层，其区分能力均较为有限；而该模型模拟的治疗后分子变化在10%假阳性率阈值下，分类准确率达到了标准临床生物标志物方法的两倍以上。此外，该平台进一步模拟了抗PD-L1与抗VEGF联合疗法在多癌种中的适应症拓展试验。团队为各大主要癌症适应症分别生成了包含100名多样化患者的虚拟队列，在数分钟内完成了计算机模拟临床试验。在具备真实临床对照数据的腺癌与鳞状细胞癌中，模拟不仅准确预测了整体缓解率，还正确判定鳞状细胞癌相较于腺癌具备更高的拓展优先级。 Synthesize Bio主张通过建立可预测的人体生物学世界模型，在未来将候选药物临床试验90%的失败率转变为90%的成功率。然而其实验数据也表明，在缺乏充足公开研究的罕见遗传病等数据稀疏领域，模型的预测精度仍受到现实数据密度的客观限制。

Detailed notes — zh

## Background 本视频由 Ai4 于 2026 年 8 月 19 日发布在 YouTube 平台上。Synthesize Bio 的人工智能副总裁兼负责人 Gregory Koytiger 在演讲中介绍了团队如何构建用于建模治疗结局的“虚拟人类”（Virtual Humans），旨在攻克临床前预测患者疗效的重大挑战并加速新药研发与个性化治疗。 ## TL;DR 针对新药临床试验高达 90% 失败率的行业瓶颈，演讲指出成败的核心在于人体生物学而非单纯的分子设计。Synthesize Bio 构建了一套生物学世界模型，将全基因组 RNA 表达丰度作为统一状态空间，采用 VAE 降维压缩结合条件 Diffusion Transformer 架构推演药物干预后的分子状态变迁。该模型支持虚拟患者群体生成与个体化疗效预测，在 IMbrave 临床试验中实现了两倍于临床现有标志物的应答预测准确率，并能在数分钟内完成高通量虚拟临床试验以指导适应症拓展。 ## Atmosphere & style 整体演讲呈现出前沿科技峰会特有的严谨、理性与前瞻性氛围。主讲人以冷静犀利的行业困境剖析开篇，随后条理分明地拆解复杂的生成式模型架构与严苛的时间切分验证，叙事节奏层层递进、论据扎实。观看过程既能感受到硬核计算生物学的严密工程逻辑，又传递出一种彻底逆转制药行业高失败率现状的颠覆性愿景与技术自信。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) 制药行业面临进入临床的药物有 90% 失败且平均上市成本达 10 亿美元的困境，失败主因是生物学层面的缺乏有效性，而非分子靶向失效。 - [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) AI 药物研发的建模重心需从单纯预测分子结构与蛋白质折叠，转向构建能够预测患者应答、试验结局及适应症拓展的生物学“世界模型”。 - [01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) 全基因组 RNA 表达丰度（mRNA Abundance）能够实时响应药物干预并表征细微分子异质性，是定义人体生物学状态的最佳统一坐标系。 - [03:51](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=231s) 采用 Agentic LLM 统一清洗和本体对齐跨研究的异构转录组数据，汇聚出覆盖 15,000 种干预手段及全人体主要组织与疾病的训练集。 - [03:51](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=231s) 模型采用两阶段生成架构：先通过变分自编码器（VAE）压缩高维表达数据过滤噪声，再利用元数据条件化的 Diffusion Transformer 预测分子变化。 - [05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) 模型支持两种工作模式：群体生成模式可生成符合流行病学与分子亚型的虚拟患者队列，参考条件模式则针对特定患者基线推演个性化疗效。 - [05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) 将药物干预按靶点、亲和力、剂量与作用时间等全生物学机制进行条件化注入，使模型能够统一表征小分子、抗体及基因疗法等多种治疗模态。 - [07:54](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=474s) 通过严格的时间切分（以 2024 年 7 月为界）与湿实验可重现性区间（dPearson 0.3-0.7）对标，验证了模型在原代临床组织上的外推泛化能力。 - [10:27](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=627s) 在罗氏 IMbrave 肝癌临床试验回溯中，该模型在 10% 假阳性率下预测疗效响应的准确度超过了临床现行金标准 PD-L1 IHC 达 2 倍以上。 - [12:32](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=752s) 通过高通量虚拟临床试验，团队能在数分钟内完成不同癌种的疗效模拟，成功在腺癌与鳞状细胞癌中复现疗效差异并辅助适应症拓展决策。 - [13:34](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=814s) Synthesize Bio 的终极愿景是使生物学本身变得可预测，将药物研发 90% 的临床失败率彻底逆转为 90% 的成功率。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) 药物研发的核心瓶颈：从分子结构走向生物学世界模型在制药行业中，进入临床试验阶段的药物有高达 90% 最终以失败告终，这使得将一款新药推向获批上市的平均成本高达 10 亿美元。导致失败的头号原因——在所有三期临床试验失败案例中占比超过三分之一——是缺乏有效性（Lack of Efficacy）。很多时候，药物在分子层面上完全达成了设计目标：精准找到了靶点并成功抑制了靶点，但对改变疾病进程或改善患者的临床结局却毫无帮助。正如演讲者所强调的，"新药研发的瓶颈不是分子本身，而是生物学。" 为了攻克这一核心瓶颈，AI 建模的思路必须发生根本转变。行业需要的不是预测靶点结合或又一个蛋白质折叠模型，而是能够回答真正决定临床成败的关键问题：谁会对这种疗法产生应答？这项临床试验能否取得积极结果？如果试验成功，接下来该用该药物拓展攻克哪种适应症？要解答这些问题，行业亟需建立一套针对人类生物学的“世界模型”（World Model）。 - **90% 临床失败率**：药物管线在临床阶段折戟的根本原因多为生物学有效性不足，而非分子靶向失效。 - **10 亿美元成本**：每款成功上市新药背后背负着高昂的试错与失败成本。 - **核心目标转变**：从单纯预测分子结构转向预测患者应答、试验结局及适应症拓展。 ```mermaid flowchart TD accTitle: 生物学世界模型工作流架构 subgraph S1["数据治理与状态表征"] A["多源异构转录组数据"] --> B["Agentic LLM 统一清洗与本体对齐"] B --> C["基线 RNA 丰度 (生物学状态)"] end subgraph S2["两阶段生成建模"] C --> D["变分自编码器 (VAE) 降维压缩"] D --> E["隐空间特征 Latent"] F["干预动作机制: 靶点/亲和力/剂量/时间"] --> G["条件 Diffusion Transformer"] E --> G end subgraph S3["下游临床转化"] G --> H["去噪生成干预后转录组状态"] H --> I["虚拟患者群体生成 (Population Mode)"] H --> J["个性化疗效预测 (Reference Conditioning)"] J --> K["虚拟临床试验与适应症拓展"] end ``` --- ### [01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) 生物学状态的数学表征：为何 RNA 丰度是最佳坐标系虽然“世界模型”这一概念如今被广泛滥用，但其核心定义非常清晰：输入一个表征系统在某一时间点全部信息的数值化“状态”（State），施加一个“动作”（Action），并预测系统在未来时间点生成的新状态。在药物研发语境下，“动作”的定义非常直观，即药物干预——无论是已在临床处方的成熟药物，还是药物化学家脑海中的新药设想。然而，如何定义人体的“生物学状态”一直是个难题。 Synthesize Bio 认为，RNA 丰度（Gene Expression / mRNA Abundance）是表征生物学状态的单一最佳载体。与终生基本保持不变的 DNA 不同，RNA 的表达丰度会随药物治疗而实时发生动态响应；与颗粒度较粗且非结构化的电子健康档案（EHR）相比，RNA 丰度直接落在基因层面，能够精确捕捉患有同一种疾病的不同患者之间的细微分子差异，从而区分出谁会起效、谁会耐药。更重要的是，基于全基因组的 RNA 丰度为所有疾病和所有药物构建了一个统一的数值坐标系。 - **动作（Action）**：涵盖临床现存药物、抗体疗法、基因干预以及处于早期概念阶段的全新化合物。 - **状态（State）**：以全基因组 RNA 丰度为统一状态空间，兼具对干预的动态敏感性与高分辨率分子表征能力。 - **建模目标**：输入任意个体、任意组织的基线状态，模拟任意药物干预，预测处理后的分子级状态转变，并进一步映射为高层次的表型与临床转归。 --- ### [03:51](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=231s) 数据治理与两阶段生成架构：VAE 压缩器与条件 Diffusion Transformer 构建全生物学尺度的预测模型面临巨大的数据挑战。训练数据散落在成千上万个不同的独立研究中，测量单位、命名规则以及 mRNA 量化方式各不相同。为此，团队开发了一套基于 Agentic LLM 的自动化数据治理系统，将异构数据集全面清洗、对齐并映射至受控词表中，并结合外部知识库进行数据增强，最终汇聚了涵盖 15,000 种不同干预手段、贯穿人体所有主要组织群和重大疾病类别的统一训练集。在模型架构设计上，团队借鉴了图像生成领域的 Latent Diffusion 两阶段思路： 1. **第一阶段（压缩器）**：采用变分自编码器（VAE），将包含大量实验噪声的高维全基因组表达量测量值压缩为紧凑的低维隐空间表征（Latent Representation），促使下游生成模型专注于高级生物学语义，避免过度拟合实验噪音。 2. **第二阶段（生成主干）**：采用元数据条件化的 Diffusion Transformer（DiT），以加噪的隐空间特征为起点，结合干预条件进行逐步去噪，从而生成干预后的分子表达谱或生成完整的虚拟患者队列。 - **数据规模**：覆盖 15,000 种生物学干预，横跨所有主要人体组织系统与疾病类型。 - **降噪机制**：通过 VAE 隐空间过滤测量噪音，保留关键分子调控特征。 - **生成引擎**：以元数据/机制为条件的 Diffusion Transformer 进行高精度去噪推理。 --- ### [05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) 两种生成模式与全机制条件注入该生成模型设计了两种核心工作模式： - **群体生成模式（Population Generation Mode）**：输入纯高斯噪声及条件元数据，模型能够生成具备真实生物学异质性、多峰分布的虚拟患者队列。例如针对乳腺癌，模型在严格遵守女性及特定年龄段流行病学约束的同时，自主生成涵盖不同分子亚型的虚拟患者群。 - **参考条件模式（Reference Conditioning Mode）**：输入特定患者治疗前的实际基因表达测量值，进行个性化干预模拟。模型能够识别治疗前极其细微的分子表达差异对最终疗效的巨大影响，精准推演用药后的分子状态变迁，进而解析药物起效的机制通路。为了让模型摆脱孤立拟合单一药物的局限，团队将药物干预以“完整生物学机制”进行条件化编码——不仅输入药物分子本身，还显式输入其生物学靶点、靶点结合亲和力（Affinity）、用药剂量（Dose）以及作用时间（Duration）。这种机制级别的统一建模，使模型能够将小分子化学药、大分子抗体乃至基因疗法无缝融合进同一套底层生物学世界模型中。 --- ### [07:54](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=474s) 时间切分严格验证与实验可重现性对标为了验证模型不仅能在训练集内插值（Interpolate），而且能对全新的前沿药物设想进行外推（Extrapolate），团队采用了极为严苛的“时间切分评估”（Temporal Data Splitting）：使用 2024 年 7 月之前的数据训练模型，将 2024 年 7 月之后产生的全新实验数据作为完全未见的测试集。由于生物学界开展新实验旨在解答未知的不确定性，这种跨时间泛化直接考验了模型对未知生物学的预测能力。在评估差异表达基因（DEG）的预测精度（dPearson 相关系数）时，团队引入了湿实验的基准线作为参照。现实中，相同实验在不同实验室间或同实验室内的可重现性 dPearson 通常落在 0.3 到 0.7 之间（完美得分为 1.0 在生物实验中并不现实）。评测结果显示： - **跨模态一致性**：无论是小分子、抗体还是其他治疗模态，模型的预测精度均稳定落在真实实验的可重现性区间内。 - **跨疾病泛化**：除罕见遗传综合征（因训练样本极少而表现稍弱）外，制药工业关注的大多数主要疾病均展现出优异预测性能。 - **临床相关性**：评测完全聚焦于临床患者样本和原代组织（Primary Tissue），而非培养皿中的贴壁细胞系，确保预测结果贴近真实人体生物学。 --- ### [10:27](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=627s) 临床转化落地：IMbrave 试验应答预测与跨癌种虚拟临床试验为了验证分子层面的表达谱变化能否转化为临床关心的应答判定，团队将模型预测的基因表达变化输入下游分类器，预测患者是否会对疗法产生应答，并在真实临床试验中进行了回溯验证。以罗氏/基因泰克（Genentech）开展的著名肝细胞癌（HCC）临床试验 IMbrave（评估抗 PD-L1 免疫疗法联合抗血管生成靶向药的组合疗法）为例：在临床现行的标准中，通常依赖免疫组化检测的 PD-L1 IHC 指标来筛选患者，但该指标在区分应答与无应答患者方面的表现极差。相比之下，Synthesize Bio 模型的预测效能显著超越了 PD-L1 表达量及基线全转录组基线模型。特别是在 10% 假阳性率（FPR）的严格筛选标准下，该模型的应答预测准确率达到了临床现有批准生物标志物的两倍以上。基于这一突破，团队进一步拓展至跨适应症模拟：为每一个主要癌症大类分别生成 100 名虚拟患者，开展高通量的计算机虚拟临床试验（In Silico Clinical Trials），系统评估该联合疗法拓展至其他癌种的应用潜力。 - **IMbrave 临床对照**：针对晚期肝细胞癌的抗 PD-L1 + 抗血管生成联合方案。 - **2 倍精度跃升**：在 10% 假阳性率下，模型区分疗效响应的准确度超过 PD-L1 IHC 临床金标准 2 倍以上。 - **规模化虚拟试验**：对全癌种构建每类 100 人的虚拟患者队列，直接在计算机内推演适应症拓展前景。 ### [12:32](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=752s) 虚拟临床试验与适应症拓展预测通过在不同的虚拟队列（Virtual Cohorts）中汇总模拟响应率，Synthesize Bio 能够系统性地评估药物分子在不同病症中的潜在疗效，进而明确判定该药物下一步应该优先拓展进入哪项适应症。这种基于虚拟人模型的计算模拟，使得团队能够在短短“几分钟内而非耗费数年时间”（"in the order of minutes instead of the order of years"）完成虚拟临床试验的筛选与验证。现实中，研究人员显然无法在真实的临床试验里逐一测试所有的药物组合与适应症分支。但在具备真实临床参考数据的领域——例如腺癌（Adenocarcinoma）与鳞状细胞癌（Squamous Cell Carcinoma）——模型的预测表现异乎寻常地精准。它不仅高度准确地预测了实际的疗效响应率，还成功将鳞状细胞癌列为优于腺癌的后续临床开发适应症。这一验证成果带来了极大的信心，充分证明了该模型在泛化推理以及预测特定疗法之适应症拓展（Indication Expansion）方面的强大能力。 ```mermaid flowchart LR accTitle: Virtual Clinical Trial Workflow subgraph S1["数据与模拟"] A["构建虚拟患者队列"] --> B["运行虚拟临床试验 (耗时数分钟)"] end subgraph S2["预测与决策"] B --> C["汇总各队列响应率"] C --> D["验证与优先级排序 (例如：鳞癌优先于腺癌)"] D --> E["指导临床适应症拓展"] end ``` - **模拟效率的跨越**：将传统临床验证从以“年”为单位的漫长周期压缩至“分钟级”的快速计算。 - **临床验证表现**：在已有真实数据对照的腺癌与鳞状细胞癌评估中，模型精准复现了疗效差异并做出了正确的优先级决策。 ### [13:34](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=814s) Synthesize Bio 的愿景：让生物学可被预测 Gregory Koytiger 阐述了 Synthesize Bio 的核心愿景：希望在这一代人的有生之年，彻底改变传统药物研发的高风险困局。当前制药行业面临着残酷的现状——进入临床试验阶段的药物中有 90% 最终以失败告终；而 Synthesize Bio 的目标是将这一体系彻底反转，推动实现“90% 进入临床的药物都能取得成功”的全新格局。要实现这一跨越式变革，其核心支柱就在于打造出能够“让生物学本身变得可预测”（"make biology itself predictable"）的底层模型。尽管目前该领域仍处于早期探索阶段，但技术演进的轨迹已经清晰可见，虚拟人类与计算生物学正在逐步将这一宏大愿景变为现实。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) 新药研发的瓶颈不是分子本身，而是生物学。 — Gregory Koytiger > [00:04](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=4s) 行业需要的不是预测靶点结合或又一个蛋白质折叠模型，而是能够回答真正决定临床成败的关键问题：谁会对这种疗法产生应答？ — Gregory Koytiger > [01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) RNA 丰度是表征生物学状态的单一最佳载体。 — Gregory Koytiger > [01:21](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=81s) 基于全基因组的 RNA 丰度为所有疾病和所有药物构建了一个统一的数值坐标系。 — Gregory Koytiger > [05:49](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=349s) 不仅输入药物分子本身，还显式输入其生物学靶点、靶点结合亲和力、用药剂量以及作用时间。 — Gregory Koytiger > [07:54](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=474s) 评测完全聚焦于临床患者样本和原代组织，而非培养皿中的贴壁细胞系，确保预测结果贴近真实人体生物学。 — Gregory Koytiger > [10:27](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=627s) 在 10% 假阳性率的严格筛选标准下，该模型的应答预测准确率达到了临床现有批准生物标志物的两倍以上。 — Gregory Koytiger > [12:32](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=752s) in the order of minutes instead of the order of years — Gregory Koytiger > [13:34](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=814s) 希望在这一代人的有生之年，彻底改变传统药物研发的高风险困局。 — Gregory Koytiger > [13:34](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=814s) make biology itself predictable — Gregory Koytiger > [13:34](https://www.youtube.com/watch?v=aj5hkhuF3eg&t=814s) 目标是将这一体系彻底反转，推动实现“90% 进入临床的药物都能取得成功”的全新格局。 — Gregory Koytiger ## Mentioned Gregory Koytiger, Synthesize Bio, Ai4, 罗氏 (Roche), 基因泰克 (Genentech), IMbrave 临床试验, Diffusion Transformer (DiT), 变分自编码器 (VAE), Agentic LLM, PD-L1 IHC ## Video description Gregory Koytiger - Vice President, Head of Artificial Intelligence, Synthesize Bio Can we predict how a patient will respond to a therapy - including novel ones not yet in the clinic? Doing so would improve personalized treatment and accelerate the discovery of new medicines but requires overcoming significant challenges. In this talk, I will cover Synthesize Bio's approach to the problem and how we are building virtual humans to model therapeutic outcomes [Source](https://www.youtube.com/watch?v=aj5hkhuF3eg)

Source quotes

[0:04] Gregory Koytiger: 新药研发的瓶颈不是分子本身，而是生物学。 [0:04] Gregory Koytiger: 行业需要的不是预测靶点结合或又一个蛋白质折叠模型，而是能够回答真正决定临床成败的关键问题：谁会对这种疗法产生应答？ [1:21] Gregory Koytiger: RNA 丰度是表征生物学状态的单一最佳载体。 [1:21] Gregory Koytiger: 基于全基因组的 RNA 丰度为所有疾病和所有药物构建了一个统一的数值坐标系。 [5:49] Gregory Koytiger: 不仅输入药物分子本身，还显式输入其生物学靶点、靶点结合亲和力、用药剂量以及作用时间。 [7:54] Gregory Koytiger: 评测完全聚焦于临床患者样本和原代组织，而非培养皿中的贴壁细胞系，确保预测结果贴近真实人体生物学。 [10:27] Gregory Koytiger: 在 10% 假阳性率的严格筛选标准下，该模型的应答预测准确率达到了临床现有批准生物标志物的两倍以上。 [12:32] Gregory Koytiger: in the order of minutes instead of the order of years [13:34] Gregory Koytiger: 希望在这一代人的有生之年，彻底改变传统药物研发的高风险困局。 [13:34] Gregory Koytiger: make biology itself predictable [13:34] Gregory Koytiger: 目标是将这一体系彻底反转，推动实现“90% 进入临床的药物都能取得成功”的全新格局。

Source quotes

[0:04] Gregory Koytiger: The bottleneck of drug discovery is not the molecule it's the biology. [0:04] Gregory Koytiger: Roughly 90% of all drug candidates that enter human clinical trials fail, resulting in an average cost exceeding $1 billion per approved drug. [1:21] Gregory Koytiger: Addressing these biological bottlenecks requires constructing a true "world model for human biology" rather than another protein folding algorithm. [1:21] Unlike DNA, which remains largely static across treatments, gene expression dynamically mirrors immediate and downstream therapeutic responses. [3:51] Compresses high-dimensional, noisy transcriptome measurements into a lower-dimensional latent space. This prevents downstream models from overfitting to batch effects and experimental noise, forcing the architecture to learn core semantic biology. [5:49] The model takes pure noise alongside conditional metadata prompts (such as disease type and demographic parameters) and synthesizes realistic, multimodal virtual patient cohorts. [5:49] It evaluates subtle pre-existing molecular features to predict that individual's unique post-treatment gene expression state, revealing the precise biological mechanisms driving therapeutic response or resistance. [7:54] Replicating the exact same biological experiment across different laboratories yields empirical correlation scores ranging from 0.3 to 0.6, while intra-lab replicates reach approximately 0.6 to 0.7. [7:54] Synthesize Bio’s model achieves prediction scores directly within this experimental reproducibility ceiling across both small molecules and antibody modalities. [10:27] Leveraging simulated post-treatment response shifts achieved superior predictive power, delivering more than twice the classification accuracy of standard methods at a 10% false positive rate threshold. [12:02] Gregory Koytiger: Transforming the biopharmaceutical paradigm within our lifetime from one where 90% of clinical candidates fail to one where 90% succeed, powered by models that make human biology predictable.

Original transcript

[0:04] Thanks everyone. Good afternoon. Um hear me? Um thank you so much for joining me today. So yeah, [0:10] let's just get into it. How do you build a virtual human for drug discovery? I [0:14] think some of you have seen a slide like this and many a deck. Uh but to recap, [0:19] 90% of drugs that enter clinical trials fail. That means it cost the pharmaceutical industry [0:25] $1 billion to bring a drug all the way to approval. And the number one reason [0:30] why drugs fail responsible for over one in three phase three failures is a lack of efficacy. The [0:37] drug does the what it was intended to do at the molecular level. It finds its target. It inhibits [0:43] its target but that does nothing to actually change the course of disease actually improve [0:48] the patient's outcome. The bottleneck of drug discovery is not the molecule it's the biology. [0:58] To build models that can actually address the core bottlenecks, [1:03] we have to think differently. We have to build models that ask that answer who will [1:08] respond to this therapy. Will this trial be positive? And if this trial is positive, [1:14] what's the next disease that I should go after with this drug? [1:21] And to do that, we don't need another, you know, protein folding model. We need a world [1:27] model for human biology. A world model is one of those terms like foundation model before it [1:33] that's gotten really popular and then lost all attachment to its original meaning. But [1:38] at its core, the original meaning is pretty straightforward. It takes a state that's a [1:43] numerical representation. It captures all of the information about a system at a point in time. [1:49] and takes an action that is performed on that state and then it generates a new state at a [1:55] subsequent point in time. Translating to drug discovery, the action part is relatively easy [2:02] to translate. Those are our drugs. Whether they're real drugs that are in the clinic being prescribed [2:07] today or whether it's a drug idea that's still only in the mind of a drug researcher, [2:12] those are the actions that are going to be driving this world model of biology. [2:17] But how do we represent the state? What does that even mean? Biological state [2:24] at Synthesized Bio, we believe that RNA abundance is the single best representation of biological [2:30] state. There's many reasons for it, but unlike DNA, for instance, it actually responds to a [2:36] therapy. Your RNA abundance changes, your DNA does not. And unlike your electronic healthcare record, [2:45] it's at the level of genes. So it captures really precise differences that differentiate people [2:52] that have the same disease, but actually might tell the difference between who will [2:55] respond to that treatment and who won't. And it also gives us a unified coordinate system [3:02] with which we can model all diseases and all drugs because it's all at the gene level. [3:08] This is not something that we can really do with electronic healthcare records. [3:16] So that's precisely the kind of model we set out to build a synthesized bio. Wanted [3:21] to create a model that can take an arbitrary tissue from an arbitrary person and simulate [3:26] the treatment with an arbitrary drug and tell you what happens afterward to that [3:31] person at the molecular level. And then we can translate those molecular responses into [3:37] high level phenotypes. This person respond to therapy. Are they moving in a healthy [3:42] direction or are they still progressing down and their disease is getting worse? [3:51] Now to build such a model, it's not easy. The first place that the training data for this model [3:57] rests in thousands of different studies. They all have their idiosyncratic naming systems. what they [4:04] capture, what they don't, how they quantify mRNA abundance. This really required us to create a [4:10] groundup system, an agentic LLM driven system that takes all these disparate data sets and harmonizes [4:18] them to a controlled vocabulary. We then enrich that controlled vocabulary with external resources [4:23] that further augment the data set. This allowed us to create a training data that covered 15,000 [4:30] different interventions spanning all the major tissue groups and all the major diseases and try [4:37] to train a single unified model across disease and drug. Now to build such a model we took [4:44] inspiration from the image model at community a latent diffusion type approach. In step one of a [4:50] latent diffusion model, which is a two-step model, we have a compressor, something that takes all [4:55] the highdimensional gene expression measurements that are full of experimental noise and compresses [5:00] them into a compact lower dimensional latent representation that allows the generative model [5:06] downstream to focus on the highlevel semantic meaning instead of getting sidetracked and [5:12] overfitting to experimental noise. And for that we use an AV variational autoenccoder. [5:24] The second stage model is the core generative model and that is our metadata conditioned [5:29] diffusion transformer. This model can then generate either whole virtual patient cohorts [5:37] or consequences of different therapies on the that patient cohort. takes the [5:42] noisy representation of a latent and dnoises it to generate the prediction. [5:49] We can operate in two different modes. In this population generation mode, we feeded [5:56] noise and it generates a full biologically like heterogeneous multimodal population that captures [6:04] a true biological heterogeneity of the data. So for instance, we can tell a breast cancer there [6:10] are multiple different types of breast cancer. The model generates all the different subtypes [6:14] while respecting that breast cancer is a disease that affects females and people of a certain age [6:19] demographic. It generates those the distribution and fits with a conditional distribution. [6:30] The second mode is uh what we call reference conditioning. It starts with a particular [6:37] person's gene expression measurement before treatment and asks how does that particular [6:42] person respond to therapy. [snorts] So now instead of generating populations, [6:46] we generate personalized predictions. The model is able to recognize subtle [6:51] differences in the pre-treated population that really have very dramatic consequences [6:56] on whether this person actually responds to treatment or not. Have generates the [7:01] post-treated gene expression profile. So you can then look at the mechanism that a drug uses [7:05] to identify the respondent population and what it predicts that responding population to do. [7:15] Now to for this model not to be a disconnected set of unique data models for unique therapies, [7:23] we also condition the therapy using its full biological mechanism. So what for a given drug, [7:31] what is its target? What's the affinity to that target? How much of this drug was [7:35] added and for how long? This allows the model to create a unified world [7:39] model biology that could integrate small molecule drugs, antibodies, [7:44] genetic interventions, all in a coherent mechanism of biology. [7:54] Now, that's all well and good, but does it actually work? But how do we know it would [7:58] work? And how do we know it would work in a way actually matters for new drug development? It [8:04] doesn't just interpolate within our training data set but actually extrapolate to new drug ideas. [8:11] To do that, we used a temporal data splitting approach. So we trained our model with all data up [8:18] to July 2024 and we held out all subsequent data from the model for evaluation. And then we ask, [8:25] can our model predict the actual genes that move in response to the therapy in this out [8:30] of time distribution? Now, people don't collect data out of time just for for fun, [8:36] right? They collected because it answers a fundamental uncertainty, [8:40] something that we don't really know about biology that we need to perform an experiment to do. [8:47] Now these numbers they're it's not obvious how do you interpret D Pearson for you know [8:54] differentially expressed genes. So we asked how well do experiments actually recapitulate each [8:59] other. So we looked at the same experimental description being performed in one lab versus [9:05] another or within the same lab. You can see these numbers range in sort of 3 to 6 [9:12] to 7 range depending on whether it's the same lab that performs the experiment or a different lab. [9:18] So we're not going to get this all the way up to 0.1 to to one per a perfect score but this [9:23] is a range that we should target and indeed that is the range that our model achieves. So whether [9:30] we look at small molecules or antibodies or any therapeutic modality we can see that our [9:35] model actually is in the range of experimental reproducibility and it doesn't really suffer [9:40] across different modalities of drug. We can slice this in a different way. You can ask how well does [9:48] our model perform across all the different diseases. Indeed, here we see a little bit [9:53] more variation. For instance, genetic syndromes which are usually rare diseases underrepresented [9:57] in our training data are not performed as well. Most of the major disease categories [10:02] that pharmaceutical industry is interested in are performed at pretty respectable rates. And here [10:08] we're only focusing on clinical and primary tissue. So, we're not interested in can our [10:12] model predict a cell line in a dish, but actually something that's really very close to the clinic. [10:27] And then we can also ask how do we translate this these molecule predictions, these gene [10:33] level changes into the things that people actually care about, which is am I responding to this drug [10:38] or am I not? who should I be recruiting for my clinical trial and who should I be excluding? [10:45] To do that, we can to take our predicted shifts or predicted responses for a given individual [10:51] and we could use that to train a downstream model that will then predict whether this [10:55] person is going to be responsive to therapy or not and we could compare it to what's going on [11:00] already in the clinic. So here we took the IM brave trial that was done by genetic that [11:06] looked at hpaticellular carcinoma and it asked the combination therapy of antiPDL1 inhibition [11:13] and anti-agision actually treat these people. If we look at what's clinically approved in the [11:20] pink line that's the PDL1 IHC that's just the the target itself. You can see that a predictor [11:28] using just this level is actually very poor at prioritizing response versus non-response. Whereas [11:35] our model is actually much stronger, including the fact that like at the 10% fall positive rate, [11:42] our model is over twice as accurate at predicting response from non-response. It's [11:47] able to really with high confidence predict who's going to respond here. [snorts] and [11:52] it's doing so at much better rates than even the full gene expression profile. [12:02] Now we could even take this one step further. We can take the model that we predicted on [12:07] hypadellular carcinoma and we could ask now that we know what is predictor of response can [12:14] we apply the predictor to every single cancer. So we simulated clinical trials across all different [12:22] major cancer indications. So we generate a 100 virtual patients for every major cancer [12:27] type and we asked is this combination therapy going to treat each of these individuals and [12:32] we're able to aggregate the response rates across these different virtual cohorts and [12:38] uh actually then prioritize which indication the Nex should go in next. Now we're able to [12:46] do these virtual clinical trials in the order of minutes instead of the order of years. [12:55] Obviously people have not gone out and tried every single one of these combinations [13:00] uh in these indications. But where we do have data is an adnocarcinoma and squamous cell carcinoma. [13:08] And here our model is surprisingly accurate in its predicted response [13:12] rate and it correctly prioritized squamous cell carcinoma over adnocarcinoma as the next [13:18] indication of the generation of one after. And so this gives us confidence that the [13:23] model then can generalize and predict um indication expansion for a given therapy. [13:34] So our vision at synthesized bio is that within our lifetime we'll go from a system [13:38] where 90% of drugs that enter clinic fail to where 90% of the drugs that [13:43] enter clinic succeed and I think that's going to get us there are models that [13:48] make biology itself predictable and I think this we're we're in a very early stage but [13:52] I think that's we're getting there and uh thank you very much for your time.