From Wet Lab to RL Env: 10,000 Drug-Discovery Campaigns in a Weekend — Ian Quigley, Leash Bio
Summary
Leash Bio discusses the convergence of large-scale biological data generation and advanced machine learning for drug discovery. The core methodology involves using DNA-encoded libraries (DELs) to screen millions of small molecules against protein targets, generating massive, reproducible datasets. The discussion highlights the need for epistemic rigor in data collection, emphasizing that the quality of the experimental readout (DNA sequencing) is as critical as the model choice. Advanced models like Hermes use cross-attention mechanisms to predict binding, but the speakers caution that models often learn 'sociology' (e.g., chemist intent or molecular weight trends) rather than true physical chemistry, necessitating complex data splits and engineered mutations to validate physical understanding.
Key takeaways
-
DELs enable massive, reproducible data generation.
3:30
DELs allow screening of up to 40 trillion molecules by attaching unique DNA barcodes to each small molecule. This process is highly scalable and reproducible, providing a robust signal for machine learning training, even when the final drug candidate must be synthesized without the barcode.
-
ML models risk learning 'sociology' instead of physics.
23:20
The speakers demonstrated that models can predict chemist identity or activity based on non-chemical factors (e.g., molecular weight or the chemist's style/intent) when trained on public data, suggesting models may learn data biases rather than fundamental chemical laws.
-
Evaluating ML models requires rigorous data splitting.
30:00
To prevent models from simply memorizing training data (overfitting), the speakers used complex data splits, such as clustering building blocks or testing on entirely orthogonal chemical libraries, which significantly degraded performance compared to training on shared components.
-
The future involves RL agents in a closed loop.
1:36:40
The most advanced approach involves using language models (LLMs) and reinforcement learning (RL) agents to iteratively hypothesize and select the next most informative experiments (e.g., selecting 10 binders and 10 non-binders) to guide drug discovery, simulating a medicinal chemist's process.
Technical details
-
DNA-Encoded Libraries (DELs)
250s
DELs use combinatorial chemistry and DNA barcodes to screen millions of molecules. The process involves binding the protein to a solid support (like a magnetic bead) and sequencing the DNA barcodes of the bound molecules, allowing for high-throughput measurement.
-
ML Model Architecture (Hermes)
3300s
The Hermes model uses a simple architecture combining two embeddings: an Evolutionary Scale Modeling (ESM) embedding for the protein side and a Kvera embedding for the chemistry side. Cross-attention is used to force interaction between the protein and small molecule structures.
-
Activity Cliffs and Protein Engineering
4800s
The concept of 'activity cliffs' involves identifying pairs of molecules or proteins that are structurally very similar but exhibit drastically different binding behaviors. This is exploited by intentionally engineering protein mutations (e.g., changing one amino acid residue) and rescreening with DELs to pinpoint functional residues.
-
Reinforcement Learning (RL) in Drug Discovery
5500s
RL agents are trained to simulate the iterative decision-making of a medicinal chemist. The agent receives a reward signal based on the experimental outcome (e.g., identifying the best binder) and uses natural language reasoning to guide the next set of experiments, creating a 'closed loop' of discovery.
Mentioned resources
- Kemble
- BindingDB
- Polaris
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.