Why 2,000 ML Teams Failed to Predict Molecular Binding
Ian Quigley shares how Leash Bio's BELKA Kaggle competition proved AI models memorize chemical fragments instead of learning physical binding.
From Wet Lab to RL Env: 10,000 Drug-Discovery Campaigns in a Weekend — Ian Quigley, Leash Bio
Ian Quigley, co-founder and CEO of Leash Bio, explains how his company combines DNA-encoded libraries (DELs) and high-throughput sequencing to generate billions of reproducible protein-small-molecule binding measurements. He shares key lessons from Leash's BELKA Kaggle competition, exposes how public chemistry databases like ChEMBL cause machine learning models to learn chemist sociology rather than physical binding principles, and details how Leash built Hermes and attention-mapping techniques with engineered protein mutations. Finally, Quigley reveals how Leash converts its experimental chemical landscapes into lightning-fast reinforcement learning sandboxes for LLM reasoning agents to conduct thousands of structure-activity relationship (SAR) medicinal chemistry campaigns.
Key takeaways
- A single researcher working in Ian Quigley's basement generated a four-billion-measurement dataset using DNA-encoded chemical libraries. Read more →
- Leash Bio trained a classifier on 1,815 prolific chemists in the public ChEMBL database and hit 60% top-five accuracy predicting which scientist built a given molecule. Read more →
- Sri Kosuri challenges the machine learning drug discovery model: if training a predictive network requires testing millions of physical molecules, you should simply develop the physical hits and skip the model. Read more →
- Hermes paired ESM protein embeddings with ChemBERTa small molecule embeddings via cross-attention, but attention maps exposed sequence memorization rather than physical binding. Read more →
- Public drug discovery databases like ChEMBL train machine learning models on human bias rather than physics because they reflect which molecules historical chemists preferred to synthesize. Read more →
- Public chemistry repositories like ChEMBL train machine learning models on human publication habits rather than physical binding mechanics. Read more →
6 articles from this episode
Why AI in Drug Discovery Predicts the Chemist, Not the Physics
Leash Bio found AI trained on ChEMBL predicts chemists, not binding affinity. Here is why your ML model might just be learning human habits.
The Kosuri Paradox: Why Train AI When Wet Labs Have the Hits?
Ian Quigley addresses Sri Kosuri's paradox on why AI drug discovery needs machine learning if wet lab screens already generate direct chemical hits.
Why Leash Bio Mutates Protein Pockets to Stop AI Memorization
Ian Quigley explains how Leash Bio caught Hermes memorizing protein nicknames and engineered pocket mutations to force true biophysical learning.
Why Leash Bio Puts 48 DNA Barcodes on a Single Molecule
Ian Quigley explains how Leash Bio fixes noisy DNA-encoded libraries and stops machine learning models from learning chemist sociology.
How Leash Bio Runs 10,000 Chemistry Campaigns in One Weekend
Ian Quigley explains how Leash Bio turns billions of wet-lab binding points into an instant reinforcement learning sandbox for LLM reasoning agents.
More Latent Space episodes
- Robots That Predict the Future: Dyna-2 & Physical AI — Jason Ma, Dyna RoboticsOct 18 · 4 articles
- How to kill your SaaS: what we learned replacing a $43K/yr tool with AI agentsOct 18 · 4 articles
- The Bitter Lesson of Biology: Why Scaling AI Isn't Enough — Pushmeet K, DeepMind & Sal C, BiohubOct 11 · 4 articles
- AI Scientists Are Here: Autonomous Labs & Synthesis Superintelligence — Periodic LabsOct 11 · 5 articles