In this episode of Latent Space, Ian Quigley, co-founder and CEO of Leash Bio, explains how his company is revolutionizing small molecule drug discovery by combining DNA-encoded libraries (DELs) with machine learning. They generate massive, high-quality datasets on protein-small molecule interactions and use them to train models that avoid common pitfalls like memorizing chemist identity rather than learning real physics.
Summarized by Podsumo
Kaggle competition revealed that even the best models failed at predicting binding for novel chemical libraries, validating Leash's data collection strategy.
Models often memorize 'chemist identity' or 'molecular weight' rather than learning biophysics, a problem Leash solves by generating engineered protein mutations.
Leash uses descriptive English language models to represent molecules and proteins, which acts as a natural regularizer against memorization.
The company has screened 5,000+ experiments on 2,000 protein constructs, building one of the largest proprietary datasets of small molecule-protein interactions.
By treating data as a fast 'compute speed' lab-in-the-loop, Leash enables AI agents to run thousands of structure-activity relationship (SAR) campaigns per week.
"If the models can identify the chemist by looking at the molecule alone, then maybe the models can infer chemist intent… they don't learn physics."
"You're reading the face, not the math — models memorize sociology, not chemistry."
"We ran a Kaggle competition that was a thousand times bigger than anything anyone had ever seen… no one did better than chance on the hardest eval."