From Wet Lab to RL Env: 10,000 Drug-Discovery Campaigns in a Weekend — Ian Quigley, Leash Bio

A DEL is a technology in small molecule drug discovery where a vast number of chemicals are attached to unique DNA barcodes. This allows for extremely high-throughput screening against protein targets by 'fishing out' binding molecules and then sequencing their DNA barcodes, significantly reducing logistical overhead compared to traditional methods.

Leash Bio acknowledges that DNA barcodes are 'hefty' and can influence binding. For machine learning purposes, they condition their models on the existence of the barcode. For drug development, they accept a certain level of attrition when manufacturing chemicals off DNA, with about 90% working for easy cases and 50% for hard ones.

Historically, DELs were seen as traumatic due to the high cost of sequencing and the focus on finding binders for very challenging targets in small numbers. Leash Bio overcame this by screening tons of proteins, characterizing chemical libraries extensively, and leveraging genomics methods development experience to tweak experimental protocols.

The 'Clever Hans' problem describes how machine learning models can learn unintended correlations, such as a chemist's distinctive signature style in molecule design, rather than the fundamental physics of chemical interactions. This leads models to infer 'chemist intent' or 'sociology' rather than ground truth, making many public datasets untrustworthy for pure chemical understanding.

The competition showed that models struggled with out-of-distribution molecules, especially when no common building blocks or central hub existed between training and test sets. This validated Leash Bio's hypothesis that current models were not very good at generalized prediction and underscored the need for more diverse data and careful evaluation to avoid memorization.

Hermes is a machine learning model developed by Leash Bio. It uses two simple embeddings: an Evolutionary Scale Modeling (ESM) embedding for proteins and a Kvera embedding for chemistry (predicts SMILES tokens). These are combined with cross-attention to predict binary binding labels, and it's designed for rapid experimental iterations and resistance to memorization.

Leash Bio experimentally engineers mutations in the binding regions of proteins and then re-screens them. By observing how small molecule binding behaviors change with these mutations, they encourage models to focus on the actual physical interactions in the binding pocket rather than 'nicknames' or non-conserved regions of the protein, preventing memorization of superficial features.

Leash Bio is transitioning to language models to represent small molecules and proteins in readable English, including their behaviors and differences. They use these models as reasoning agents in a 'lab-in-a-loop' setup, allowing the AI to propose hypotheses, 'run experiments' against collected data, and refine its understanding of optimal binders, significantly accelerating drug discovery campaigns.

DELs can be constrained by building block availability and chemical geometries, often leading to flatter or Y-shaped molecules. While Leash Bio optimizes for drug-likeness and expands chemical space by adding new libraries and diverse building blocks, there are still corners of chemical space that are hard to access, necessitating ongoing innovation in library design.

The 'Kosari Paradox' questions the business model of ML in drug discovery: if you need vast amounts of high-quality data to train a good model, why not just use that data to find drugs directly without the model? Leash Bio argues that models enable generalization across all targets, faster optimization of early materials, and broader applications beyond drug discovery, like agriculture or material science.

Beyond cheaper sequencing, Ian Quigley would wish for better, super-parallelizable tools for purifying small molecules at a massive scale. This would allow for clean, purified compounds to be used in orthogonal assays, enabling robust validation of measurements and accelerating the next stages of drug development.