LeashBio releases Hermes model and LLM agents for autonomous chemistry

LeashBio has introduced Hermes, a model designed to predict small molecule binding to protein targets, alongside a new workflow integrating large language models for autonomous reasoning. The Hermes architecture combines ESM embeddings for proteins and ChemBerda embeddings for molecules, using cross-attention to predict binary binding labels. The model was trained exclusively on internally generated data from DNA-encoded library screens, a decision driven by LeashBio’s analysis that public datasets like ChEMBL are heavily biased by the sociological habits of medicinal chemists rather than underlying physics. In benchmarks across approximately 13 million test cases, Hermes performed comparably to the state-of-the-art structural model Boltz, though Boltz outperformed Hermes on specific targets where it had likely seen the exact complex in its training set. To prevent memorization of non-conserved sequences, the team trained Hermes on engineered protein mutations that disrupted binding sites, forcing the model to attend to functional residues.

The company is also deploying large language models, including Qwen and GLM, to perform structure-activity relationship reasoning. These agents operate in a "lab-in-a-loop" environment, proposing hypotheses and testing them against a database of millions of screened molecules. By using natural language to describe molecular behaviors, the system forces causal reasoning rather than pattern matching. LeashBio has extracted over 10,000 such campaigns from its data, a volume far exceeding what a human chemist typically encounters. The company has released its raw data to the public via Polaris to encourage further research, arguing that physical grounding through experimental data is essential for generalizable AI in drug discovery.