Leash Bio

Leash Bio builds experimental datasets for predicting whether small molecules bind proteins. Its approach starts in the laboratory: screen many DNA-tagged molecules together, use sequencing to measure which ones associate with a target, and design controls that help distinguish real chemical binding from artifacts.[1]

In a DNA-encoded library, a chemical carries a DNA barcode that records its identity. Leash exposes a pooled library to a target protein, recovers the bound material and sequences its barcodes. Enrichment relative to the starting pool suggests which chemicals may bind. Combinatorial synthesis also creates related compounds, giving the team ways to examine what happens when a small part of a molecule changes.[1]

The barcode can interfere with the measurement: a protein may stick to the DNA rather than the chemical. Leash describes putting dozens of barcodes on the same chemical and looking for sequence-specific enrichment. It also uses negative controls, including beads without the target, to detect sticking to the experimental apparatus and other nonspecific effects.[1]

The company built a Kaggle competition around a chemical library of about 130 million molecules and evaluated models on several increasingly unfamiliar splits. On the hardest split, a separate library had no shared central scaffold or building blocks and used different attachment chemistry. Leash reported that, among roughly 2,000 teams, nobody beat chance on that split. That result is specific to this evaluation; it does not establish that all predictions on new chemistry fail.[1]

Leash also investigated a shortcut in public chemistry data. In an analysis of ChEMBL, the team predicted which chemist likely made a molecule and represented the molecule by those probabilities. In its reported activity prediction test, that representation performed about as well as a representation based on chemical formulas. The exercise shows why models can look useful while picking up patterns in what researchers chose to make and measure.[1]

Its Hermes model combines protein and chemical representations to predict binding, and the company has tested whether it attends to the relevant binding pocket by changing protein residues. These are ways to probe model behavior, not proof that every learned prediction is mechanistic or will transfer to every target.[1]

By October 2026 Leash had begun using its accumulated measurements for an agent sandbox. A model receives examples of binders and non-binders, chooses candidates to test, and gets answers by looking up experiments already run. It can revise its hypothesis and select again within a budget. The immediate loop is restricted to molecules already screened; ordering and testing new choices was described as a future extension.[1]