Leash Bio’s Lab in the Loop at Compute Speed

Millions of molecules in one tube

A useful drug candidate has to do more than bind a protein, but binding can be a productive place to start. In a Latent Space conversation with Leash Bio’s Ian Quigley, the company’s approach begins by measuring huge numbers of small-molecule interactions, then uses those measurements to train models and test scientific questions.

A DNA-encoded library puts many different small molecules in the same tube. Each molecule carries a DNA barcode that identifies its chemical construction. Add a target protein, recover the molecules retained with it, and sequence their barcodes: the readout tells the experimenters which library members were enriched. This replaces a separate physical well for every candidate with a pooled screen.

The library is built combinatorially from a smaller collection of chemical building blocks. Related candidates can differ in just one component, making it possible to compare which changes accompany stronger or weaker measured binding. That convenience comes with a constraint: the space of molecules depends on the available building blocks and chemistry.

Controls turn a signal into a useful measurement

Picture a protein attached to a magnetic bead and dipped into the mixture. After the bead is pulled out and washed, the retained barcodes are amplified and sequenced. Enrichment is a comparison, so a bead without the target and other negative controls help reveal molecules that stick to the support or to proteins generally rather than to the target of interest.

There is another failure mode: a protein may bind the DNA tag instead of the small molecule. The episode describes assigning 32 or 48 different barcodes to copies of the same chemical. If only particular barcodes are enriched, their DNA sequences can reveal a tag-binding preference; agreement across tags gives a different pattern.

The tag also matters when a hit is taken forward. The model can be trained with the presence and position of the DNA attachment represented, but a drug candidate must eventually be made without it and tested again. The interview reports that about 90% of easier examples and about 50% of harder ones worked in those off-DNA follow-ups in the team’s hands; those figures are context-specific, not a general success rate.

Repeated screening provides another diagnostic. Leash describes screening each library against many proteins and noticing that separate productions of an ostensibly identical library behaved differently. That comparison helps surface manufacturing or chemistry quirks that a single screen might miss. The goal is reproducible measurement across both chemical and protein variation.

A harder test than a random split

Leash’s Kaggle challenge asked whether models could predict held-out interactions across increasingly unfamiliar chemistry. One split shared a building block with training molecules; another held out the blocks but retained the same central chemical hub; the hardest used a separate library with a different hub, blocks and attachment chemistry.

The conversation reports that the winning entry did reasonably on shared building blocks, with target-dependent performance on the shared-hub split, while no team beat chance on the different-library test. That result concerns this particular held-out setup; it shows why a good score among familiar variants is not the same as proven generalization to new chemistry.

There is a practical budget decision behind those splits. With fixed sequencing resources, a team can favor more chemical candidates, more protein targets, or better characterization of the combinations already measured. The episode describes choosing broader variation on both axes and more carefully characterizing a smaller library rather than relying only on a huge number of chemicals against a few proteins.

What a model might learn instead of chemistry

A model can exploit patterns in how data was produced. The conversation describes a test on ChEMBL in which a classifier predicted the likely chemist from the molecule; a representation made from predicted creator probabilities was then used to predict reported activity. In that reported analysis, it performed about as well as a representation based on chemical formulas. The concern is that an activity model can learn which candidates people chose to make and test, not only how molecules interact with proteins.

The choices have a plausible path into the data: a chemist may repeatedly elaborate a promising series or favor a building block that worked before. Models trained on the resulting records can pick up those decisions. The experiment raises a concrete shortcut risk for the datasets studied.

Hermes, variants and broken pockets

Hermes began as a small, fast experimental model. The episode describes combining a protein embedding with a chemistry embedding using cross-attention, then predicting a binary binding label. Cross-attention offers a way for information about parts of the molecule and protein to interact.

The team tested chemical splits that held out clusters of similar building blocks, not just identical blocks. It reports that a simple XGBoost baseline degraded as the splits grew more distant while Hermes was steadier across those regimes. In a separate comparison on public data, the discussion says Hermes trained on internal data performed about as well as Boltz on the particular binding task, while acknowledging that the models have different capabilities and may have seen different training distributions.

The protein side has its own shortcut. In one analysis, Hermes attention appeared to concentrate on variable stretches outside the expected functional region. The team’s interpretation was that the model could use those stretches as a protein “nickname” instead of attending to the pocket. Attention maps are a clue for investigation.

To challenge that shortcut, the team describes physically changing amino acids in a binding pocket and running the screen again. A useful cross-check first adds a known competing molecule: candidates displaced by the competitor are compared with candidates that disappear after the pocket mutation. If other binding series remain, that provides evidence against the simple explanation that the whole protein was broken. The team also observes that some mutations can increase binding for other molecules.

This creates examples of activity cliffs: small changes to a molecule or protein can cause large changes in observed binding. Similar chemical series and repeated protein variants help locate the differences worth investigating. After adding mutant data, the team reports attention becoming more evenly distributed; that observation alone does not establish complete physical understanding.

A lab-in-the-loop sandbox

The later experiment changes the model’s job from predicting a single label to choosing what to test. The team describes giving a language-model agent examples of binders and non-binders, asking it for a hypothesis in plain language, and allowing it to select a few molecules from a catalog to test that hypothesis. Results return, the agent updates its hypothesis, and the cycle repeats under a limited budget.

The “lab” in this loop is a lookup into measurements already collected, rather than a newly run wet-lab experiment at each step. The episode describes starting with ten binders and ten non-binders, allowing 100 selections, and reports finding strong binders among 6.5 million candidates in that setting. The reachable choices are limited to what was screened; the discussion proposes testing new catalog choices physically in future.

The appeal is speed and repeatability: a large existing measurement corpus can supply many feedback sequences, including poor choices a chemist may never have selected. The team hopes natural-language reasoning will reduce some memorization, but that is a research hypothesis. Fast access to a recorded answer does not by itself establish that the agent generalizes beyond the measured library or makes a safe drug.

From binding to a medicine

A binder still has to survive a much longer development process. The episode reports an inflammation-program series moving from library hits to mouse pharmacokinetic and pharmacodynamic results in six months: the team measured how the compound reached tissues and tracked cytokine markers. That is a reported early result in mice, not evidence of safety or efficacy in people; further manufacturing and safety work remains.

The same measurement discipline limits what can be inferred from a DEL screen. Repeated screens may recover hits in a similar order, but the team says sequencing reads are not yet a reliable direct conversion to a binding constant, and it is exploring separate biophysical measurements. Its wish for massively parallel purification would make it easier to test many compounds in a cleaner, independent assay.

The broader bet is that measured examples across diverse molecules and proteins can help models optimize chemical material or generalize to new problems. The interview also acknowledges the tradeoff: if one screen already produces a useful candidate for one target, screening itself has created value. The model’s additional value depends on whether it can reliably help with the next choices and targets beyond what was directly measured.