AI FOR SCIENCE · INTERVIEWS
Quigley says that on medicinal-chemistry data a model can look predictive while having learned only who made the molecule and a 'nickname' for the protein. Leash tries to close those shortcuts with large DNA-encoded-library (DEL) screens, pocket mutants and controls, then turns the measured results into environments where language-model agents practise structure-activity (SAR) reasoning.
The interview strings Leash's method into one line: Clever Hans (1,815 chemists, authorship guessable from the molecule alone), the BELKA competition (no one beat chance on a new chemical library), Hermes, pocket mutants, and then the agent sandbox, released the same day as Minos. On scale, he says ten people have made about 2,000 protein constructs and run about 5,000 screens, each across at least 6.5 million molecules.
The performance claims are almost all Leash's own evaluations, with no outside group yet checking them in a new physical experiment; Minos covers only molecules already screened and only binding, not pharmacokinetics or toxicity. No development candidate, its target or a partner is disclosed, the inflammation programme has reached mice only, and DEL reads cannot yet be turned into affinities (Kd).
When Quigley and co-founder Andrew Blevins were at Recursion, a medicinal chemist walked past a molecule on screen and said it looked like a Stuart Schreiber molecule. If a model can recognise the chemist, he reasoned, it may be inferring the chemist's intent, not physics. They took every ChEMBL chemist with at least 30 papers and 600 molecules, 1,815 of them, and trained a classifier to predict the author from the molecule alone, reaching about 60% top-5 accuracy. He says a chemist's style passes on to students and that reactions that work get reused, so these signatures are what a model picks up.
“If the models can identify the chemist by looking at the molecule alone, then maybe the models can infer chemist intent.”Ian QuigleyTwo years ago Leash put about 4 billion measurements into a Kaggle competition (BELKA). It had three evaluation tiers, with shared building blocks, with a shared core but different blocks, and an entirely different chemical library, none announced to entrants. About 2,000 teams competed and no one beat chance on the hardest tier; he says three of the top five were the noisy tutorial notebook. He says they gave it a 5% chance that co-folding models would make the company redundant, then released all the data and invited anyone who thinks they are good to show it.
Hermes is deliberately simple: an ESM embedding for the protein, a ChemBERTa embedding for the molecule, cross-attention, and a bind-or-not output. He says that as splits move further away XGBoost steadily drops while Hermes stays roughly flat. The Hermes preprint gives the evaluation's size: about 13.8 million examples in four sets, covering 164, 59, 403 and 26 targets; on these Hermes performs close to Boltz, then the leading structure model, each better on some protein families, and the post itself says it does not claim Hermes is the better model. He cites BRD4: the same molecule is in the PDB, Boltz called it perfectly and Hermes badly, and he suspects Boltz had trained on that complex.
One DEL problem is that a protein may bind the DNA. Leash gives each molecule 32 or 48 different barcodes; if enrichment differs widely between barcodes for the same molecule, that suggests the protein is binding DNA, not the molecule. On the pocket side, he says current co-folding models still place the molecule where a digitally broken pocket used to be, and public structure data lacks such negatives. Leash mutates the pocket physically: a competition screen with a known ligand, then a swap of the pocket's amino acids and a rescreen, to see whether the same molecules drop out. Hermes first looked at short non-conserved stretches of the protein, a 'nickname'; trained with mutants its attention became 'much more even', though he gives no evidence beyond the attention maps.
Leash taught open-source language models to describe molecules and the differences between them in English, then gave them 10 binders, 10 non-binders and a budget of 100 molecules to 'run experiments' among about 6.5 million, with results returned at once from data already measured, like a coding sandbox. He says the data holds choices a human would not make, so an agent can make mistakes and get feedback; one weekend they pulled more than 10,000 SAR campaigns out of it, against the few dozen a medicinal chemist sees in a career. The Minos post of the same day names it: the flagship task has no starting points, a budget of 100 and 33 targets, about 100,000 trajectories so far, and a release of some environments and thousands of reasoning traces. It also lists how agents fail: staying too long in low-potential regions, retesting the same molecule, not using the whole budget, and failing to update after a contradiction.
“It's got to have the counterfactuals. It's got to make mistakes.”Ian QuigleyHe says that made off DNA about 90% of hits on easy targets work and about half on hard ones; DEL molecules tend to be flat and depend on the building-block catalogue, so some chemical space will be done badly; and reads cannot yet be turned into Kd. Screening leaves starting material that looks selective for about 700 targets, and the direction he names is cytokine targets now drugged only by antibodies, reached instead with a pill. An inflammation programme went from library hit to mouse PK/PD in six months, and the company is talking about partnering such programmes out into new companies.
Source: the Blevins and Quigley preprint (arXiv 2512.20924); abstract, methods and both result tables read. The figures are the paper's own, measured on one curated ChEMBL slice and averaged over five random scaffold splits.
Commentary recorded, on 2026-10-14, Quigley's argument from this episode about ChEMBL: public data teaches models who made a molecule.
The interview gives the whole ladder of tests: quantify the authorship shortcut first (1,815 chemists, about 60% top-5), then test generalisation on BELKA's new chemical library and Hermes' outside sets, then break the protein-side shortcut with physical pocket mutants.
The timeline records OpenBind-0 (2026-08-21): a fully open co-folding model that rivals AlphaFold3 on standard benchmarks, while its top-25 success falls to single digits on flexible targets.
Quigley offers a test from the experimental side: physically mutate the pocket and see whether the molecules leave. He says current co-folding models still place the molecule where a broken pocket used to be, without naming which models or papers.
In interviews on this site, Lila's CTO uses real experiments as the reinforcement-learning verifier (Latent Space, 2026-07-16) and Periodic's co-founder says the deciding superconductivity measurement takes about an hour a sample (Latent Space, 2026-10-08); in both, feedback waits for the experiment to run.
Quigley proposes a third way: measure the whole experimental space first and let agents get feedback from it at once, with the outcomes of the molecules they did not pick also known; the price is choosing only among molecules already screened, and only binding.
A larger dataset does not by itself bring better chemical understanding. Leash's steps, authorship style, BELKA and pocket mutants, all do one thing: build controls that make a shortcut useless. Reading a scientific model's score starts with whether its test has been attacked this way, not with the accuracy alone. That is our reading.
Public structure and binding data cover negatives, failed experiments and untried choices thinly. Because Minos measured the whole space in advance, it shows not only what an agent picked but what the ones it skipped would have returned, so its decisions can be judged: how long it stays in a poor region, whether it changes its mind after a contradiction, whether it spends its budget. For closed-loop science that points at an evaluation question: are we scoring the final answer, or how an agent uses evidence and budget to choose the next step? That is our reading.
Lila and Periodic have agents direct real experiments, each step taking days to months; Minos runs the experiments in advance, so each step is a lookup. It does not simulate the rules of chemistry but replays a finite space already measured, so it can only choose among molecules already screened and sees binding only, not pharmacokinetics or toxicity. Releasing the environments and traces is not independent verification.
About ten people and little money raised; some 2,000 protein constructs and 5,000 screens to date, each across at least 6.5 million molecules; 100 constructs screened in a big week.
Company-reported
His account, with no public document; he estimates about $100K a screen at a contract research organisation.
From 1,815 ChEMBL chemists, a classifier predicting the author from the molecule alone reaches about 60% top-5 accuracy.
Company-reported
Matches the '1815 chemists' in Leash's February 2026 post; the paper was not read here, and the explanation, a signature style, is his conjecture ('we think it is very likely').
BELKA: about 4 billion measurements and about 2,000 teams, none beating chance on an entirely different chemical library.
Company-reported
The competition and data are public; this page did not reread the results, and the figures are his.
On Leash-built evaluations (four sets, about 13.8 million examples, covering 164, 59, 403 and 26 targets), Hermes, trained only on Leash's data, performs close to Boltz, each better on some protein families.
Company-reported
The size comes from Leash's post; he says about 13 million and that he does not recall it exactly. Leash ran the comparison and the post says it does not claim Hermes is better; the paper was not read.
On one BRD4 molecule Boltz was perfect and Hermes poor, because the complex is in the PDB.
Guest's judgment
Quigley suspects Boltz had trained on it ('we suspect'); no check of overlap with its training data was made.
Each molecule carries 32 or 48 different barcodes; large differences in enrichment between barcodes suggest the protein is binding DNA.
Company-reported
A description of method; he adds that enriched sequences can recover a transcription factor's known DNA motif, as a sign the protein was made right.
After training with pocket mutants, Hermes' attention no longer concentrates on short non-conserved stretches and is 'much more even'.
Company-reported
From attention maps shown on screen, with no figures; more even attention does not show the model understands the binding mechanism.
Current co-folding models still place the molecule where a digitally broken pocket used to be.
Industry context
He relays others' work ('a couple different groups have shown this') without naming it.
Given 10 binders and a budget of 100, agents find the best binders among about 6.5 million.
Company-reported
Recorded before the preprint; the same day's blog defines the flagship task as no starting points, a budget of 100 and 33 targets, and says stronger models can do it 'at least a little bit', weaker than the interview's wording.
Some Minos environments and reasoning traces are released; about 100,000 SAR trajectories and about 2 new environments a day.
Company-reported
From Leash's own blog; release is not independent verification, and the preprint was not read here.
Made off DNA, about 90% of hits on easy targets still work, about 50% on hard ones.
Company-reported
'In our hands', he says, adding that others report differently; what counts as working is not defined.
An inflammation programme went from library hit to mouse PK/PD in six months with blood cytokines moving as expected; starting material looks selective for about 700 targets.
Company-reported
No development candidate, target or data disclosed; mice only; he hedges the 700 ('I think like 700') and does not say how selectivity is judged.
DEL reads give a relative ranking but cannot yet be turned into Kd; SPR measurements are being added to build a potency ladder.
Not yet closed
'I won't say that we're good at it yet,' he says.
The only scalable measurements in the life sciences are imaging and sequencing; a better-funded lab could not quickly replicate this, because the moat is a culture and tooling for chasing down experimental cheating modes.
Guest's judgment
He calls the first a hot take and expects proteomics people to disagree; the second is his view of his own company's advantage.
Stronger models abandon unproductive neighbourhoods faster; weaker ones keep tweaking the same few molecules.
Company-reported
Said in the follow-up episode with no figures; he says he does not know why.
Bayesian optimisation was not used as a baseline because the landscape is too spiky and too high-dimensional.
Guest's judgment
His judgment, untested, so 'stronger models search better' has not been set against a traditional search method.
Medicinal chemists who looked at transcripts found the agents' hypotheses plausible; no person was run on the same task.
Company-reported
The impression of a few experts, 'not a ton', with no figures; he says a human version would need far fewer building blocks.
About 1,000 proteins screened; two new environments a day now, with twenty a day wanted.
Target
He says twenty 'without too much lifting'; the date is unclear in the captions ('26 or 27') and is not set as a deadline.
Whether a group outside Leash evaluates or trains agents on the released Minos environments and tests them in a new physical campaign.
Would show up as An outside paper naming the environments used, hit rates on a new target or chemical library, and a baseline without Minos.Whether a model trained only on Leash's DEL data beats chance on someone else's chemical library.
Would show up as A public score on BELKA's hardest split or a comparable outside split, run by someone other than Leash, with a stated check of overlap between training and test data.Whether the inflammation programme or another yields a named development candidate or a partnered new company.
Would show up as An announcement or paper naming the target, the molecule class and the mouse PK/PD data.Whether DEL reads can be calibrated to affinities.
Would show up as DEL enrichment set against SPR-measured Kd for the same molecules, with the correlation and the exceptions stated.Whether, in Minos, language-model agents choose molecules better than a search method that uses no language model.
Would show up as Results for Bayesian optimisation or a similar method on the same environments and the same 100-molecule budget, set beside the agents' and saying who ran them.Whether reinforcement-learning progress made in these environments carries to a biology problem it has not seen.
Would show up as Experimental results on a new target or chemical library, against the same model not trained on Minos.Later evidence that tests or limits what was said, and targets whose date has passed.
The interview was recorded about two months earlier; the day it went out, Leash released Minos as a blog post, a preprint and a partial data release. Against the interview, the sandbox now has a name and a fixed flagship task, trajectories have gone from 'more than 10,000 in a weekend' to about 100,000 in six weeks, and the claim is more careful, that stronger models can do it 'at least a little bit'. These remain the company's own results.Leash Bio, 'Making reality replayable' ↗
Latent Space added a 16-minute follow-up with Quigley after Minos was released. Against the main interview: trajectories are now about 90,000; stronger models leave unproductive chemical neighbourhoods faster while weaker ones keep tweaking the same few molecules, and he says the reason is not known; he did not use Bayesian optimisation as a baseline because he thinks the landscape too spiky and too high-dimensional, 'mostly think it's not going to work'; a few medicinal chemists looked at transcripts and called the hypotheses plausible, but no human was run on the same task; about 1,000 proteins have been screened, two environments are made a day now and he wants twenty; whether reinforcement-learning progress made in these environments carries to new biology he says is not yet known. These remain the company's own statements.Latent Space, 'An Eval for AI Scientists That Knows the Answer Key' ↗
Read 2026.10.15 The Substack machine transcript attached to the Latent Space episode page, with numbered but unnamed speakers, read in full; the audio and video were not heard. The hosts are Brandon Anderson (Atomic AI) and a co-host. The episode page says it was recorded about two months before release, so 'recently' and 'last week' mean around August. Figures were checked against Leash's own blog, read in full: the Minos post of the same day, the February 2026 Hermes post and the December 2025 post on protein 'nicknames'; the Hermes preprint of 9 February 2026 was read, and so were the abstract, methods and result tables of the Clever Hans preprint (arXiv 2512.20924). Timestamps are the machine transcript's, from a recording of the same length as the YouTube episode (1:49:56); they were not checked against YouTube playback. The 16-minute follow-up of 2026-10-26 was read in full as the timestamped YouTube-caption transcript the author supplied; its audio was not heard. Open the interview ↗Latent Space episode page (with transcript) ↗