Latent Space

Ian Quigley on why drug-discovery AI learns sociology, not physics

Ian Quigley· Co-founder and CEO of Leash Bio at Leash Bio
·~110 min·English·Latent Space
AI CompanyTrainingReasoningOpen Source
TL;DR

Leash Bio's Ian Quigley argues AI drug discovery is bottlenecked by data, not models: public chemistry datasets teach a model which chemist made a molecule rather than the physics of how it binds, so Leash generates its own causal data at scale and lets reasoning agents explore it.

01Core Mental Model

Models Learn Sociology, Not Physics

Trained on public chemistry data, a model can identify the human chemist who made a molecule from its structure alone, which means it is learning the fingerprints of the people, not the physics of binding.

We really think that it's learning a lot of sociology. It's not learning ground truth.

— Ian Quigley, Latent Space
Key Insight
Both routes reach the same score on familiar data, so a leaderboard cannot tell them apart. The sociology route only fails on a molecule from a chemist the model has never seen, which is exactly the case that matters for discovering a genuinely new drug.
02The Evidence

Two Thousand Teams, No Better Than Chance

Leash ran a Kaggle competition with three tiers of difficulty, and on the hardest tier, an entirely unfamiliar chemical library, none of the 2,000 teams beat random guessing.

And then when you go to the most challenging eval, no one did better than chance.

— Ian Quigley, Latent Space
Key Insight
A random train-test split would have hidden this, because familiar building blocks let a model score well by memorizing. Leash's tiered design isolated true generalization and showed the field does not yet have it, which is a diagnosis of the data far more than of the contestants.
03The Fix: Make Your Own Data

Four Billion Measurements, Made in a Basement

Rather than trust public data it cannot vouch for, Leash generates its own at enormous scale using DNA-encoded libraries, where every molecule carries a DNA barcode so a single sequencing run reads out millions of binding results at once.

And so we released a data set that was about four billion physical interactions and it was all measured by a single person in the basement of my house.

— Ian Quigley, Latent Space
Key Insight
The scale is striking, but the real move is control: because Leash designs and runs the experiment, it knows the negatives, the controls, and the batch conditions. That provenance is what makes four billion of its measurements more trustworthy than three million aggregated from strangers.
04Killing Memorization

Break the Pocket to Prove the Physics

Leash mutates a single amino acid to deliberately break a protein's binding pocket and re-runs the screen, so that the exact chemical series that disappears is causal proof of where it was binding, a test a memorizing model cannot fake.

If you use current state of the art models to go see what the model will do, it will shove the small molecule in where the pocket used to be.

— Ian Quigley, Latent Space
Key Insight
Attention maps revealed why this matters: Hermes had learned a shortcut, memorizing a short, variable stretch of each protein as a nickname rather than the binding site. Engineered mutations remove the shortcut by making nearly identical proteins behave differently, forcing the model to look at the chemistry instead.
05Where It Is Going

A Physically Grounded Playground for Agents

Leash turned its screened data into a fast reinforcement-learning environment where a language-model agent proposes a binding hypothesis in plain English and tests it as an instant database lookup, running thousands of optimization campaigns in a weekend.

right now we give these agents 10 binders, give them a budget of a hundred and they'll get the best binders out of six and a half million in that budget.

— Ian Quigley, Latent Space
Key Insight
The stored screens are not just training data but the environment the agent acts in, which is why each test resolves instantly and the reward stays physically grounded. Quigley's bet is that reasoning in plain English, which he calls a regularizer, resists the memorization that beat earlier architectures.
06The Business Model

If You Have the Data, Why the Model?

The paradox is that if you can collect enough data to train a model, you may already have the chemical material you were after, so the model only earns its keep when it generalizes to targets, and even whole industries, you never screened.

if we get really good at chemical reasoning then we can get really good at lots of different industries and their applications without having to physically go and screen in all those different domains.

— Ian Quigley, Latent Space
Key Insight
This reframes the asset. The screens are the cost of training, but the durable product is a model that prices in chemical reasoning once and then spends it anywhere, which is why Leash can describe the same engine reaching into agriculture or materials without new lab work.
07The Payoff

A Pill Instead of an Infusion

The prize Leash is chasing is a small molecule that hits the same target as a proven antibody drug, turning a cold-chain infusion into a pill that can enter cells and reach patients far more cheaply.

folks are using antibodies to drug them and going after a small molecule therapeutic that you could take as a pill instead of an infusion

— Ian Quigley, Latent Space
Key Insight
Many high-value targets already have a working antibody, so efficacy is not the question; delivery is. A pill is far cheaper to make and ship than a cold-chain infusion, so converting those targets into small molecules is where Leash's ability to screen hard-to-drug proteins would matter most to patients.
08The Moat

The Edge Is Process, Not Scale

Asked what stops a better-funded lab from simply out-scaling it, Leash's answer is that the moat is years of chasing down the subtle cheating modes in the data, work that depends on culture and attention rather than raw compute.

it takes kind of a unique culture and patience and willing to make these mistakes and then have the team like really think very deeply about how to address them.

— Ian Quigley, Latent Space
Key Insight
A telling example: Leash noticed chemicals in its own library that resemble lumacaftor, a cystic-fibrosis drug that acts on misfolded proteins, lighting up as hits. The team turned that accident into a signal that a protein likely misfolded and its screen should be rerun, the kind of hard-won instinct that raw scale does not confer and a rival would have to rediscover the slow way.