Ian Quigley
Related people
Episodes
Notes
Latent Space: The AI Engineer Podcast · 🔬 Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
-
How Does Leash Bio Run Lab-in-the-loop Without a New Experiment?
Ian Quigley says Leash Bio runs lab-in-the-loop at compute speed by letting a language model test a binding idea against measurements the lab already collected, and get the answer back instantly.
They taught open language models medicinal chemistry in readable English, using activity cliffs: the same chemicals measured across many proteins. The model describes a structure, states the difference between two molecules, and says why the binders might be binding, including hard negatives. It then sees 10 binders and 10 non-binders that are close, writes a reason, and may pick five or 10 molecules from a catalog of millions. It looks those up in data Leash Bio already has and updates the idea. T
-
Why Does the Hermes Binding Model Nickname a Protein?
Ian Quigley says the Hermes binding model nicknames a protein by memorizing a short variable stretch of amino acids, instead of looking at the binding pocket or the conserved domain.
Hermes has cross-attention, so a map can show which amino acids the model uses. The first hope was the binding pocket. A more cautious hope was every residue, because the model did not yet know what mattered. It did neither. It looked at short sequences away from the functional part and away from the stretch that looks similar across many proteins. Biologists use that conserved domain to group families and guess function. Hermes looked outside it, at non-conserved stretches where the letters vary more. He think
-
Why Do Structure Models Put a Drug Back Into a Filled Pocket?
Ian Quigley says current co-folding models still place a small molecule where a pocket used to be after that pocket has been filled in, removed, or broken.
Start with a protein the model has already seen, a pocket, and a chemical sitting in it. Change the amino acid sequence so the pocket is filled, as if with cement, or so the pocket is gone or much bigger. In the trials his group has run with current state-of-the-art models, the model shoves the molecule back into the old spot. He reads that as memory of the location, not as an understanding of the physics. A physical pocket depends on a small number of amino acids with hydrogen-bond positions sticking out, plus a bit of pi stacking. If t
-
What Is a Magic Methyl in Drug Design?
Ian Quigley says a magic methyl is the medicinal-chemistry story that one tiny thorn on a molecule can make it work or make it fail, and he calls that story kind of mythical rather than a law.
He is describing an activity cliff: two things that look almost the same except for a small change that drives an outsized change in behavior, on the protein or on the small molecule. Chemists tell the magic-methyl version as one little thorn that is present or absent, with the molecule either working well or not working at all. If those cases are real, a model has a reason to care about very small edits. DNA-encoded libraries already contain some of that pattern, because two of three building blocks
-
Why Do Binding Models Learn the Chemist Instead of the Chemistry?
Ian Quigley says a model can predict ChEMBL activity from who probably made a molecule about as well as it can from the chemical formula, so it is learning the chemist rather than the chemistry.
At Recursion, a medicinal chemist walked past a screen Ian Quigley was looking at with his co-founder and called the molecule a Stuart Schreiber molecule. If a model can name the chemist from the structure alone, it can infer that chemist's intent and skip the physics. They took every ChEMBL chemist above a bar he remembers as more than 30 papers and 600 molecules. That was 1,815 chemists. A classifier that saw only the molecule picked the chemist with a top-five accuracy of 60%. Confusion showed up
-
Did Anyone Beat Chance on the Hardest Leash Bio Binding Test?
Ian Quigley says that on the hardest split of Leash Bio's public binding contest, nobody did better than chance, and three of the top five entries were a noisy tutorial notebook.
Leash Bio generated the set and released it. BindingDB, then the biggest public collection of small molecules binding protein targets rather than cells, held a little under 3 million physical interactions. The release was about 4 billion interactions, measured by one person in the basement of his house. Xiaoji Lu of AlphaMa in Shanghai provided a library of about 130 million molecules. Every molecule shared a center hub, a motif called a triazine, with three spokes. The contest hid three tests. An easier one shared
-
Which Lab Measurements Can Actually Scale?
Ian Quigley says that if life science is going to be paired with machine learning, the only scalable readout is taking pictures, and sequencing DNA is another way of taking pictures.
Recursion takes pictures of cells, and other groups do that too. Spatial omics and similar methods mix pictures and sequencing, which he still counts as pictures. He expects people who work on proteomics to disagree, and he argues the scale is not there yet. He treats that disagreement as the first hot take of the episode. Sequencing earns the role because it is extremely cheap, well established, and getting cheaper again after a stretch where the price drop plateaued. New companies are pushing the price down.
-
How Many Molecules Can One DNA-encoded Library Hold?
Ian Quigley says the largest DNA-encoded library produced holds 40 trillion molecules, while the biggest one-chemical-at-a-time collections sit in the low single-digit millions.
The slow path puts different chemicals in different tubes and dunks the protein in one tube at a time. A DNA-encoded library puts many chemicals in one tube, each carrying a small DNA barcode. The protein goes in like a fishhook, the binders come back out, and the barcodes are sequenced, which skips the overhead of huge plate runs. He says there are problems when a library gets to 40 trillion. The mixtures are combinatorial. Thousands or tens of thousands of building blocks, squared or cubed depending on the design,
-
Who Hosts Latent Space AI for Science?
RJ Honicky hosts Latent Space AI for Science with Brandon Anderson.
RJ Honicky opens the episode as CTO of MiraOmics and introduces his co-host, Brandon Anderson, who develops RNA therapeutics at Atomic AI. The guest in the studio is Ian Quigley. The opening calls Ian Quigley the co-founder and CEO of Leash Bio, and describes the company as one of the favorite young tech-bio companies of the last few years. The reasons it gives are epistemic rigor, data collection at a large scale, a commitment to open science, and a run of blog posts about how to do machine learning in a hard domain.
Ian Quigley thanks them. He says he and his colleagues founded Leash Bio after trying to predict which sma
-
What Small-molecule Cleanup Does Ian Quigley Want?
Ian Quigley wants small-molecule cleanup that can run at massive scale. The gizmos that purify well, including HPLC, are not highly parallel.
Making a molecule means putting material in a pot and heating it, or using some other conditions. The output lands at an efficiency that might be great or might be poor. Either way, and especially when the result is poor, another set of gizmos has to clean it up. People use HPLC and similar tricks, and none of the really good ones are super parallelizable. His lab runs many molecules through a funky assay already described as giving people lots of trauma. Large numbers of molecules that are clean, purified, and easy to work with could then go into a d
-
Why Does Copying Leash Bio Still Take Time?
Ian Quigley says a better funded lab still needs time to outscale Leash, because the advantage is a long trail of laboratory mistakes, metadata discipline, and processes rebuilt from scratch.
At Leash and at Recursion, his team made many mistakes by bumping into cheating modes that come out of ordinary laboratory work. Much of the defense is tracking batch effects with metadata. The organization has to build a culture and frictionless tools so the people doing the work actually record the details, then go back through the founding history of the company and check that results stay consistent. When big changes appear, the team has to identify what drove them and mitigate them. Leash also reb
-
How Many Proteins Has Leash Bio Screened?
Ian Quigley says Leash Bio has developed and screened about 2,000 different protein constructs and about 5,000 screens in the history of the company, each across at least 6.5 million molecules, and has not raised a lot of money. Chemical material cast across many targets is meant to build a machine understanding of chemical behavior, so early materials can be optimized much faster, especially at the low capital infusions the company has used.
Strong chemical reasoning would support work in agriculture, veterinary science, and material science immediately, without a physical screen in every new domain. Protein-folding teams already chase protein targets that have not been screened yet, or th
-
Why Does a Chemical Library Favor Some Proteins?
Ian Quigley says chemical library design depends on the building blocks, a whole library made in one geometry might be really good for some types of protein targets and really crappy for other types, and some corners of chemical space are probably weak. The spread is in how the library performs across target types.
A design is usually limited to two or three building blocks stitched together. One design run a lot right now is a forked, Y-shaped molecule. Other choices are linear molecules and circular, macrocycle-shaped ones. DNA has to sit somewhere on the molecule, and that placement can cause problems. Commercial availability of building blocks and the chemistries required to stitch them
-
How Does Breaking a Protein Pocket Prove Binding?
Ian Quigley says breaking one pocket with a mutation shows which chemicals were binding there. Chemicals a competitor had already kicked out also disappear once the pocket is broken, while chemicals binding everywhere else on the protein stay intact, so the mutation did not simply destroy the whole protein.
On big weeks the lab screens 100 protein constructs. The libraries are combinatorial, so five different chemical series can come to hundreds of molecules combined, and another group can be the same hundreds of molecules with a single change in one building block. One small side group can shift how a large population behaves, and five independent examples of that shift make a clear case t
-
Did Hermes Match Boltz on Binding Prediction?
Ian Quigley says Hermes did about as well as Boltz, then the state of the art structural model, on public data sets for binding prediction. The models were trained only on internal data and scored external molecules they had never seen. In his count the Hermes paper covered something like 13 million tests across about 500 protein targets.
Large collections of external molecules were predicted, with positives and negatives chosen carefully. Other labs usually screen fewer protein targets, so outside sources are tempting, and he has not seen another group train for the task only on internally generated data. Hermes cannot do a lot of what Boltz can do, and a fair comparison needed some tricks
-
What Is a Small Molecule Drug?
Ian Quigley calls a small molecule a tiny chemical that can sit in a pill, go inside a cell, and affect how a protein behaves. Small molecules make up most drugs, and pills in a house are almost certainly small molecules. His company is a machine learning small molecule company. It collects massive numbers of small molecules binding to protein targets, uses the measurements to drive machine learning, and pushes small molecule material into partners and toward patients.
Trying to drug something starts with a protein of interest in the body and an attempt to affect its behavior. Antibodies, a common treatment already on the market, are made by cells. They are large molecules, big enough that
-
How Reproducible Are Leash Bio Binding Screens?
Ian Quigley says about 90% of chemicals that are easy in their hands still work after they are made off DNA and run in functional assays, and about 50% of the hard ones do.
For machine learning, the chemicals stay in the tube with DNA barcodes. Some bind and some do not, and the binding is signal measured from the world. Models are conditioned on the barcode, with a placeholder that marks where the DNA sits. Leash does not pretend the DNA is missing. Fresh protein batches and a new chemical library, rerun over many months, bring back the same hits in the same order for the most part. The screen is extremely reproducible and robust. Making the molecule without the DNA for a drug is where the
-
How Does Leash Bio Find Binding Chemicals?
Ian Quigley says Leash Bio finds binding chemicals by fixing a protein target on a magnetic bead, dunking the bead in a barcoded mixture, and sequencing the DNA that stays stuck after a rinse.
The target is a protein or a protein complex on a solid support. Leash Bio usually uses a magnetic bead. Amino acid chains click onto the ends of each other, fold, and either bind the target or miss it, and each component carries a barcode. In an easy case the protein has a distinctively shaped hole. A tiny fraction of the chemicals have the shape and electronics to fit the hole. The protein is pulled back out and rinsed a couple of times, so only the chemicals in the hole come along. A little PCR amp