Latent Space: The AI Engineer Podcast
Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
How Leash Bio tamed the notoriously frustrating DEL (DNA-Encoded chemical Library) assays, used that data to catch models “cheating” through shortcuts, and why this data is uniquely suited for training AI agents to think like chemists.
3 hours ago
A few years ago, I stumbled upon Leash Bio’s Substack and was instantly hooked. From their Kaggle competition (which proved that we haven’t actually solved protein-ligand binding at all), to their striking discovery that simply knowing who authored a molecule might be enough to predict its activity. Their Substack—or at least Owl Posting1’s deep dive on them—should be required reading for any scientist attempting to apply AI to chemistry.
So I was thrilled to finally sit down and learn from Ian Quigley of Leash Bio. Leash focuses on DNA-encoded chemical libraries (DELs), a method capable of gathering protein–small molecule interaction data at massive scale.2 Prior to learning about Leash, all I had heard from DEL practitioners were horror stories and complaints. Yet Ian and his team somehow tamed this beast, extracting clean, reproducible data that serves as prime material for training large-scale models. Without giving away too many spoilers, the secret comes down to epistemic rigor and an unwavering determination to “Do The Thing.”
And the best was yet to come! Toward the end of the episode, Ian gave a sneak peek into a new direction his team was experimenting with at the time of recording: an AI playground for verifiable scientific reasoning by agents. They realized that the data collected over the past few years could simulate tens of thousands of “lab-in-the-loop” iterations—exactly what agents need for scientific reinforcement learning (RL). Two months later, that concept matured into Minos, and today they released the full preprint paper, blog post, and open-source dataset. We’ll also be dropping a quick follow-up pod in a few days covering the latest developments, so stay tuned!
We had an absolute blast talking with Ian, and this is easily one of our most fascinating episodes to date. Whether you’re a newcomer to chemistry or a seasoned veteran in AI x Bio, we’re sure you’ll find plenty of takeaways.
Enjoy watching/listening!
Here is what we covered:
- Why DELs cause so much pain for anyone who has ever worked with them, and how Leash attaches dozens of barcodes to each molecule to filter out proteins binding to DNA rather than the drug.
- Ian’s scaling law of biology: biology scales in only two ways—taking pictures (imaging) and sequencing DNA.
- The Kaggle competition: ~4 billion measurements generated in a basement, drawing ~2,000 teams, none of which could beat random guessing on unseen chemical libraries. Plus, why Leash quietly worried co-folding models might make their work obsolete.
- The Clever Hans effect in ChEMBL: models can infer which chemist made a molecule just from its structure, and “who made it” predicts activity nearly as well as the structure itself.
- How Hermes, trained solely on Leash’s proprietary data, kept pace with Boltz on public benchmarks. Plus, the BRD4 molecules Boltz had already seen in the PDB.
- Breaking the pocket: after discovering Hermes had memorized proteins’ “nicknames” rather than learning true binding pockets, they embarked on massive protein mutagenesis campaigns.
- How Lumacaftor-like hits inadvertently turned into detectors for misfolded proteins.
- The Agent Sandbox: LLMs reasoning about structure-activity relationships (SAR) in natural language to search for the best binder out of 6.5 million molecules on a budget of just 100 experiments.
- The Recursion mafia, and why companies founded by its alumni share similar traits.
- The Kosuri paradox: if you can just gather the data directly, why do you need a model at all?
- How an anti-inflammatory pipeline went from a DEL hit to solid in vivo mouse PK/PD data within six months, and the rationale for pursuing oral small molecules over antibodies against cytokine targets.
- What prevents better-funded labs from outscaling them, and the “massively parallel purification” Ian dreams of willing into existence.
Hands down the best biotech blogger on the internet.
As always, the secret sauce is mapping the problem to a combinatorial library and reading it out using next-generation sequencing (NGS). We dive deep into this on the pod!
