Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio

What it claims

These are show notes for a podcast with Ian Quigley of Leash Bio. The page gives topic bullets, not results. Beyond the summary, it adds a few specifics an engineer would want to know before trusting protein-ligand models or DEL data:

• Data leakage in Boltz. A BRD4 molecule used in comparison had already appeared in the PDB, so Boltz may have seen it during training.
• Hermes memorized proteins, not pockets. Leash caught Hermes learning "nicknames" for proteins instead of binding sites. Their fix was to engineer protein mutants at scale to "break the pocket."
• Clever Hans in ChEMBL. Models can infer which chemist made a molecule from its structure. Knowing "who made it" predicts activity about as well as the structure does.
• The Kaggle benchmark. It had roughly 4B measurements and about 2,000 teams. Nobody beat chance on an unseen chemical library.
• Minos sandbox. The existing DEL data turned out to supply tens of thousands of "lab-in-the-loop" simulations for agent RL. Agents reason about SAR in plain English.
• Drug program. On an inflammation program, Leash went from a DEL hit to good mouse PKPD in six months. Quigley argues for pills over antibodies on cytokine targets.

How it works

The page describes DEL hygiene only at bullet level:

• Each molecule carries dozens of barcodes. This lets Leash catch proteins that bind the DNA tag rather than the small molecule, which is the classic DEL artifact.
• Lumacaftor-like hits turned up and act as an accidental detector of misfolded proteins. Treat that chemotype as a signal about protein quality, not as genuine binding.
• The general recipe is to map the problem to a combinatorial library with a next-generation-sequencing readout. Quigley's rule is that biology scales only through imaging and DNA sequencing.

Limits

• These are episode notes. The page gives no benchmark names, metrics or numbers for the Hermes-vs-Boltz comparison, the mutant experiments or Minos.
• The Minos preprint, blog and dataset are mentioned but not linked or described here. Their size, licence, access and evaluation protocol must come from those sources.
• The Minos discussion was recorded two months before release, when it was an early prototype. A follow-up lightning episode is promised.
• The Clever Hans and leakage points mean you should audit any ChEMBL- or PDB-trained binding model's evaluation splits for author and structure overlap before relying on reported scores.