Key Takeaways

  • Sri Kosuri challenges the machine learning drug discovery model: if training a predictive network requires testing millions of physical molecules, you should simply develop the physical hits and skip the model.
  • Ian Quigley counters that physical assays only solve the exact target screened, whereas machine learning extracts generalized chemical reasoning applicable across agriculture, animal medicine, and hard-to-screen human cytokines.
  • Leash Bio compressed the timeline from raw library hits to validated mouse PKPD data down to six months, needing only a handful of medicinal chemistry cycles.
  • High-throughput wet lab screening acts as a filter for biological artifacts; observing clusters of hits that mimic lumacaftor, a known cystic fibrosis chaperone drug, flags misfolded protein assays before bad data pollutes the model.

The Disagreement

Brandon Anderson put the core paradox directly to Quigley: “what is the business model in machine learning for drug discovery? Because in order to build a good machine learning model, you need lots of data and good high quality data at scale. You clearly have that. But then if you need that to train the model, like what do you need the model for?”

This is the Kosuri paradox, named after Octant Bio founder Sri Kosuri. The argument is simple: wet labs are expensive. If you run DNA-encoded libraries against a cancer target and uncover three tight-binding chemical series, you already hold the lead molecules. Spending millions of dollars on compute clusters to predict binding affinity after you already observed it looks like pure tech-theater.

Quigley acknowledges the critique when applied to a narrow scope. “I think if your only rationale is to go off and get good chemical material for a particular target then of course Sri is right,” Quigley admitted. "You would have it."

The divergence comes down to scale and transferability. Quigley argues that screening thousands of diverse proteins does something physical chemistry alone cannot: it teaches an algorithm how small molecules behave across biology. “Like 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.”

Who's Right (and When They're Wrong)

Kosuri is right if you run a single-asset biotech company. If your goal is taking one kinase inhibitor into clinical trials for a specific oncology indication, building a foundation model is an expensive distraction. Take your physical hits from the assay plate, send them to a contract research organization for lead optimization, and run your animal studies. The model adds overhead without reducing clinical risk.

Quigley is right when you look past the initial hit to lead optimization and out-of-domain targets. Finding a molecule that binds a protein is relatively easy. Turning that hit into a drug that survives metabolic breakdown in the liver and reaches target tissue without toxic side effects is where biotechs spend five years burning cash. Quigley noted Leash moved “directly from chemical library hits to good PKPD in a mouse in 6 months with a very small number of medchem cycles.” That speed comes from predictive reasoning, not repeated wet lab trial and error.

A predictive model also lets companies tackle biological targets where physical screening stalls, such as unstable human cytokines or veterinary pathogens.

The real moat is not code; it is laboratory process hygiene. High-throughput data contains silent biological errors. Quigley pointed to molecular chaperones, compounds that stabilize misfolded targets: “There are drugs whose job it is to refold misfolded proteins called molecular chaperones. There's a very famous one that treats cystic fibrosis patients called lumacaftor. We have a bunch of chemicals in our chemical library that look like lumacaftor. And so now we can use it as one filter: if we get a bunch of lumacaftor hits, that protein probably didn't fold right.”

Without that experimental awareness, an engineering team will train its models on misfolded junk data, generating confident predictions that fail completely in living animals.

What to Do With This

Audit your data acquisition pipeline before buying your next batch of compute. Pick 50 unexpected positive results from your latest experimental run and trace their physical lineage back to the raw instruments. If you cannot explain whether an outlier stems from authentic signal or an assay artifact like protein misfolding, pause model training until you build automated sanity checks.