Key Takeaways

  • Public drug discovery databases like ChEMBL train machine learning models on human bias rather than physics because they reflect which molecules historical chemists preferred to synthesize.
  • DNA-encoded libraries can scale up to 40 trillion molecules in a single tube, but historical screens failed because teams treated them as noisy, one-off gambles on hard targets.
  • Leash Bio attaches 32 to 48 unique DNA barcodes to the exact same small molecule to separate genuine protein binding from false-positive DNA interactions.
  • Rerunning experiments across fresh protein batches and chemical batches over months yields the exact same hits in the exact same rank order when physical DNA tags are treated as ground-truth training signals.
  • High-throughput chemical binding assays can be converted into fast reinforcement learning sandboxes where LLM reasoning agents run automated structure-activity relationship campaigns.

How Public Data Teaches Chemist Sociology Instead of Physics

Most computational chemistry efforts stall because of the data feeding them. Academic repositories like ChEMBL contain millions of experimental measurements, but those measurements do not represent a neutral sample of physical reality. They represent human research trends. Chemists choose what to make based on patent incentives, cheap starting reagents, and personal intuition. When you train a deep learning model on public chemistry data, the model does not uncover the underlying laws of molecular binding. It memorizes which functional groups academic chemists published between 1990 and 2020.

Ian Quigley and Leash Bio bypass this trap by generating their own physical data from scratch. Instead of trusting historical literature, Leash runs physical assays that test massive combinatorial collections against thousands of engineered protein constructs. The goal is not finding one magic bullet for a single diseased tissue. The goal is mapping raw chemical affinity across biology without human editorial selection.

As Quigley put it: “What we have done that's very very different is we screen tons and tons and tons of proteins and we are less driven by trying to find individual binders for a particular protein target that might be of high interest for a program. We just want to get tons and tons of active chemical material across large numbers of proteins.”

Turning Chemistry into DNA Sequencing Images

DNA-encoded libraries (DELs) solved the collection problem decades ago by tagging synthetic chemicals with short DNA identifiers. In theory, you can test vast chemical spaces in one tube. In practice, early pharmaceutical adopters hated the technology because the experiments were dirty, expensive, and filled with non-reproducible hits.

“With DELs, the largest one that has been produced is 40 trillion molecules,” Quigley noted. “There's some problems when they get that big, but that is the kind of scale that you could try to attempt with this technology.”

The breakdown happens when the target protein binds to the DNA barcode rather than the small molecule attached to it. When that occurs, you sequence a hit that has zero medicinal value. Leash fixes this by building redundant verification into the library design.

“One thing that we do is we have large numbers of barcodes that each encode one particular type of small molecule,” Quigley explained. “You say here's one small molecule. That's great. Now we have 32 or 48 different barcodes all encoding this one small molecule.”

If all 48 distinct DNA sequences show up in the sequencer, the small molecule binds the protein. If only two barcodes light up, the protein bound the nucleic acid backbone, and the algorithm drops the false positive.

By treating DNA sequencers as high-volume optical sensors, Leash turns biochemistry into an information pipeline. Quigley summarized the mental model: “Sequencing DNA is another way of taking pictures. It's basically pictures and DNA sequencing. And if you can shovel your challenging problem into that using one of those approaches as a readout,” you unlock massive scale. When Leash runs fresh protein batches against new chemical batches over months, Quigley found the data reproduces the exact same hits in the exact same rank order.

What to Do With This

Audit your primary model training set this week. If you rely on scraped public benchmarks or human-labeled historical records, isolate 500 edge cases and tag whether the label reflects real physical performance or human workflow habit. If your data tracks human behavior rather than objective mechanics, design a hardware or software loop that generates machine-readable ground truth before you train your next model weights.