Brandon Anderson, who interviewed Ian Quigley on the Latent Space podcast, opened with a confession. Every person he had met who worked with a DNA-encoded library, he said, had "come out severely traumatized." When a mutual friend told him he had invested in a company built on the technique, Anderson joked that it was "money you'll never see again." In the episode he told the audience he had been wrong.
The company was Leash Bio, and his guest, Quigley, is one of its co-founders. Leash runs these screens at large scale to produce data for machine-learning models that predict which small molecules stick to which proteins. Over the conversation, Quigley explained why the method has such a bad reputation and what his team changed to make its results trustworthy enough to train on.
A fishing trip in one tube
Most drugs are small molecules. Quigley described the goal of small-molecule drug discovery as getting "a dinky chemical" to bind to a protein thought to be causing a disease. The slow way to find one is to keep chemicals in separate tubes and test the protein against them one tube at a time. A DNA-encoded library, or DEL, puts a huge mixture of chemicals in a single tube, each attached to a short DNA barcode that records what it is.
The target protein is attached to a magnetic bead and dipped into the mixture "kind of like a fishhook," Quigley said. Molecules with the right shape and electrical properties settle into a pocket on the protein. The bead is pulled out and rinsed, and the attached barcodes are copied with PCR, the standard DNA-copying reaction, then read on a DNA sequencer. The barcodes that show up most often identify the molecules that bound.
Pooling makes very large libraries possible. Quigley said the biggest collections tested one tube at a time are in the low single-digit millions of molecules, while the largest DEL ever produced held 40 trillion. He added that "there's some problems when they get that big." The libraries are built from building blocks in combination, so thousands or tens of thousands of blocks can produce many millions of molecules. Google's account of its DEL work explains why this is efficient: two stages with N choices each give N² combinations from roughly 2N chemical steps.
The building-block design restricts what chemicals a library can contain, Quigley said, but it suits machine learning. Many molecules differ from their neighbors by only one or two pieces. That lets a model work out which part of a molecule drives binding and which parts can change without much effect.
The readout matters as much as the chemistry. Quigley gave what he called a "party line": for combining life science with machine learning, there are really only two scalable ways to measure things, taking pictures and sequencing DNA, and "sequencing DNA is another way of taking pictures." He conceded that some proteomics researchers, who measure proteins directly, would disagree, but argued that their methods are not yet at the same scale.
Controls, and proteins that grab the barcode
Counting barcodes means nothing without a comparison, Quigley said. A team has to build negative controls, such as dipping the bead alone into the library, and has to check for molecules that stick to many proteins. Leash's public BELKA dataset separates two kinds of background measurement: sequencing the unscreened library to see what it contained to begin with, and bead-only runs to catch molecules that cling to the capture material. Its documentation records one case where the company switched bead types because some building blocks bound strongly to the original beads.
The bigger hazard comes from the protein side. Many proteins exist to grab DNA and switch genes on or off. Put one of them in a tube full of DNA barcodes and it may latch onto the tag instead of the drug-like molecule, creating false positives. Quigley named this as one of the problems behind people's DEL nightmares.
Leash's answer is redundancy. Each small molecule in its libraries is encoded by 32 or 48 different barcodes. If a molecule really binds, all of its barcodes should rise together. If one barcode stands far above its siblings, even though they all carry the same chemical, the protein is probably binding that DNA sequence. Each copy also carries a stretch of unique sequence, Quigley said, so the many copies PCR makes of one captured molecule can be traced back to a single binding event. Some DNA sequences copy faster than others, and this guards against that bias.
The DNA-binding problem can also work as a check. When a transcription factor, a gene-regulating protein known to prefer a short DNA pattern of six or eight letters, is screened, Leash can read which barcodes it pulled out and see whether that pattern reappears. If it does, Quigley said, "I think our protein probably got made right." In a blog post on its screening methods, Leash also describes testing DNA-binding proteins with added DNA fragments that carry their preferred sequence and compete with the barcodes.
"We don't pretend that it's not" there
The conversation then turned to a familiar objection. Every molecule in a DEL has a DNA tag hanging off it, so does the tag change how the molecule binds? Quigley said the answer depends on why the experiment is being run.
If the goal is a drug, the tag does matter. Leash has to make the promising chemicals again without DNA and test them in functional assays. Quigley said about 90% of them work for targets that are easy in Leash's hands, and about 50% for hard ones, adding that "other people have different experiences."
For machine learning, he argued, the tag is not a flaw. The tagged molecules were physically in the tube, and some bound while others did not. "That is real signal that you were physically measuring from the world," he said. Leash's models include a placeholder that marks where the DNA is attached. When the conversation raised the possibility that a tag might block a pocket and hide a real binder, Quigley said that is not a hidden false negative for the model: "We know it's there and we don't pretend that it's not."
The evidence he cited for trusting the data is reproducibility. When Leash reruns experiments with fresh batches of protein and a newly made batch of the chemical library, months apart, "we get back the same hits in the same order for the most part."
Why the old way produced trauma
Quigley traced DELs' reputation to the conditions they were invented under. Sequencing was expensive, so each screen was a big logistical and financial bet. Teams usually turned to DELs for one very hard target after other methods had failed. They might spend two to six months growing and purifying the protein, then try to improve their odds by using the biggest, most diverse library they could get.
Leash reversed that pattern. It is less interested in a single binder for one high-value target than in active chemicals across many proteins, so each of its chemical libraries is screened against many hundreds of targets. That shows how each library behaves. Quigley described one library that Leash had manufactured more than once: it was the same chemical library each time, but it behaved differently. "We did not know that until we screened hundreds of proteins against it," he said.
Quigley said he and several colleagues came from genomics methods development rather than chemistry. Before Leash, he and his co-founder Andrew Blevins spent years at Recursion, a company that produces large amounts of cell-imaging data. There, Quigley said, the team made "a lot of mistakes" finding the ways models could cheat on experimental data. One test carried over: "Can you predict, you know, the instrument that this was run on? And if you can, that's really bad." If a model can tell which machine produced a measurement, it may be learning lab quirks instead of biology.
The lumacaftor accident
Asked how Leash knows its proteins are any good, Quigley gave his favorite current check. A protein that folds badly can sometimes be caught on a gel, a lab test that sorts proteins by size, but a slightly misfolded protein is hard to spot.
Some drugs help misfolded proteins fold properly. Quigley named lumacaftor, a well-known cystic fibrosis treatment. Leash's libraries contain chemicals that resemble it. "Guess what kind of proteins it hits when we screen them," he asked. Anderson guessed cystic fibrosis. "Misfolded proteins," Quigley replied.
He said the team found this by accident. The same lumacaftor-like chemical series kept appearing in screens of many unrelated proteins. When the team checked the protein production records, the hits tended to show up when production had gone badly. Leash now uses the pattern as a filter: if a screen returns many lumacaftor-like hits, the protein probably did not fold correctly. The team keeps that screen out of the training data and reruns the protein under conditions that are kinder to it. The link is Quigley's empirical observation from Leash's own screens. Leash's screening post describes a related cross-check: closely related proteins should often prefer similar chemicals, so a hit profile that looks unexpectedly different raises suspicion.
What a richer lab can't simply buy
Anderson put a pointed question to Quigley: "What stops a better funded lab from outscaling your Dell engine?"
Quigley's answer was patience and culture rather than equipment. Laboratory work naturally produces "cheating modes," he said, and many of them come from batch effects, the differences between one run and another. Catching them means tracking metadata about every experiment and building tools that are easy enough that the people doing the work will actually record it. It also means going back through data from the company's founding to check for consistency, and tracing every large change to its cause. Leash rebuilt its selection process and much of its protein production from scratch, he said, and added other measures to judge whether a protein preparation went well.
Quigley was sympathetic to the labs that never did this. A team that only wants one or two good hits to start a drug program has no reason to chase down every aspect of the assay. "But like if you're motivated for it to be machine learning fodder then like yeah you have to pay attention all this stuff," he said. Anderson's verdict: "You're also traumatized in a good way."
What the data cannot do yet
Anderson asked whether the screens could measure potency, meaning how strongly a molecule binds. Chemists often express that as a dissociation constant, or KD. Quigley said relative ranking is "quite doable," since repeated screens return the same hits in roughly the same order. Turning sequencing read counts into actual KD values is "not really" possible yet. Some proteins that are not described as DNA binders have turned out to bind the barcodes, which could distort such numbers. Leash has started using biophysical measurements, including surface plasmon resonance (SPR), a technique that measures binding directly, to build a "potency ladder" of reference molecules. Quigley said the team hopes for a "sloppy translation" into KD values but "wouldn't say that we're there yet."
Asked what bottleneck he would remove by fiat, Quigley first pointed to one already shrinking: sequencing costs, which he said had fallen by "many many many orders of magnitude." Anderson suggested that a human genome had gone from a billion dollars for the first one to about a thousand. Quigley thought it might now be about a hundred. Both figures were offered as rough recollections. Quigley also said there were stretches when prices fell faster than Moore's law, and the US National Human Genome Research Institute's cost data shows that pattern from 2008, when sequencing centers moved to next-generation machines.
His real wish was for massively parallel purification. Chemists can clean up a newly made molecule with tools such as HPLC, a column-based separation method, but Quigley said none of the very good options are very parallelizable. As Waters' guide to preparative HPLC explains, throughput depends on factors such as column capacity, solubility and how finely compounds must be separated, so scaling up takes more than running more samples. Cheap, large-scale purification would let Leash take many molecules from the "funky assay" that has traumatized so many people and test them in a second assay that is "really clean and purified and easy to work with." That, Quigley said, would let the team "orthogonally validate," or independently check, the measurements it makes.