Evidence at a glance
More Data Makes the Question of Learning Harder
Leash Bio has released a preprint, a blog post, and an open dataset for Minos. Minos organizes the team's accumulated DNA-encoded library, or DEL, screening records into an environment where AI agents can explore structure–activity relationships and test scientific reasoning. The aim is to let agents adjust their search repeatedly in response to experimental feedback. This is not a general question-answering task. It addresses a harder drug-discovery question: can a model's judgment about a molecule be checked against experimental records or subsequent experiments?
The work starts from a gap between data volume and generalization. A Kaggle competition involving Leash included about four billion measurements and roughly 2,000 teams. The task was to predict outcomes for an unseen chemical library, and according to the materials, no team beat random guessing. That does not prove the data were useless. It does warn research leaders that a large training set does not mean a model has captured relationships that hold across chemical libraries.
DEL Readouts Must First Survive Experimental Checks
DEL achieves high throughput through a parallel screening process. Small molecules attached to DNA barcodes are pooled and incubated with a target protein. After washing, sequencing reads out which barcodes became enriched, offering clues about which molecules may interact with the protein. Sequencing allows many candidates to be screened together, but a barcode signal does not automatically mean that the small molecule bound. A protein may be misfolded, it may bind directly to the DNA, and noise can also make enrichment look like a hit.
Leash describes using several controls to check the screening results themselves. Adding a known competitor should reduce the expected binding signal. A mutant that disrupts the binding pocket should change the corresponding signal. In-house proteins are compared with commercially available proteins whose function has been validated, and related proteins should show plausible relationships in their results. The team says it still discards some failed screens after these checks. The point is not to label data “high quality,” but to make the credibility of an experiment something that can be questioned and evidenced.
Barcodes and Mutations Can Expose Shortcuts
Barcode design makes one hidden experimental problem more observable. The program description says Leash assigns dozens of barcodes to each molecule. If different barcodes for the same compound produce inconsistent readings, that may point to nonspecific DNA-related signals or other screening problems. The public description does not provide the full barcode design or decision rules, however, so “multiple barcodes” should not be taken to mean that every source of interference has been eliminated. It is better understood as evidence that can help locate a problem, not an automatic quality guarantee.
Models can also exploit shortcuts in experimental data. Leash's analysis of Hermes found that, in some protein predictions, it paid more attention to non-conserved regions. The team suspected that the model might be identifying protein identity rather than judging interactions through the binding pocket. After adding single-amino-acid mutation data to training, Leash reported that the model paid more attention to conserved functional regions. This is a useful diagnostic result, but a change in attention alone does not prove that the model has learned the physics of binding.
Minos Turns Historical Measurements into a Testbed
Minos connects the other end of this quality-control chain to agents. The materials describe agents reasoning about structure–activity relationships in plain language and searching among about 6.5 million molecules for binders, with a stated search budget of 100. The value is not simply that a model can “find an answer.” Researchers can compare candidate rankings with existing screening results and observe how the agent changes its search in response to feedback. Leash presents this kind of environment as a foundation for scientific reinforcement learning: experimental records provide checkable feedback, letting agents try repeatedly without paying the cost of running every experiment.
But checkable does not mean equivalent to a real experiment. Minos is built from existing DEL data, so historical screening noise, biases, and gaps in chemical-library coverage may enter the environment. If evaluation data are not sufficiently separated from training data, agents may also score well by exploiting familiar records. The materials mention an open dataset and a preprint, but do not provide enough information here to assess leakage controls, licensing details, or transfer to real experiments. A high score in the simulation therefore first shows that a model can perform in that environment. It does not establish that it has found a molecule that will work reliably in the lab.
Use Computational Scores to Set Priorities, Not Conclusions
The comparison between Hermes and Boltz should be read within the same limits. The materials say Hermes was trained only on Leash's own data and kept pace with Boltz on public benchmarks. They also note that Boltz had previously seen a BRD4 molecule in the PDB. This example shows how benchmark results can depend on training-data coverage and the particular test case. A score alone cannot establish which model has a stronger understanding of chemistry. The more useful question is whether a model remains reliable on an unseen chemical library, under protein mutations, and in other conditions that disrupt superficial correlations.
For a drug-discovery team, Minos is better treated as a low-cost way to prioritize candidates and diagnose models than as a replacement for wet-lab work. Leash also mentions an inflammation program that went from a DEL hit to good mouse PK/PD results in six months. That shows experiments still determine whether a candidate advances, but the materials do not attribute that progress to Minos or to agents. A practical approach is to establish how training and evaluation are separated, test whether agents rely on shortcuts using mutations or unseen data, and send a small number of high-priority candidates to the lab. Minos becomes part of a genuine research loop only if its simulated advantage is reproduced in independent experiments.