Three hundred lab measurements sharpen virtual screening

Virtual screening sifts through vast compound libraries to find molecules worth lab testing. Budgets are limited, so the goal is not full classification but how early truly active compounds appear in the ranking. A paper by Japanese authors asks whether Boltz-2 can be prepared for this task using a few dozen to a few hundred measurements from a specific assay.

How much data is enough

The authors fine-tuned Boltz-2's affinity heads using between 40 and 300 binary activity labels. They evaluated the model retrospectively on eight targets from the MF-PCBA set. With 300 measurements, the number of active compounds in the top one percent of the ranking rose by a geometric mean factor of 1.77 compared to the untuned variant. Average precision improved 2.14-fold. The result depends on the target and training set composition, so the authors treat it as a signal rather than a rule.

The second experiment addresses compute cost. Instead of rescoring the entire candidate set with the tuned head, the researchers limited rescoring to the roughly ten percent of proposals that Boltz-2 had scored highest. Even with this narrowing, hit recovery remained comparable to full rescoring. The practical takeaway is simple: if the model already has a good sense of the coarse scale, refining the tail of the list is sufficient.

The authors note that the study is retrospective. The labels come from existing datasets rather than new experiments, so the numbers describe method performance on known data, not its effectiveness in an unknown lab. The practical conclusion remains useful. Instead of training a model from scratch on tens of thousands of measurements, a team can fine-tune the heads on a set that fits a single project's budget and still improve candidate ordering.

Data processing in proteomics

The measurement side is developing in parallel. Another paper presented ProteoEM, an open Python library that estimates protein and proteoform abundance from single-molecule affinity traces. The problem is that imperfect, nonspecific probe binding allows one trace to fit several possible identifications. Rather than assigning a trace to a single candidate, the author proposes splitting the uncertainty by weight in an expectation-maximization scheme, using pre-calibrated probe response rates. The model reports indistinguishable proteoforms as groups instead of forcing an arbitrary choice.

Both approaches share one idea: in biomedical research, value lies not in abstract accuracy but in how a model's decision translates into the next, costly lab step. Consequently, model evaluation is shifting from average quality across the whole set to performance in the narrow window that actually matters.