Deep learning has become a standard tool in early drug discovery for two jobs: predicting how a small molecule sits in a protein's binding pocket, known as pose generation, and predicting how tightly it binds. Dozens of scoring functions now claim state-of-the-art results. In a Perspective in Nature Machine Intelligence, Viet-Khoa Tran-Nguyen and Pedro Ballester argue that those claims have become hard to compare, and set out guidelines for how such models should be evaluated.
Their list is short and practical. Training data, code and procedures should be disclosed so others can reproduce, reuse and scrutinise the work. Studies should guard against data leakage, where test molecules or proteins are too similar to training ones, against sampling bias and against flawed designs. Comparisons should use common data splits, check how much results vary from one training run to the next, and include proper baselines, so that a new method is measured against something simple as well as against the latest rival.
The paper builds on a body of work that has exposed the problem. Ballester's group has previously shown that scaffold splits, a common way of separating training and test molecules, overestimate virtual screening performance, and a 2026 benchmark called NextTopDocker, co-authored by both researchers, found limitations in current end-to-end machine-learning docking and strength in hybrid approaches that rescore classical results. Work on randomness in neural network comparisons, which the authors also cite, shows that a single run can flatter or bury a method.
For readers choosing a docking or affinity model, the guidelines double as a checklist for reading papers and model cards. A claim without shared splits, repeated runs and a simple baseline deserves less weight, however large the headline gain.
The limits are those of any Perspective. It proposes norms rather than presenting new results, and we could read only its abstract and reference list, not the full text, so we cannot report how it ranks specific tools. Guidelines also depend on enforcement: unless journals, reviewers and benchmark maintainers require these checks, the incentive to publish the most flattering number will remain.