Summary
A bioRxiv preprint introduces PLABench, a benchmark comparing nine deep-learning methods for predicting protein–ligand binding affinity. Its authors report that strong overall accuracy did not translate into consistent performance across targets or protein families.
A bioRxiv preprint by Lyuwei Wang and Jianlin Cheng introduces PLABench, a framework for comparing deep-learning methods that predict how strongly a small molecule binds to a protein. In tests spanning blind CASP16 targets, leakage-controlled ChEMBL35 sets, and the Davis and KIBA datasets, the authors report that model performance varies substantially by target and protein family—even when structure-based methods perform well overall.
Binding-affinity prediction is used in computational drug discovery to help assess and prioritise candidate molecules. It estimates an interaction property; it is not, by itself, evidence that a molecule will be an effective or safe medicine. Comparing prediction methods is therefore useful, but the result depends on whether the evaluation resembles the challenge of predicting interactions for unfamiliar targets.
How PLABench compares methods
The authors benchmarked nine deep-learning methods, including sequence-based and structure-based approaches. They describe the framework as target-centric: instead of relying only on aggregate scores, it examines performance across individual targets and protein families.
The evaluation combines several kinds of test sets. These include blind CASP16 targets, ChEMBL35 sets designed to control data leakage, and the established Davis and KIBA datasets. Data leakage occurs when information related to test examples enters model training or evaluation, making performance on a benchmark less representative of genuinely new cases.
For structural comparisons, the authors used AlphaFold 3 to standardise structural input. They also compared predicted and experimental conformations, asking whether more experimentally accurate structures consistently improved affinity predictions.
Strong averages did not mean consistent generalisation
The preprint reports that pretrained structure-based models had the highest overall accuracy, but their results varied markedly between targets. Greater structural fidelity—from predicted to experimental conformations—did not produce consistent gains. On some targets, sequence-based methods outperformed particular structure-based models.
Evaluations grouped by protein family also showed uneven performance. The authors report that models had complementary strengths: one method could perform well where another did not, a pattern that can be concealed by a single overall score. Their findings point to a limitation of treating model scale or structural detail alone as a guarantee of reliable predictions for new targets. They argue that future approaches need more context-aware modelling of protein–ligand interactions.
The authors present PLABench as an extensible, open-source platform for further comparisons. The work is posted as a bioRxiv preprint. Its abstract gives qualitative performance findings rather than numerical scores or sample counts, so the size of the reported differences cannot be assessed from the abstract alone.