Executive Summary

Predicting whether a protein and a ligand (a candidate drug molecule) bind, and how tightly they stick, is at the heart of computational drug discovery. Benchmarks so far have focused on measuring those two things. One is binary binding prediction, the task of deciding whether a given protein and ligand bind at all. The other is affinity regression, which predicts the strength of binding as a number and checks how well it correlates with real experimental values. The headline numbers on the leaderboards come mostly from these two metrics.

The trouble is that these metrics only confirm the fact that binding happens; they never ask why and where the model decided it would bind. Binding occurs when a drug enters a specific pocket on the protein and forms chemical interactions such as hydrogen bonds or hydrophobic contacts. Deciding whether binding happens and pinpointing the location and type of those interactions are entirely different abilities. The former can be faked with statistical patterns alone; the latter requires understanding the mechanism of molecular recognition.

InteractBind, submitted to the 2026 NeurIPS Datasets and Benchmarks track, takes direct aim at exactly this gap. Its title is a question: do protein-ligand models learn binding sites, or just binding likelihood?

InteractBind is built from 99,391 protein-ligand pairs drawn from the Protein Data Bank (PDB). It holds 11,473 unique proteins and 9,017 unique ligands, and each binding event is annotated with six kinds of non-covalent interaction: hydrogen bonds, salt bridges, van der Waals contacts, hydrophobic contacts, π-π stacking, and cation-π interactions. The authors re-evaluated eight models on the same data, including three sequence-based models and five interaction-aware ones.

The contrast in the results is stark. FusionDTI, the top performer, scored AUROC 98.3% on binary binding prediction. On the question of whether binding happens, it is close to perfect. Yet the same model's ability to pinpoint the binding site — the share of cases where its single best candidate matches a real binding residue (BRHR@1) — was only 21.6%. Even widening the field to five candidates (BRHR@5) reached just 35.6%, less than half. Line up the top three models and they all sit below half: FusionDTI 35.6%, DrugBAN 31.4%, GraphBAN 30.4%.

This low localization accuracy was also uneven across types of binding. The paper reports that when performance was broken out by the six non-covalent interaction types — hydrogen bonds, salt bridges, hydrophobic contacts, and the rest — it varied markedly from one type to another. Some binding modes are at least caught; others are almost entirely missed. A model that had learned the general principles of molecular recognition should hit them evenly regardless of interaction type, but in practice performance was concentrated on the specific binding modes seen most often in the data.

The interpretation converges on one point. If a model that nearly perfectly calls whether binding happens misses where that binding takes place most of the time, what it has learned is likely not the chemical mechanism of molecular recognition but a statistical distribution — "this kind of protein tends to attract this kind of ligand." The paper's concluding sentence says as much: strong binary binding performance does not translate into reliable binding-site localization.

If InteractBind showed what the models miss, the Kopko team (CIIRC at the Czech Technical University, ETH Zurich, and Masaryk University) quantified when they break down. They evaluated four learnable scoring functions two ways. One is CASF-2016, the industry-standard benchmark; the other is a similarity-controlled split that isolates genuinely novel targets with low pocket similarity to the training data.

On the standard benchmark, the models were excellent. GEMS scored a Pearson correlation of 0.815, GEMSATOMICA 0.808, and GenScore 0.807. But move them to the similarity-controlled split and the same models fell to roughly half. GenScore dropped to a mean R of 0.422, and on its worst target it returned 0.047 — effectively no correlation at all. A model that boasted 0.8 on the standard benchmark was, in front of a pocket it had never seen, not much better than a coin flip.

The team's conclusion is aimed at methodology. Standard benchmarks split data heavily toward similar targets, so they fail to reflect the real difficulty of generalizing to novel targets. In other words, CASF-2016's 0.8 is closer to a measure of how much the evaluation set resembles the training set than to the performance a model would deliver in the field. As a remedy, they propose the wide adoption of leakage-free, similarity-controlled splits.

The two papers come from different tasks, different data, and different teams. InteractBind reports a failure of binding-site localization; the Kopko team reports a collapse of performance on novel targets. Yet both symptoms share the same root. Today's models learned the distribution of the training data rather than the mechanism of binding, so they post good scores only when the evaluation set resembles the training set. A benchmark score cannot separate a model's skill from "benchmark-training set similarity."

Could better representation learning then close the gap? The Kopko team combined self-supervised pretraining embeddings (ATOMICA) with GEMS. It helped. On unseen pockets, mean R rose from 0.470 to 0.498, about 6%, and the worst-case R on the hardest target improved from 0.236 to 0.311, about 32%. The direction is right. But even the improved figures fall far short of the 0.8 seen on the standard benchmark. Pretraining narrows the gap without closing it.

The competence of scientific AI is decided by two kinds of data: what it trained on, and what it was measured against. The higher the cost of failure in a domain, the more the second one matters — and few domains carry higher failure costs than drug discovery. The expensive collapse of a promising candidate in the clinic often begins with mistaking an inflated benchmark score for real skill. What the two papers target is not a defect of the models but the design of the evaluation set.

So the definition of "AI-Ready Data" widens by one notch. It no longer stops at whether the training data is clean and the labels are accurate; it now includes whether the evaluation set overlaps with the training set in time or structure. Was it split by similarity control? Does it verify not just whether binding happens but where the binding site is? Was it tested on targets the model has genuinely never seen? Designing a leakage-free evaluation set lifts trustworthiness far more than scaling the model up one more time.

The same problem repeats as you move across domains: the publication bias in drug-discovery AI that overrates models trained only on successful experiments, the hidden-pocket case where a binding site missed by structure prediction was recovered from measurement data, and the LLM benchmark contamination where evaluation questions leaked into the training data. The causes differ a little, but the conclusion converges. Scores are made not by the model alone but by the quality of the data that measures them.

• 1.Meng, Z., Bai, Z., Yuan, K., Ounis, I., Meng, Z., Xu, H., & Loscalzo, J. (2026). "A Large-Scale Dataset and Benchmark: Do Protein-Ligand Models Learn Binding Sites or Just Binding Likelihood?" NeurIPS 2026 Datasets & Benchmarks Track. arXiv:2605.24045. 99,391 pairs, 8 models re-evaluated; FusionDTI AUROC 98.3% vs. top-1 binding site 21.6%.
• 2.Kopko, J., Graber, D., Eyrilmez, S. M., Mazurenko, S., Bednar, D., Sedlar, J., & Sivic, J. (2025). "Generalization Beyond Benchmarks: Evaluating Learnable Protein-Ligand Scoring Functions on Unseen Targets." arXiv:2512.05386. Similarity-controlled split; CASF-2016 R 0.807 collapses to a worst-case R of 0.047 on unseen pockets.