The scoreboards are mis-measuring: three audits of molecular-model benchmarking
VHH-Bench grades six structure predictors on nanobody complexes; a PPI study shows pooled metrics conflate target effects with mutation ranking; a DTI audit finds standard filters discard 78.4% of negative data.
- In a preprint, VHH-Bench evaluated six leading structure-prediction models on nanobody-antigen complexes; Protenix-v2 achieved the highest overall docking accuracy, while all six performed worse on kinked CDR3 conformations. Self-reported m:…0606
- In a preprint on PPI ΔΔG benchmarking, moving from random splits to family-out-of-distribution evaluation cut pooled Pearson correlation from 0.6844 to 0.1508, with the between-target component accounting for 81.1% of that change. Self-reported m:…4910
- In a preprint, standard low-bias filtering of the BindingDB drug-target interaction benchmark excludes 78.4% of measured negative pairs, and models scored 0.072-0.136 lower pooled AUROC on those excluded regions. Self-reported m:…9914
A scoreboard built for nanobodies
In a preprint posted this week, researchers Nehad el Salamouni, Qiang Zhu and Haibo Yu introduce VHH-Bench, a benchmarking framework purpose-built for nanobody-antigen complex structure prediction. Rather than borrowing a generic protein benchmark, they curate a non-redundant dataset stratified by CDR3 conformation - the hypervariable loop that largely determines what a nanobody binds - and run six leading models through it: AlphaFold3, Boltz-2, Chai-1, ESMFold2, HelixFold3, and Protenix-v2. The workflow is standardized and reproducible, with model confidence analysis folded into the evaluation so developers can see not just who scored best but where confidence and correctness diverge. m:…0606
Protenix-v2 came out on top for overall docking accuracy and success rate. The more durable finding, though, is the failure mode all six share: reduced accuracy and higher failure rates on complexes containing kinked nanobody CDR3 conformations, compared with extended ones. The implication the authors press is that a benchmark which does not stratify by loop geometry lets models look competent on average while failing exactly where therapeutic design gets hard. That is my read of the result, not the paper's stated claim: nanobody CDR3s are the business end of the molecule, so a scoreboard blind to their shape is a scoreboard blind to the hard cases. m:…0606
Pooled metrics mix two different signals
A second preprint, from le Yu, takes aim at the ΔΔG prediction benchmarks used to rank mutations in protein-protein interactions - the scores drug teams use to decide which variants to make. The core claim is methodological: pooled correlation metrics conflate between-target signal with within-target signal. A model can post a strong pooled correlation largely because it correctly orders targets from easy to hard, while ranking mutations within a single target poorly - which is the actual job. The author demonstrates this with an exact covariance decomposition that separates the two components. m:…4910
The measured gap is large. Under random splitting, pooled Pearson correlation reaches 0.6844; under family-out-of-distribution evaluation it falls to 0.1508, and the between-target component accounts for 81.1% of that absolute change. Separately, the preprint finds that 40.8% of PPI targets with at least five mutations sit in a degenerate regime for fixed-K shortlist metrics, meaning the metric stops discriminating for nearly half the test cases. The practical read: if a pipeline selects models by leaderboard, it is mostly selecting for target-difficulty memory, not mutation ranking skill. m:…4910
The negatives the filters threw away
The third preprint, from Justin Li, reconstructs the eligibility rules behind standard low-bias filtering of the BindingDB drug-target interaction benchmark and measures what the filters discard. The answer: 78.4% of measured negative pairs - the experimentally confirmed does-not-bind data that drug pipelines depend on. When leading models (DrugBAN, MolTrans, BioT5) are evaluated on those excluded regions, pooled AUROC drops by 0.072 to 0.136 relative to the standard benchmark regions. The benchmark's headline numbers, in other words, are measured on the data the filters kept, which is also the data the models handle best. m:…9914
The paper's closing recommendation is procedural rather than technical: "Benchmark reports should specify eligibility, coverage at the evaluated unit, metric aggregation, and performance on excluded regions where labels are available." That is a reporting standard, not a new model - and it is the cheapest fix of the three papers' proposals. m:…9914
The pattern across all three
None of these three preprints claims AI models are getting worse at molecular tasks. The shared argument is narrower and harder to dismiss: the rulers are bent. Do not stratify by conformation, do not decompose pooled metrics, do not report performance on the excluded negatives, and every scoreboard reads high. For teams choosing models for a discovery pipeline, the practical question is not what the top score is but what the benchmark chose not to measure. All three are preprints, so treat the numbers as reported findings awaiting independent replication, not established facts. m:…0606 m:…4910 m:…9914
What is still unresolved
- Whether VHH-Bench rankings hold once peer-reviewed and extended to more recent model versions.
- Whether the covariance decomposition approach transfers to other ΔΔG and affinity benchmarks beyond PPI.
- Whether benchmark maintainers adopt the proposed reporting standard for excluded regions.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.