{"id":"eadae3b9-f542-4d1f-b006-04c59314af2b","arxiv_id":"2507.10877","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"BioScore unifies affinity prediction, pose ranking, docking, and virtual screening for diverse biomolecular complexes in one graph neural network model.","lead":"BioScore is a newly proposed machine learning scoring function that predicts binding strength for many kinds of biomolecular complexes, including proteins, nucleic acids, small molecules, and carbohydrates, using one unified representation. The authors report it matches or beats 70 existing tools across 16 benchmarks and introduce a new protein-protein interaction benchmark, but the full method and data are not yet released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PPI screening success is driven almost entirely by the interaction-edge-count term, and the self-built screening set does not size-match pseudo-ligands; without a size-controlled check, the unified-screening claim remains unverified.","rationale":"I read the central claim as: one model can replace system-specific scoring functions for scoring, ranking, docking, and screening across proteins, nucleic acids, small molecules, and carbohydrates. The strongest independent evidence is the breadth of benchmarks and the explicit leakage controls (PDB ID and sequence-based deduplication, macrocycle exclusion, cluster splitting). I do not see a mathematical inconsistency in the dual-tower construction, and the ablations are a point in the authors' favor because they isolate the edge-count term's contribution. The weak point is whether that term's contribution to PPI screening is physical or artifactual. The reader identified the self-built PPI Benchmark; I partly agree but would sharpen the concern: the load-bearing component inside that benchmark is log(n_ij), and the benchmark does not control for interface size. A size-matched reanalysis would settle whether BioScore is a general screening function or a buried-surface-area proxy. This does not overturn the overall verdict; the PLI/CASF results and the nucleic-acid and cyclic-peptide transfer experiments are independent of the PPI screening protocol. It does mean the strongest 'first to unify all four capabilities' claim should stay conditional until either the data and code are released or the size-matched screening check is run.","tokens_in":34369,"tokens_out":10431,"duration_ms":127505,"concrete_test":"Recompute the PPI screening metrics on the existing benchmark after stratifying the pseudo-ligands per receptor by chain length, heavy-atom count, and predicted buried surface area so that their distributions match the true binder's distribution, and compare BioScore with its own Ablation 3 (no edge-count term). If SR1% drops from 15.19% toward the 1.27% ablation level once interface size is matched, the screening claim is an artifact of the log(n) confidence term; if SR1% remains near 15%, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is not only that the new PPI Benchmark is fair, but that the screening advantage it reports for BioScore reflects binding discrimination rather than an interface-size prior. In Table S3, removing only the interaction-edge-count confidence term (Ablation 3 vs BioScore) drops PPI screening SR1% from 15.19% to 1.27%, SR5% from 32.91% to 12.66%, and SR10% from 41.77% to 18.99%. That term is log(n_ij | d_ij < cutoff) in Eq. 25. The PPI screening set (SI §3) is built by cross-docking the 79 PDBbind partners against each receptor and taking 100 K-means-clustered ZDOCK poses per ligand, with pseudo-ligands excluded only by 90% sequence clustering via CD-HIT. Nothing in the protocol size-matches pseudo-ligands to the true binder or controls for interface area. If the true binder has systematically more contacts than the cross-docked pseudo-ligands, BioScore's screening power is largely an interface-size signal. This matters because the PPI Benchmark is the only evaluation in which BioScore alone shows effective PPI screening; the central 'unify scoring, ranking, docking, screening' claim depends on it. The absence of error bars and the lack of released pose sets make the reported SR1% value unverifiable from the manuscript alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces BioScore, a dual-scale geometric graph neural network that aims to serve as a foundational scoring function for diverse biomolecular complexes. The model combines an equivariant Transformer encoder with a mixture density network pretrained on structural distances, followed by task-specific fine-tuning. It claims to unify scoring, ranking, docking, and screening across proteins, nucleic acids, small molecules, and carbohydrates, and it introduces a new protein-protein interaction (PPI) benchmark built from PDBbind with scoring, docking, and screening subsets. The authors evaluate BioScore on 16 benchmarks, reporting state-of-the-art or competitive performance against 70 methods, and they demonstrate gains from cross-domain pretraining, zero/few-shot transfer, and specialized chemical spaces such as cyclic peptides and carbohydrates. The manuscript also includes ablation studies examining the contributions of interface masking, the statistical potential formulation, and the interaction-edge-count confidence term.","tokens_in":34708,"tokens_out":6992,"duration_ms":83669,"significance":"If the reported results are robust, BioScore would be an important contribution: it is the first model to combine scoring, ranking, docking, and screening in one framework for general biomolecular complexes, and the cross-domain pretraining results are compelling. The paper also provides a carefully deduplicated PPI benchmark, with sequence-similarity-based splits, and it includes extensive ablations. The authors are transparent about using MDN-based statistical potentials as motivation rather than as a rigorous PMF derivation, and the deduplication protocol is a clear strength. The significance is tempered, however, by the benchmark-construction choices and comparison protocol, which are the focus of the major comments below.","major_comments":[{"comment":"The PPI screening result is the only evidence that BioScore can discriminate true binding protein partners from non-binders, but the benchmark does not size-match pseudo-ligands to the true partner or control for interface area. Ablation 3 in Table S3 shows that removing the log(n_ij | d_ij < cutoff) term collapses PPI screening SR1% from 15.19% to 1.27%, while BioScore with that term achieves 15.19%. Because the true partner in its native complex will generally have more inter-molecular atom pairs within 8 Å than arbitrary cross-docked partners, the reported screening power may reflect an interface-size prior rather than binding specificity. The authors should report the distribution of n_ij for true binders versus pseudo-ligands, include a size-matched negative set, and repeat the ablation under size-matched conditions. They should also state how zero-contact poses are handled, since Eq. (25) is undefined for n_ij = 0.","section":"PPI Benchmark construction and Eq. (25), Table S3"},{"comment":"The baseline comparison protocol is inconsistent: GET and MINT are retrained on BioScore's training data, while GNN-DOVE and DProQA are evaluated with original published weights; many other comparisons (e.g., CASF-2016 in Table S8) rely on published results. This heterogeneous protocol makes the claim of 'consistently outperforms or matches 70 methods' difficult to assess, because some baselines are at a disadvantage from not being retrained on the same data. The authors should either retrain all baselines under the same data and task conditions where feasible, or provide a rigorous justification for using original weights and published results, and they should clearly distinguish 'retrained' from 'as-published' columns in every comparison table.","section":"Table 1 and its footnote"},{"comment":"The newly proposed PPI Benchmark is a central contribution, but the decoy pose sets (7,979 docking decoys and 613,900 screening conformations) are not released, and no error bars or statistical significance tests are reported for any of the metrics in Table 1 or the external benchmarks. Without the pose sets, other groups cannot reproduce or extend the benchmark; without confidence intervals, the small margins in several metrics cannot be interpreted as meaningful differences. The authors should deposit the exact benchmark poses and provide standard deviations or confidence intervals, at least for the PPI benchmark and CASF-2016 screening metrics.","section":"PPI Benchmark reproducibility and statistical reporting"}],"minor_comments":[{"comment":"There are numerous typos and language issues that should be corrected: 'techqniue' in the Introduction, 'resulintg' in the Results section on protein-ligand tasks, 'assements' in the Abstract and Discussion, 'Perfermence' in the heading, and 'exicting' in the Discussion.","section":"Throughout"},{"comment":"The term log(n_ij | d_ij < cutoff) is called a 'confidence' term, but it is simply the logarithm of the edge count and is not a confidence measure in any statistical sense. A more descriptive name, such as 'interaction-edge-count prior' or 'size term,' would be clearer.","section":"Eqs. (25) and (27)"},{"comment":"The paper repeatedly calls the PPI benchmark 'comprehensive' and 'full-spectrum,' but the SI correctly notes that a ranking-power benchmark could not be constructed for PPIs. The main text should temper these claims accordingly.","section":"Abstract and Discussion"},{"comment":"The validation curves for interface-masking versus no-interface-masking are mentioned as evidence of structural overfitting, but the figure is not described in the main text. Please add a sentence explaining what is plotted and what the reader should infer.","section":"Figure S1"}],"recommendation":"major_revision","confidential_remarks":"This is a strong empirical paper with a useful new benchmark and a well-executed cross-domain pretraining study. The main technical risk is the PPI screening evaluation, which may be dominated by an interface-size prior; this needs to be addressed before the paper can be accepted. The baseline comparison protocol also needs to be made fully consistent or clearly justified. If the authors can provide the additional analyses and release the benchmark data, the paper would be a valuable contribution to the field."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on BioScore. The paper is worth knowing about, but the headline screening claim is on shaky ground.\n\nWhat's new: BioScore is a genuine attempt at a generalist scoring function—one model for affinity, ranking, docking, and screening across proteins, nucleic acids, small molecules, macrocycles, and carbohydrates. The dual-tower design with a mixture-density-network statistical potential plus an edge-count confidence term is a reasonable engineering choice, and the interface-masking ablation shows real gains. The cross-domain pretraining results (e.g., PPI affinity, antibody-antigen, zero-shot NLI) are suggestive, and the authors made a real effort to deduplicate training data by PDB ID and sequence similarity. The new PPI benchmark is useful, at least as a template.\n\nSoft spots: The biggest problem is the PPI screening benchmark. The authors show that removing the edge-count term (Ablation 3 in Table S3) collapses screening success rates from 15.19% to 1.27% at SR1%, with similar drops at SR5% and SR10%. BioScore is the only method that discriminates on this benchmark, and the benchmark's pseudo-ligands are cross-docked partners with no size matching to the true binder. So the screening advantage could be mostly an interface-size prior—bigger interfaces have more contacts, and log(n_ij) rewards that. The paper doesn't provide a control showing that size-matched decoys behave differently. That undercuts the 'unified screening' claim, which is central to the 'first to unify all four tasks' narrative.\n\nOther issues: baseline comparisons are inconsistent (GET and MINT retrained on BioScore's data; others use published numbers). No error bars anywhere. No code or data released, so the reported numbers, especially the screening set, can't be independently checked. These are fixable, but not minor.\n\nWho's this for? People building or benchmarking scoring functions. If you work on protein-protein or nucleic-acid scoring, the PPI benchmark and the dual-tower idea are worth discussing. But I wouldn't use BioScore's screening results as evidence for a generalist claim until the size confound is addressed and the artifacts are released.\n\nRecommendation: Send it to serious peer review—the idea is timely and the architecture is plausible—but insist on major revision: size-matched screening controls, error bars, consistent baselines, and code/data release. As it stands, the central screening claim is unverified.\n\nThat's my two cents.","headline":"BioScore is a plausible generalist scoring function, but its PPI screening success appears to be driven by an interface-edge-count prior, so the 'unified screening' claim is not yet verified.","tokens_in":35248,"tokens_out":3086,"would_cite":false,"duration_ms":37265,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"BioScore claims to be the first scoring function that unifies affinity scoring, ranking, docking, and virtual screening across proteins, nucleic acids, small molecules, and carbohydrates in one framework, consistently matching or beating…","keywords":["foundational scoring function","biomolecular complexes","protein-protein interaction benchmark","binding affinity prediction","structure-based virtual screening","mixture density network","equivariant transformer","cross-domain pretraining"],"falsifier":"Run BioScore on an independent PPI screening set built with decoys from multiple docking engines, or with experimentally characterized non-binders, and check whether true binders are consistently ranked above decoys; if its screening discrimination collapses or becomes comparable to baselines, the PPI screening claim fails. Alternatively, recompute PPI docking success rates after removing the interaction-edge-count confidence term; if the docking gain disappears, the dual-tower mechanism is the load-bearing component.","tokens_in":34203,"feed_emoji":"🧬","tokens_out":6075,"duration_ms":63764,"temperature":0.7,"pith_summary":"This paper tries to establish that a single scoring function can handle structural assessment for all major classes of biomolecular complexes: protein–ligand, protein–protein, protein–nucleic acid, nucleic acid–ligand, and less standard systems like cyclic peptides and carbohydrates. If correct, that would let drug-discovery and structural-biology pipelines replace many system-specific scoring tools with one pretrained model that also works in zero- and few-shot settings. The authors support this with 16 benchmarks and 70 baselines, and they introduce a new protein–protein benchmark to fill the gap in standardized evaluation.","feed_headline":"One model scores and screens nearly all biomolecular complexes","feed_subtitle":"BioScore claims to match or beat 70 existing methods across 16 benchmarks spanning proteins, nucleic acids, drugs, and sugars.","key_machinery":"The load-bearing mechanism is a dual-scale geometric graph in which the nodes are block units (atoms for small molecules, amino acid residues for proteins, nucleobases for nucleic acids) carrying all-atom features and coordinates, processed by an E(3)-equivariant transformer. Two design choices carry the argument: interface-masking encoding, which hides inter-molecular edges during representation learning and is shown by ablation to improve distance-distribution fitting, and a dual-tower scoring module that separates docking/screening from scoring/ranking. The docking/screening tower builds a learned statistical potential through a mixture density network, replaces the traditional sum of log-probabilities with a mean, and adds an interaction-edge-count-aware confidence term, giving a score of the form $\\mathrm{Score} = -U_{\\mathrm{cplx}} + \\log n_{ij}$; the scoring/ranking tower fits an MLP to pairwise block features and also adds the edge-count confidence term. A pretraining–fine-tuning workflow trains the density network on unlabeled structures first, then adapts to labeled affinity data per system.","core_discovery":"The central claim is that BioScore is the first model to unify scoring, ranking, docking, and screening functionalities within a single framework for general biomolecular complexes, and that it consistently outperforms or matches 70 traditional and deep learning methods across 16 benchmarks. The authors report that pretraining on mixed protein–ligand and protein–protein structures improves protein–protein affinity prediction by up to 40% and antigen–antibody binding correlation by over 90%, that zero- and few-shot transfer to nucleic acid systems improves Pearson correlation by 15% and 71%, and that the unified representation improves protein–cyclic peptide affinity prediction by over 60%. Along the way, the paper introduces a PPI Benchmark with scoring, docking, and screening subsets built from a single data source, and reports that BioScore is the only tested method that effectively discriminates true binders in the PPI screening task.","pith_inferences":["Editorial inference: if the interface-masking finding generalizes, then generalist biomolecular encoders that emphasize interaction edges may be systematically limiting for scoring tasks; a testable extension is ablating inter-molecular edges in other generalist models.","Editorial inference: the PPI screening claim rests on decoys produced by one docking engine (ZDOCK) with K-means sampling, so a stronger test would use decoys from multiple docking engines and experimentally confirmed non-binders.","Editorial inference: the reported transfer gains suggest that as labeled data for RNA–ligand, glycan, and macrocycle systems accumulate, the same pretrain–fine-tune recipe can absorb them; this is a directly testable scalability prediction.","Editorial inference: if BioScore were wired into structure prediction and binder-design loops, it could serve as the filter that selects designed binders, but that would require confirmation that its ranking quality holds on predicted rather than experimental conformations."],"forward_implications":["One pretrained model can serve as a drop-in scorer for protein–ligand, protein–protein, protein–nucleic acid, nucleic acid–ligand, cyclic peptide, macrocycle, and carbohydrate complexes without system-specific re-engineering.","Cross-domain pretraining supplies transferable structural features that compensate for scarce affinity labels, enabling zero-shot prediction on nucleic acid–ligand and cyclic peptide systems and few-shot gains on protein–nucleic acid complexes.","The new PPI Benchmark offers a common evaluation standard with scoring, docking, and screening subsets built from one data source, so future protein–protein scoring functions can be compared on the same footing.","The interface-masking result implies that explicitly exposing inter-molecular edges during representation learning can hurt statistical-potential-based scoring, contradicting a common assumption in generalist biomolecular representation models.","The edge-count-aware confidence term converts the number of interaction edges into a discriminative signal, which is what the paper credits for its docking and screening gains across complex types."],"supporting_citations":[{"why":"Supplies the PDBbind v2020 structural and affinity data used for pretraining, fine-tuning, and several test sets.","marker":"[44]"},{"why":"Defines the CASF-2016 benchmark protocol and evaluation metrics that structure the PLI and PPI evaluations.","marker":"[43]"},{"why":"Provides the GET equivariant transformer encoder that BioScore adopts for unified feature extraction.","marker":"[41]"},{"why":"Is the prior RTMScore model whose residue–atom distance likelihood approach BioScore extends to general complexes and compares against.","marker":"[38]"},{"why":"Is GenScore, the prior PLI scoring framework BioScore builds on and surpasses, and the source of external screening docking poses.","marker":"[55]"},{"why":"Is ZDOCK, the docking engine used to generate decoys for the new PPI docking and screening benchmarks.","marker":"[45]"},{"why":"Supplies SAbDab antigen–antibody complexes used as an external PPI scoring test.","marker":"[53]"},{"why":"Provides the ITN peptide–MHC screening benchmark used for zero-shot PPI screening evaluation.","marker":"[54]"},{"why":"Is the CPSet protein–cyclic peptide benchmark used to test zero-shot generalization on macrocyclic ligands.","marker":"[66]"},{"why":"Supplies the CSM-carbohydrate training set and independent test set used for protein–carbohydrate evaluation.","marker":"[71]"}],"fun_headline_variants":["BioScore: one model for all biomolecular scoring","BioScore matches or beats 70 methods on 16 benchmarks","Unified scoring for proteins, nucleic acids, and sugars","BioScore tackles complexes from proteins to cyclic peptides","From drugs to sugars: BioScore scores it all"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The screening claims for protein–protein complexes depend on the newly built PPI Benchmark, whose decoys come from one docking program (ZDOCK) and K-means sampling, and on which BioScore is the only tested method that discriminates actives; if that protocol accidentally favors BioScore, the benchmark's screening evaluation and the generality claims built on it lose force.","fun_headline_variants_meta":{"raw":{"variants":["BioScore: one model for all biomolecular scoring","BioScore matches or beats 70 methods on 16 benchmarks","Unified scoring for proteins, nucleic acids, and sugars","BioScore tackles complexes from proteins to cyclic peptides","From drugs to sugars: BioScore scores it all"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000702,"raw_usage":{"total_tokens":3176,"prompt_tokens":959,"completion_tokens":2217,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":2139}},"tokens_in":575,"tokens_out":2217,"duration_ms":18833,"temperature":1.0,"reasoning_tokens":2139,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:23:00.217930+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BioScore on an independent PPI screening set built with decoys from multiple docking engines, or with experimentally characterized non-binders, and check whether true binders are consistently ranked above decoys; if its screening discrimination collapses or becomes comparable to baselines, the PPI screening claim fails. Alternatively, recompute PPI docking success rates after removing the interaction-edge-count confidence term; if the docking gain disappears, the dual-tower mechanism is the load-bearing component.","supporting_citations":[],"review_version":1}