{"id":"10c2e6fc-613e-4e97-ac4d-966aea813dae","arxiv_id":"2504.18646","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A new descriptor pipeline based on persistent Stanley-Reisner invariants reports slightly better protein-ligand and metalloprotein-ligand affinity predictions than selected older models.","lead":"This paper applies a new mathematical tool from commutative algebra, persistent Stanley-Reisner theory, to predict how tightly small molecules bind to proteins. It reports small gains over a few older models on two binding-affinity benchmarks, but the comparison is incomplete and the paper contains an internal dataset inconsistency.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PDBbind-v2016 test-set contradiction (§2.1: 290 vs Table 3: 195) makes the headline R=0.858 incomparable to the cited 0.848 baseline until the actual split is confirmed.","rationale":"The paper's mathematical contribution is plausibly novel, and a code link is provided, which is real evidence. I did not find an internal mathematical contradiction in the PSRT construction that would invalidate the method. The central empirical claim, however, is a comparison of R values against published baselines, and that comparison is only meaningful if the test sets are identical. The manuscript contains a direct contradiction: §2.1 describes the PDBbind-v2016 split as 3,768/290, while Table 3 reports 1,300/1,105/195. The reader spotted exactly this. I add that Table 1 and §2.1 also disagree on CAML(ES,CS) R (0.845 vs 0.853), and that the metalloprotein dataset citation differs between [40] and [48] in Table 3. These are not proof that the result is wrong, but they make the benchmark identity the least secure part of the argument. A single reproducible run of the released code, or even an inspection of the split files, settles it. Because this is addressable and does not attack the method's validity, the reader's CONDITIONAL verdict is appropriate; I would not change it.","tokens_in":14171,"tokens_out":6971,"duration_ms":64210,"concrete_test":"Clone the CAML repository at https://github.com/WeilabMSU/CAML and execute its PDBbind-v2016 pipeline without modification, recording the train/test partition used by the data-loading code. Then recompute the final CAML(ES,CS)+Transformer R and RMSE on the 290-complex core set used by TopBP-DL [23] and on the 195-complex split implied by Table 3. If R=0.858 occurs on the 195-complex split, the comparison to 0.848 is invalid; if it occurs on the 290-complex set, Table 3's counts are wrong and must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a comparative benchmark result, so it stands or falls on the test-set identity. Section 2.1 says PDBbind-v2016 has 'clearly defined training (3,768 complexes) and test sets (290 complexes)', but Table 3 lists PDBbind-v2016 as 1,300 total, 1,105 training, 195 test. These cannot both describe the same experiment. If the 195-complex split is what was actually used, the R=0.858 result is not measured on the same PDBbind-v2016 set on which TopBP-DL [23] reports R=0.848, and the claimed improvement is not established. The paper also contains a second numeric inconsistency: Section 2.1 gives CAML(ES,CS) R=0.853 while Table 1 lists 0.845. The metalloprotein comparison has the same vulnerability: Section 2.2 and Section 3.1 say the dataset is from [40], but Table 3 cites [48], the JPH-GBT baseline; if the split follows [40] rather than [48], the comparison to JPH-GBT R=0.742 is not apples-to-apples. These issues do not require any intent; each could be a typo, but as submitted the reproducibility trail does not pin down the benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces commutative algebra machine learning (CAML) for protein-ligand and metalloprotein-ligand binding affinity prediction. Building on persistent Stanley-Reisner theory (PSRT), the authors propose three featurization schemes: element-specific commutative algebra, category-specific commutative algebra, and commutative algebra on bipartite complexes. These descriptors are combined with gradient-boosted decision trees and, in the final PDBbind model, with transformer-based sequence embeddings (ESM-2 and a small-molecule transformer). The authors report Pearson correlation R = 0.858 and RMSE = 1.669 kcal/mol on PDBbind-v2016, and R = 0.755 on a metalloprotein-ligand benchmark, claiming state-of-the-art performance against persistent homology and persistent spectral baselines. The manuscript also provides a GitHub link for code and data availability.","tokens_in":14422,"tokens_out":2562,"duration_ms":24744,"significance":"If the reported benchmark results are valid, this paper demonstrates that commutative-algebra descriptors, specifically PSRT-derived facet Betti numbers, add predictive signal for structure-based affinity prediction beyond the compared persistent homology and persistent spectral methods. The work introduces a genuinely new descriptor class and pairs it with established machine learning tools, and it reports averages over 20 independent runs for the main results. However, the central claim is a comparative benchmark claim, and it currently rests on unresolved inconsistencies about the exact test set used for PDBbind-v2016 and about the metalloprotein dataset split. These issues must be resolved before the claimed state-of-the-art status can be accepted.","major_comments":[{"comment":"The test-set definition for the PDBbind-v2016 benchmark is internally contradictory. Section 2.1 states that PDBbind-v2016 has clearly defined training (3,768 complexes) and test sets (290 complexes), but Table 3 lists PDBbind-v2016 as 1,300 total, 1,105 training, and 195 test. These two descriptions cannot describe the same evaluation. If the actual split is the 195-complex one in Table 3, then the reported R = 0.858 is not obtained on the same PDBbind-v2016 test set used by TopBP-DL (R = 0.848), and the headline comparison is not established. The authors must clarify which split was used and, if the 195-complex split is correct, rerun or explicitly compare on the standard 290-complex refined-core test set.","section":"Section 2.1 and Table 3"},{"comment":"There is a numeric inconsistency for the CAML(ES,CS) model. The text in Section 2.1 reports R = 0.853 for this model, while Table 1 lists 0.845. Since the paper describes CAML(ES,CS)+Transformer as a consensus of CAML(ES,CS) and the Transformer model, both the component performance and the final consensus value need to be unambiguous. This discrepancy should be corrected and the reported values verified against the code outputs.","section":"Section 2.1 and Table 1"},{"comment":"The metalloprotein-ligand benchmark has the same kind of comparability problem. Section 2.2 and Section 3.1 state that the dataset is from reference [40], but Table 3 cites reference [48] for the metalloprotein-ligand row. If the training/test split follows [40] while the JPH-GBT baseline in Table 2 is taken from [48], the comparison to JPH-GBT R = 0.742 may not be apples-to-apples. The authors need to state explicitly which split was used and confirm that the baseline values correspond to that same split.","section":"Section 2.2, Section 3.1, and Table 3"},{"comment":"Table 1 reports only average values over 20 runs, without standard deviations or any measure of run-to-run variability, for the PDBbind-v2016 results. The paper's metalloprotein Table 2 includes error bars, so the omission in Table 1 is noticeable. Given that the claimed improvements over TopBP-DL (0.858 vs 0.848), PerSpect-ML (0.843), and PPS-ML (0.840) are small in absolute terms, reporting variability is necessary to assess whether these differences are meaningful.","section":"Section 2.1 and Table 1"}],"minor_comments":[{"comment":"The manuscript contains several typographical errors, including 'Stanely' in the Section 3.2 heading, 'comutative' in the Section 3.3 heading, 'Rensiner' in Section 2.1, 'dscriptors' in Section 3.5, 'o.745' in Section 2.2, and 'we we' in Section 3. These should be corrected.","section":"Throughout"},{"comment":"The RMSE formula appears malformed as printed: the display reads sqrt(1/n M sum ...), which is not a well-formed expression. The intended formula is presumably sqrt((1/M) sum (y_e - y_p)^2), and this should be typeset correctly.","section":"Section 3.6"},{"comment":"The sentence 'The final molecular descriptors are obtained by first vector among the 256 embedding vectors' is unclear. The authors should specify whether they take the first token embedding, an average over all token embeddings, or some other aggregation.","section":"Section 3.4.2"},{"comment":"The description of the PCA-based vectorization would benefit from an explicit statement of which features are concatenated for each atom-group point cloud, e.g., facet Betti numbers for 0- and 1-simplices and their rates, and how the filtration parameters (1 to 12 Å, step 0.5 Å) map to the final feature vector length.","section":"Section 3.3"},{"comment":"The consensus weighting between CAML(ES,CS) and the Transformer model is not described. Since the final CAML model is defined as the consensus of these two models, the paper should state how the consensus prediction is computed (e.g., simple averaging of predicted values, weighted averaging, or rank averaging).","section":"Section 3.5 and Table 5"}],"recommendation":"major_revision","confidential_remarks":"The two benchmark-identity inconsistencies (PDBbind test-set size and metalloprotein dataset citation) are the main obstacles to publication. They appear fixable by a careful re-analysis and by reporting results on the standard splits used by the cited baselines. I would recommend major revision rather than rejection, provided the authors can confirm the exact splits and recompute the comparisons on a well-defined common test set."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things up front. First, this is a real methodological novelty: persistent Stanley–Reisner facet invariants have not been applied to protein–ligand affinity prediction before, and the three vectorization variants (element-specific, category-specific, bipartite) are concrete additions. Second, the paper's headline result is on shaky ground because the data account is internally inconsistent. Section 2.1 says PDBbind-v2016 has 3,768 training and 290 test complexes, but Table 3 lists 1,300 total, 1,105 training, and 195 test. Both cannot describe the same experiment. If the 195-complex split is what was actually used, the reported R=0.858 is not measured on the same set as TopBP-DL's 0.848, so the claimed improvement is not established.\n\nWhat the paper does well: the PSRT framework is explained with enough detail to follow the math, the feature construction is clearly specified (cutoffs, filtration ranges, facet Betti numbers for 0- and 1-simplices), and the authors provide a GitHub link for code and data. Applying the same pipeline to metalloprotein–ligand binding with element- and category-specific interaction types is a sensible stress test, and the gains over JPH-GBT (0.755 vs 0.742) are plausible if the splits are aligned.\n\nThe soft spots are proportionate to the claim. The test-set contradiction is the big one. There is also a smaller numeric mismatch: Section 2.1 says CAML(ES,CS) has R=0.853, but Table 1 lists 0.845. The metalloprotein section cites reference [40] as the dataset source, while Table 3 attributes it to [48]; if the split follows JPH-GBT rather than MetalProGNet, the comparison to JPH-GBT is not apples-to-apples. The PDBbind results are reported without error bars, even though they are averages over 20 runs, and the state-of-the-art comparison omits several recent scoring functions. The consensus rule that combines CAML(ES,CS) and Transformer is described only as consensus without the weighting procedure, which is a reproducibility issue, though not a dealbreaker.\n\nNone of these flaws require the method to be wrong; they are all addressable. But as submitted, the benchmark trail does not pin down the actual split, and the headline comparison depends on exactly that.\n\nWho is this for? Computational chemists and TDA researchers who follow structure-based affinity prediction. It deserves a serious referee: the descriptor family is new, the baseline comparisons are mostly to independently published results, and the code is promised. I would condition my own verdict on the authors reconciling the split, adding error bars, and specifying the consensus rule. My recommendation: send to peer review, but make the dataset accounting a hard requirement before acceptance.","headline":"CAML introduces a genuinely new descriptor family for affinity prediction, but the central benchmark claim is currently unverifiable because the paper contradicts itself about which PDBbind-v2016 split was used.","tokens_in":15017,"tokens_out":1363,"would_cite":false,"duration_ms":14733,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["13P10","55N31","92C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"Commutative algebra descriptors push binding-affinity prediction to R=0.858 on a standard benchmark.","keywords":["persistent Stanley-Reisner theory","facet persistence barcodes","commutative algebra machine learning","protein-ligand binding affinity","metalloprotein-ligand binding","gradient boosting decision trees","topological data analysis","structure-based drug discovery"],"falsifier":"Run the released code on the PDBbind-v2016 split stated in Section 2.1 (3,768 training, 290 test) and check whether R=0.858 and RMSE=1.669 kcal/mol are reproduced; if the correlation falls below the R=0.848 of the persistent-homology baseline on that split, the claimed state-of-the-art result collapses. Similarly, reproducing R=0.755 on the metalloprotein split behind the JPH-GBT comparison would confirm the metal-ion result.","tokens_in":13954,"feed_emoji":"🧬","tokens_out":9806,"duration_ms":87637,"temperature":0.7,"pith_summary":"The paper proposes commutative algebra machine learning (CAML), a featurization strategy for predicting how tightly small molecules bind to proteins. The central move is to encode a protein-ligand complex as a family of simplicial complexes and track algebraic invariants of the associated Stanley-Reisner rings as a filtration parameter grows, producing descriptors that feed gradient-boosted decision trees. On the PDBbind-v2016 benchmark, the full model—element-specific and category-specific commutative algebra descriptors fused with transformer sequence embeddings—reaches Pearson correlation R=0.858 and RMSE=1.669 kcal/mol; on the metalloprotein-ligand benchmark, the category-specific model reaches R=0.755. These results are presented as surpassing earlier persistent homology and persistent spectral models. If the comparisons hold, commutative algebra contributes a new class of geometric descriptors for structure-based drug discovery.","feed_headline":"Commutative algebra tops binding-affinity benchmarks","feed_subtitle":"Persistent Stanley-Reisner descriptors plus sequence transformers reach R=0.858 on the standard protein-ligand set.","key_machinery":"The carrying object is the facet persistence barcode from persistent Stanley-Reisner theory. From a point cloud of atom coordinates, the method builds Rips or bipartite simplicial complexes, filters them by interatomic distance, and at each filtration level records the facet prime monomial ideals of the Stanley-Reisner ring; the persistent facet Betti number counts how many facet ideals survive from one level to the next. These survival counts, together with persistent graded Betti numbers via Hochster's formula and persistent f- and h-vectors, form the molecular descriptor. Descriptors are computed per element group, per amino-acid category, and for bipartite protein-ligand and metal-involving interactions, then fed to gradient-boosted decision trees; pretrained protein and small-molecule transformers supply a separate sequence-based model whose consensus with the structural model gives the final CAML prediction.","core_discovery":"The paper's claim is that persistent Stanley-Reisner theory, the commutative algebra of squarefree monomial ideals determined by the facets of a simplicial complex, can serve as a practical featurization engine for machine learning on biomolecular structures. Tracking how facet ideals, graded Betti numbers, and f- and h-vectors change under a distance filtration yields a vectorized fingerprint of a molecular complex, and the paper introduces three algorithmic variants—bipartite complexes, element-specific atom pairings, and category-specific amino-acid groupings—to capture the physics of protein-ligand and metalloprotein-ligand interactions. Combined with a consensus fusion of transformer-based sequence embeddings, this pipeline is reported to give state-of-the-art binding affinity predictions on both benchmarks.","pith_inferences":["If the conflicting split counts in the paper (Section 2.1: 3,768 training / 290 test; Table 3: 1,105 training / 195 test) mean the final model was evaluated on a split different from the baselines', the reported gain over TopBP-DL may not be reproducible on the baseline split; the paper does not reconcile the two counts.","A reader cannot tell from the paper alone how much of the 0.858 result comes from the commutative-algebra descriptors versus from the transformer sequence model or from consensus averaging; ablating each channel on the same split would separate those contributions.","The bipartite and category-specific constructions should transfer naturally to other two-component biomolecular systems, such as protein-protein, protein-DNA, or protein-RNA complexes, where pairwise interaction categories matter but standard persistent homology treats all atom pairs alike.","The persistent facet Betti curves are continuous functions of the filtration; making them differentiable inputs to a deep network is a natural next step that the paper does not explore."],"forward_implications":["On the standard protein-ligand benchmark (PDBbind-v2016), the full CAML model reaches Pearson R=0.858 and RMSE=1.669 kcal/mol, above the compared persistent homology model TopBP-DL (R=0.848), persistent spectral model PerSpect-ML (R=0.843), and path-spectral model PPS-ML (R=0.840).","On the metalloprotein-ligand benchmark, CAML(CS) reaches R=0.755 and CAML(ES) R=0.745, exceeding the compared JPH-GBT model at R=0.742.","The element-specific and category-specific variants are independently strong (R=0.836 and R=0.834), their consensus rises to R=0.845, and adding the transformer sequence model lifts the consensus to R=0.858, indicating that structural and sequence information are complementary.","Because the descriptors are built directly from atom coordinates and sequences, the same CAML pipeline transfers to other biomolecular property prediction problems without redesigning the featurization.","For metalloprotein complexes, the bipartite-complex construction explicitly represents protein-metal and metal-ligand interactions, allowing the model to use 138 atom-pair combination types rather than treating all atoms alike."],"supporting_citations":[{"why":"It supplies the persistent Stanley-Reisner theory that every CAML descriptor in the paper is built from.","marker":"[31]"},{"why":"It provides the protein-ligand benchmark dataset and its experimental binding affinity labels.","marker":"[39]"},{"why":"It supplies the persistent-homology TopBP-DL baseline (R=0.848) that the CAML protein-ligand claim must beat.","marker":"[23]"},{"why":"It supplies the PerSpect-ML persistent spectral baseline (R=0.843) used in the comparison.","marker":"[26]"},{"why":"It supplies the PPS-ML path-spectral baseline (R=0.840) used in the comparison.","marker":"[41]"},{"why":"It provides the metalloprotein-ligand dataset and the MetalProGNet baseline model.","marker":"[40]"},{"why":"It supplies the JPH-GBT metalloprotein baseline (R=0.742) that CAML claims to exceed.","marker":"[48]"},{"why":"It provides the pretrained protein language model used to build the sequence-based transformer descriptor.","marker":"[51]"},{"why":"It provides the pretrained small-molecule language model used to embed ligand sequences.","marker":"[52]"},{"why":"It established the element-specific interaction modeling strategy that the CAML element-specific descriptors extend.","marker":"[37]"}],"fun_headline_variants":["Commutative algebra wins binding-affinity race","Stanley-Reisner fingerprints sharpen binding predictions","CAML: algebra-based AI for protein-ligand affinity","New algebraic method tops drug-binding benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported numbers were computed on the same benchmark split and with the same metric conventions as the earlier models they are compared to; the paper's own Section 2.1 (3,768 training, 290 test) and Table 3 (1,105 training, 195 test) give conflicting split counts for the protein-ligand benchmark, so if the actual split differs from the baselines' split, the headline R=0.858 is not comparable.","fun_headline_variants_meta":{"raw":{"variants":["Commutative algebra wins binding-affinity race","Stanley-Reisner fingerprints sharpen binding predictions","CAML: algebra-based AI for protein-ligand affinity","New algebraic method tops drug-binding benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000356,"raw_usage":{"total_tokens":1874,"prompt_tokens":832,"completion_tokens":1042,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":982}},"tokens_in":448,"tokens_out":1042,"duration_ms":10223,"temperature":1.0,"reasoning_tokens":982,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:13:07.701946+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on the PDBbind-v2016 split stated in Section 2.1 (3,768 training, 290 test) and check whether R=0.858 and RMSE=1.669 kcal/mol are reproduced; if the correlation falls below the R=0.848 of the persistent-homology baseline on that split, the claimed state-of-the-art result collapses. Similarly, reproducing R=0.755 on the metalloprotein split behind the JPH-GBT comparison would confirm the metal-ion result.","supporting_citations":[{"cited_title":"Pdb-wide collection of binding data: current status of the pdbbind database","cited_arxiv_id":null,"evidence_quote":"It provides the protein-ligand benchmark dataset and its experimental binding affinity labels."},{"cited_title":"Persistent Stanley--Reisner Theory","cited_arxiv_id":"2503.23482","evidence_quote":"It supplies the persistent Stanley-Reisner theory that every CAML descriptor in the paper is built from."},{"cited_title":"Persistent path-spectral (pps) based machine learning for protein–ligand binding affinity prediction","cited_arxiv_id":null,"evidence_quote":"It supplies the PPS-ML path-spectral baseline (R=0.840) used in the comparison."},{"cited_title":"Metalprognet: a structure-based deep graph model for metalloprotein–ligand interaction predictions","cited_arxiv_id":null,"evidence_quote":"It provides the metalloprotein-ligand dataset and the MetalProGNet baseline model."},{"cited_title":"Join persistent homology (jph)-based machine learning for metalloprotein–ligand binding affinity prediction","cited_arxiv_id":null,"evidence_quote":"It supplies the JPH-GBT metalloprotein baseline (R=0.742) that CAML claims to exceed."},{"cited_title":"Extracting predictive representations from hundreds of millions of molecules.The journal of physical chemistry letters, 12(44):10793– 10801, 2021","cited_arxiv_id":null,"evidence_quote":"It provides the pretrained small-molecule language model used to embed ligand sequences."},{"cited_title":"TopologyNet: Topology based deep convolutional and multi- task neural networks for biomolecular property predictions","cited_arxiv_id":null,"evidence_quote":"It established the element-specific interaction modeling strategy that the CAML element-specific descriptors extend."}],"review_version":1}