{"id":"82fa2e39-ee94-4971-b0a0-ac1df529bbcf","arxiv_id":"2607.25115","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Persistent manifold learning with Boundary-Induced Graph Laplacians plus language-model features beats prior SOTA Pearson correlation on metalloprotein–ligand and SKEMPI wild-type protein–protein affinity benchmarks.","lead":"A geometry-plus-topology method turns binding interfaces into multiscale manifolds and predicts how tightly proteins bind ligands or other proteins better than prior scores on two standard benchmarks. Drug discovery and protein design both need cheaper affinity ranking; a single shape-based descriptor that works for metal pockets and flat protein surfaces is practically useful.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline claim is a geometric-framework claim, but no ablation ever separates the manifold features from the ESM-2/ChemBERTa embeddings they are concatenated with — the 0.008 (MPLI) and ~0.03 (PPI) PCC edges over prior SOTA could be carried entirely by the language-model features and GBDT.","rationale":"The reader's weakest_assumption already names this concern in its second half (\"if performance is driven mainly by LM embeddings/GBDT rather than the manifold features, the SOTA claim does not establish the geometric framework\"), and my read confirms it is the right load-bearing point rather than the spectral-truncation half. On truncation: §5.3 gives a plausible, checkable defense (high modes approach grid spacing ℓ and are discretization-contaminated; retaining more eigenvalues was tried and degraded performance), so that assumption is weaker as an attack surface. The missing feature-source ablation, by contrast, is unaddressed anywhere in the paper and directly determines whether the central claim — a geometric framework, not just a strong feature soup — is established. I keep the verdict at CONDITIONAL rather than REJECT because: (a) the concern is a missing experiment, not a demonstrated error — the ablation could well come out in the paper's favor, and the manifold-only arm would itself be informative; (b) the math exposition (§2) is careful and standard, and the BIG-Laplacian pipeline has independent published support ([41], [47]); (c) the empirical results, while metric-selective on PPI (worse RMSE than PLNet) and thin on MPLI (+0.008 PCC), are reported with seed-level variance and are internally consistent enough to be real. The feature-count inconsistency (3 vs 6 per manifold, §3.4 vs §5.2) and the CV-vs-test-set ambiguity (§4 vs §4.1.1) are reporting-quality issues that reinforce the conditional grade and should be corrected, but they do not independently overturn the claim. If the proposed ablation shows the LM-only arm matching the full model, the verdict should drop; if it shows additive contribution from manifold features on both benchmarks, the paper's central claim is substantially strengthened and could merit ACCEPT.","tokens_in":23585,"tokens_out":2605,"duration_ms":103906,"concrete_test":"On the MPLI benchmark, retrain the identical GBDT (Table 1 hyperparameters, same 10 seeds, same train/test split from [53]) on three feature sets: (i) ESM-2+ChemBERTa embeddings only, (ii) manifold embedding features only, (iii) the concatenation. If (i) reaches PCC ≥ 0.745 (CAML's number), the SOTA edge is attributable to the LM features and the geometric claim is unsupported; if (iii) exceeds both (i) and (ii) by more than the seed-level std (~0.002), the manifold features contribute genuinely additive signal. Repeat with the same folds on SKEMPI-WT, where the manifold:LM feature ratio is most unfavorable (486 vs 2560).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is not merely \"we built a good BA predictor\" but that a *manifold-based geometric representation* drives the SOTA result — the Relevance statement says \"much of what determines binding strength is encoded in the shape of the interface itself,\" and §5 is devoted to explaining why the manifold features work. Yet the only ablation reported (§4.1.2, §4.2) holds the feature set fixed and varies the regressor (GBDT vs SVM). There is no experiment with LM-embeddings-only, manifold-features-only, or leave-one-source-out. This matters quantitatively: on MPLI the margin over CAML is 0.753 vs 0.745 PCC — a 0.008 edge that ESM-2(1280-d)+ChemBERTa(384-d)+GBDT with 10,000 trees could plausibly produce on its own, since large LM embeddings are known to be strong BA predictors. On PPI the imbalance is worse: 486 manifold features vs 2560 ESM-2 features, and §5.2 concedes the ESM-2 embeddings get \"proportionally greater weight.\" The claim that PML \"outperforms PLNet... particularly for flatter interfaces\" attributes the gain to manifold geometry, but PLNet uses persistent-Laplacian features with a neural net and (as far as stated) no ESM-2 — so the comparison confounds representation and feature source simultaneously. Secondary cracks: §3.4 says \"6 features per manifold\" then computes \"3×9×40,\" while §5.2 uses 6×9×9=486 — an internal inconsistency in the basic feature accounting; §4 states metrics come from 10-fold CV, but §4.1.1 reports a fixed external test split from [53], and for SKEMPI-WT no split protocol for the PLNet comparison is specified at all; and on PPI, PML's RMSE (2.051) is substantially worse than PLNet's (1.533), so \"outperforms\" holds only for PCC. None of these is fatal, but the missing ablation is the load-bearing gap: without it, the paper demonstrates a good ensemble, not a validated geometric framework.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript introduces Persistent Manifold Learning (PML) for binding affinity (BA) prediction. A binding interface is converted, via element-specific Gaussian (FRI) density fields, into a filtration of sublevel-set manifolds on a Cartesian grid; the Boundary-Induced Graph (BIG) Laplacian L_{3,n} under normal boundary conditions is assembled on each manifold, and β0 plus the first two nonzero eigenvalues per manifold are used as features. These are concatenated with ESM-2 (protein) and ChemBERTa (ligand) embeddings and regressed with GBDT. On a metalloprotein–ligand benchmark (fixed 1845/618 split) PML reports average PCC 0.753 vs 0.745 for the prior SOTA (CAML); on SKEMPI-WT protein–protein BA it reports average PCC 0.713 (best run 0.731) vs 0.681 for PLNet. Section 2 reviews de Rham–Hodge theory, the Friedrichs theorem, Eulerian DEC discretization, and persistent Hodge Laplacians; the mathematical review is competent and correctly cited.","tokens_in":24182,"tokens_out":5750,"duration_ms":191071,"significance":"If the results hold, the paper demonstrates that a single geometric representation (BIG-Laplacian spectra on density sublevel sets) transfers across two structurally distinct interaction classes — compact metal-coordination sites and flat PPI interfaces — which would be a useful unification. Strengths worth naming: the empirical protocol is better than much of this literature (fixed external test split for MPLI, 10 seeds with mean±std, 10-fold CV, an SVM regressor ablation, tabulated baselines); the mathematical development is a correct, well-cited recall of Hodge/Friedrichs theory and the BIG approximation; and the benchmark claims are concrete and falsifiable. However, the framing claim — that interface geometry drives the SOTA result — is not isolated from the contribution of the language-model embeddings, and on the PPI benchmark PML is worse than PLNet on RMSE, so the current evidence supports a narrower claim than the abstract and Relevance statement make.","major_comments":[{"comment":"No ablation separates the manifold features from the LM embeddings they are concatenated with. The only ablation varies the regressor (GBDT vs SVM) with the feature set held fixed. Yet the paper's framing claim is geometric ('much of what determines binding strength is encoded in the shape of the interface itself'; §5 attributes the PPI gain to manifold geometry). On MPLI the margin over CAML is 0.008 PCC (0.753 vs 0.745, Table 2), which a 1664-d ESM-2+ChemBERTa input to a 10,000-tree GBDT could plausibly produce alone; on PPI, PLNet [58] apparently uses no ESM-2, so the comparison confounds representation and feature source, and §5.2 concedes ESM-2 gets 'proportionally greater weight'. LM-only, manifold-only, and leave-one-source-out runs (or GBDT feature-importance per source) are required to support the geometric claim; these are cheap given the existing pipeline.","section":"§4.1.2, §4.2, §5.2; Relevance statement"},{"comment":"On SKEMPI-WT, PML does not outperform PLNet on RMSE: 2.051±0.046 (best 1.996) vs PLNet's 1.533±0.021 — roughly 34% worse — and PML-SVM is worse still (2.127). The statement in §4.2 that 'PML with GBDT consistently achieves the best performance' and the abstract's unqualified 'outperforms state-of-the-art methods' are therefore supported only for PCC. The large RMSE regression despite higher PCC (a calibration/shape issue, possibly from GBDT on 343 samples) needs explicit reporting, qualification of the SOTA claim, and discussion.","section":"Table 3, §4.2, Abstract"},{"comment":"The feature accounting is internally inconsistent. §3.4 states 'β0 together with the first 2 non-zero eigenvalues... giving 6 features per manifold' — that is 3 numbers, not 6 — and then computes '3×9×40' (=1080) features per MPLI complex, while §5.2 uses 6×9×9=486 for PPI and quotes 2160 for MPLI (=6×9×40). Which is it, and where does the factor of 2 come from? Similarly §3.2 says PPI pairs are restricted to {C,N,O}×{C,N,O} (=9) but then writes 'pairs between {C,N,O,S} and {C,N,O,S}' (=16). These are basic counts needed to reproduce the input dimensionality.","section":"§3.4 vs §5.2; §3.2"},{"comment":"§4 states that 'PCC and RMSE are computed via 10-fold cross-validation rather than on a single train-test split,' but §3.1 and §4.1.1 describe a fixed 1845/618 train/test partition with 10-fold CV used only for hyperparameter tuning, and Table 2 is a test-set comparison. For SKEMPI-WT the evaluation protocol (CV folds? a held-out split? alignment with PLNet's protocol in [58]) is never stated. Please state precisely which protocol produced each number in Tables 2 and 3, and confirm that all baseline numbers were obtained on identical splits and label conventions (the Table 2 caption says RMSE is on raw pKd labels while the column is labeled kcal/mol).","section":"§4 (metrics paragraph) vs §3.1/§4.1.1"},{"comment":"Sections 2.3–2.4 develop the genuinely persistent Hodge Laplacian ∆_{i,j} (biharmonic extension, commutative diagrams, discrete assembly), but Remark 2.13 states that only the i=j special case is used, stacked over 9 isovalues. The implemented method is therefore single-scale Hodge/BIG spectra evaluated repeatedly along a filtration — closer to the Eulerian method of [47] than to the persistent theory presented. Either the framing/title should be aligned with what is computed, or the authors should justify why stacking i=j spectra constitutes 'persistent' manifold learning in the sense of §2.3 (e.g., what the persistent construction would add and why it is unnecessary here).","section":"§2.3–2.4, Remark 2.13; title/abstract"}],"minor_comments":[{"comment":"The scale parameter τ in ρ(x,τ) is never assigned a value, and the source of the van der Waals radii r_i is not given; both are needed for reproduction. Also, the GBDT implementation is described only as 'the Python library' (§3.6) — name the package and version.","section":"§3.2, Eq. (3.1)"},{"comment":"The claim that extending the spectrum to ten eigenvalues 'degrades performance on both benchmarks' is asserted without data. Either show the experiment (a small table would suffice) or soften the claim.","section":"§5.3"},{"comment":"Table 2 caption says RMSE is computed on raw pKd labels but the column header and figures use kcal/mol; Figure 3(b) caption reads 'PCC 0.756 kcal/mol' — PCC is unitless. Also, comparing the best of 10 seeds (0.756, 0.731) against baselines' reported means is generous; consider a significance test on the average-run margin (0.753±0.002 vs 0.745±0.001).","section":"Table 2 / Figure 3"},{"comment":"§3.3 builds the bounding box per complex from atom coordinate min/max, while §3.4 says 'a fixed Cartesian grid that remains the same across all complexes.' These appear contradictory; clarify whether boxes (and hence grid spacing ℓ) vary per complex, and if so, why spectra remain comparable across complexes given different ℓ.","section":"§3.3 vs §3.4"},{"comment":"'We evaluate the level-set function ρ(x,τ) from Eq. (2.1)' — wrong cross-reference; ρ is defined in Eq. (3.1). Eq. (2.1) is dd=0.","section":"§3.3"},{"comment":"With 343 SKEMPI-WT complexes and ~3000 input features, please state whether model selection used nested CV to avoid leakage, and report the PLNet protocol explicitly enough that the comparison is apples-to-apples (PLNet [58] is the authors' own prior method).","section":"§4.2 / Table 3"},{"comment":"Code is 'available from the authors upon reasonable request.' For a method paper whose claims are empirical, a public repository (GitHub/Zenodo with a DOI) including the feature-extraction pipeline and trained-model scripts would substantially strengthen reproducibility.","section":"Declarations"},{"comment":"Typographical: 'Metallprotein-ligand' (§4.1 heading); 'RMSE are computed' (§4); 'protein-protein' vs 'protein–protein' dash inconsistency; Figure 1 is low-resolution and its panel labels (e.g., 'Hodge Laplaciand') run together.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The SOTA baselines on both benchmarks are largely the group's own prior methods (PLNet [58], CAML [16], and the JPH-GBT lineage [53,54]), and the citation pattern is heavily self-referential; this is not disqualifying — the methods lineage is genuine — but the 'new state of the art' claims rest on margins (0.008 PCC on MPLI) that are small relative to protocol differences, and the headline numbers quoted are best-of-10-seed runs. The missing LM-only ablation is the one experiment I would insist on before acceptance; it is inexpensive and directly tests the paper's framing claim."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a solid methods-and-benchmarks paper in the group’s persistent de Rham–Hodge / BIG-Laplacian program, applied cleanly to two BA settings. The SOTA edges are real on the numbers they report, but small, and the claim that interface shape is doing the work is not isolated from the language-model features.\n\nWhat is new is the packaging, not the operators. Element-specific FRI sublevel sets, Eulerian BIG L_{3,n} (β0 plus two nonzero eigenvalues, nine isovalues), ESM-2/ChemBERTa, GBDT — all built on their cited stack ([6], [41], [47], PLNet, etc.). The useful move is running one geometric pipeline on both compact metal sites and flat PPIs, with an honest discussion in §5 of why packing/coordination might show up in the filtration. Math sections are careful: Hodge/Friedrichs, BIG approximation, spectral reduction to singular values. Empirical protocol is better than average for this subfield — fixed MPLI split, 10 seeds, CV on train, SVM-vs-GBDT ablation, tabulated baselines.\n\nSoft spots, in proportion. The stress-test lands: there is no manifold-only, LM-only, or leave-one-source-out run. On MPLI the margin over CAML is ~0.008 PCC; on PPI, manifold features are outnumbered by ESM-2 (486 vs 2560) and they concede LM gets more weight, while RMSE is worse than PLNet’s. So “manifold framework drives SOTA” is an ensemble result, not a dissected one. Minor: feature-count wording (3×9×40 vs 6×9×9), “outperforms” leaning on PCC, code on request, free parameters (τ, grid, truncation) not swept hard. None of that sinks the paper; it just caps how much geometric credit you can give.\n\nWho it’s for: people already in topological/spectral biomolecular ML or BA feature design. Not a foundational biology result. Math and citations look solid within their line; circularity is low (external affinity labels).\n\nI’d send it to peer review. Ask referees for the ablation and a clearer split protocol on SKEMPI-WT. Worth engaging if you work this area; not urgent if you don’t.","headline":"Competent dual-benchmark extension of their Hodge/BIG-Laplacian line; real but small gains, and the missing LM-vs-manifold ablation undercuts the geometric story.","tokens_in":24413,"tokens_out":586,"would_cite":false,"duration_ms":19312,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92E10","55N31"],"pacs":[],"model":"grok-4.5","headline":"A single multiscale manifold description of binding interfaces predicts affinity for both metal pockets and flat protein surfaces better than prior methods.","keywords":["persistent manifold learning","de Rham-Hodge theory","Hodge Laplacian","Boundary-Induced Graph Laplacian","binding affinity prediction","metalloprotein–ligand","protein–protein interactions","Betti numbers"],"falsifier":"Retrain and retest on the same held-out metalloprotein–ligand and SKEMPI-WT splits using only the language-model embeddings (no manifold spectra) or only higher/truncated spectra; if the reported PCC gains disappear, the geometric claim is not supported.","tokens_in":23955,"feed_emoji":"🧬","tokens_out":887,"duration_ms":17928,"temperature":0.7,"pith_summary":"How tightly two biomolecules stick together is hard to predict from structure alone, especially when the contact region can be a tight metal-coordinated pocket or a broad, nearly featureless protein surface. This paper argues that much of what sets binding strength is encoded in the shape of that interface, and that one geometric description can serve both classes without hand-crafted, system-specific features. It turns each interface into a nested family of manifolds built from element-specific Gaussian density fields, then reads topology and geometry from the spectra of Boundary-Induced Graph Laplacians (a discrete form of de Rham–Hodge theory). Those manifold embeddings are concatenated with sequence embeddings from protein and molecular language models and fed to gradient boosting trees. On standard metalloprotein–ligand and wild-type protein–protein benchmarks the resulting model beats previously reported state-of-the-art scores, supporting the claim that a shared manifold view of interface shape is enough to drive competitive affinity prediction across dissimilar interaction types.","feed_headline":"One manifold view beats prior affinity models on metal and PPI interfaces","feed_subtitle":"Multiscale geometry plus language embeddings predicts binding for both pocket types without hand-tuned features","key_machinery":"Boundary-Induced Graph Laplacian (BIG Laplacian) on degree-3 forms under normal boundary conditions: a discrete de Rham–Hodge operator whose zero eigenvalues recover the 0-th Betti number and whose leading nonzero eigenvalues supply multiscale geometric descriptors of each manifold in the filtration.","core_discovery":"Persistent manifold learning (PML) represents a binding interface as a filtration of manifolds from element-specific atomic density fields and extracts both Betti numbers and nonharmonic spectral geometry via Boundary-Induced Graph Laplacians; when these features are combined with language-model embeddings and gradient boosting, the same pipeline outperforms prior methods on both metalloprotein–ligand and protein–protein binding-affinity benchmarks.","pith_inferences":["If the truncated spectra are doing real work, ablating element-specific pairing or the metal-coordination shell cutoff should hurt metalloprotein performance more than PPI performance.","The framework’s reliance on static crystal coordinates suggests a natural next test on ensembles or flexible docking poses to see whether leading eigenvalues remain stable.","Because language-model embeddings already carry substantial signal, the manifold block is most valuable where sequence alone under-resolves interface packing."],"forward_implications":["One geometry-first pipeline can replace separate feature engineering for metal-centered pockets and broad protein–protein interfaces.","Affinity ranking for virtual screening can be driven by multiscale manifold spectra rather than pocket-specific handcrafted descriptors.","The same level-set plus BIG-Laplacian construction is in principle reusable for other structure–sequence quantitative tasks beyond binding free energy.","Spectral truncation to leading modes is presented as a resolution-matched regularizer, not a temporary limitation."],"fun_headline_variants":["PML manifolds beat prior models on metal and PPI affinity","Multiscale manifold features top metalloprotein and PPI benchmarks","Persistent manifolds plus language embeddings lift binding prediction","One manifold pipeline outperforms affinity models across interface types","Boundary-induced Laplacians capture interface geometry and topology"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That the 0-th Betti number plus only the first two nonzero eigenvalues of these Laplacians, taken on nine fixed isovalues of element-pair Gaussian densities, already carry enough binding-relevant shape information for both metal sites and flat protein interfaces.","fun_headline_variants_meta":{"raw":{"variants":["PML manifolds beat prior models on metal and PPI affinity","Multiscale manifold features top metalloprotein and PPI benchmarks","Persistent manifolds plus language embeddings lift binding prediction","One manifold pipeline outperforms affinity models across interface types","Boundary-induced Laplacians capture interface geometry and topology"]},"model":"grok-4.5","effort":"low","cost_usd":0.002558,"raw_usage":{"total_tokens":904,"prompt_tokens":662,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":25584000,"prompt_tokens_details":{"text_tokens":662,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":181,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":662,"tokens_out":61,"duration_ms":5032,"temperature":1.0,"reasoning_tokens":181,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T00:47:58.403994+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain and retest on the same held-out metalloprotein–ligand and SKEMPI-WT splits using only the language-model embeddings (no manifold spectra) or only higher/truncated spectra; if the reported PCC gains disappear, the geometric claim is not supported.","supporting_citations":[],"review_version":1}