{"id":"0d1e9fd6-9692-4fd5-8cb9-a1229646e7c0","arxiv_id":"2506.06294","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GLProtein pre-trains a BERT-style protein model with triplet contrastive learning on structural similarity, 3D distance encoding, and substructure-based amino acid embeddings, improving contact and PPI prediction.","lead":"Protein language models usually ignore how a protein folds and how its shape resembles other proteins. GLProtein adds both global structure similarity and local amino acid details to a protein language model, and reports gains on contact and interaction prediction tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3) triplet loss compares whole-sequence matrices of differing row counts; without a specified pooling or length-alignment step the global pretraining objective is undefined for varying-length triplets.","rationale":"I agree with the reader's identification. The central claim that GLProtein improves contact and PPI prediction rests on the joint objective in Eq. (5) being computable and on the global triplet loss contributing. The shape mismatch is the most load-bearing concern because it is not a probabilistic claim or a tuning issue; it is a mathematical well-definedness problem in the core formula. It is also fixable: a sentence specifying mean pooling would resolve it, and the empirical results could still hold under that interpretation. Therefore the correct verdict remains CONDITIONAL, meaning the reader's verdict should be unchanged. The paper needs a correction and reproducibility artifacts before the claims are fully trusted. I would not escalate to REJECT because the idea is plausible, the limitations are candid, and the missing detail is likely an omission rather than a fundamental flaw. However, without the clarification the method is currently unimplementable as written.","tokens_in":20354,"tokens_out":5288,"duration_ms":55113,"concrete_test":"Take two Swiss-Prot sequences of different lengths (e.g., 100 and 150 residues) that form a positive pair, and compute Eq. (3) exactly as written after the encoder; the operation ||E_p - E_pos_p||_2 fails because E_p is R^{100 x D} and E_pos_p is R^{150 x D}. To settle whether a hidden pooling exists, re-run the pretraining loop with the authors' code (when released) on a small subset containing triplets of unequal lengths without adding any pooling; a runtime shape error confirms the objective is undefined as stated. If code is unavailable, request the authors to specify whether mean pooling, residue-wise alignment, or length filtering is used; the absence of such a description makes Eq. (5) non-reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The global structure information modelling component in Section 3.2 defines the triplet loss L_PTL = max(||E_p - E_pos_p||_2 - ||E_p - E_neg_p||_2 + epsilon, 0), with E_p, E_pos_p, E_neg_p in R^{L_p x D}. The paper states that L_p is the length of the amino acid sequence, and positive and negative proteins are sampled from the database independently of the anchor's length. For any triplet whose three proteins have different lengths, the matrix differences and the L2 norms are dimensionally invalid: subtracting R^{150 x D} from R^{100 x D} is undefined. No pooling, padding, masking, or length-restricted sampling is described in Section 3.2 or in the implementation details, and the appendix's complexity analysis (Section A.9) treats the triplet loss as O(3Ld), implicitly assuming a single length L. Since L_PTL is the only global-structure term in the joint objective Eq. (5), the pretraining pipeline as written cannot be executed or reproduced on real data of varying lengths. This is not a cosmetic typo: the gradient of L_PTL with respect to the encoder is the mechanism intended to inject global structure similarity into the representation, and its shape behaviour determines whether the method works at all. The limitation section honestly reports weaker results on binding affinity and fluorescence, and the ablation study suggests the global component matters, but those results are uninterpretable until the length-alignment step is specified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GLProtein, a protein representation learning framework that combines a ProtBert-style masked language model with two structure-aware modules. Global structure information is injected through a triplet loss on protein representations, with positive and negative proteins selected by TM-Vec based on TM-score structure similarity; local structure information is injected through AlphaFoldDB-derived 3D distance encodings and mol2vec substructure encodings in the decoder. The model is pre-trained on Swiss-Prot and evaluated on contact prediction, PPI, TAPE, function prediction, binding affinity and stability, reporting strong results on contact and PPI and mixed results on other tasks, which the limitations section acknowledges.","tokens_in":20650,"tokens_out":7030,"duration_ms":66899,"significance":"If the objective definitions are corrected, the paper's contribution is a plausible and potentially useful integration of global structural similarity and local molecular detail into a single protein pre-training pipeline. The evaluation is broad, includes ablations and parameter sensitivity studies, and the authors are transparent about tasks where GLProtein does not improve over baselines. The use of TM-Vec as an external structural similarity oracle is a legitimate design choice and not circular. However, the two mathematical errors in the objectives are load-bearing and currently prevent the method from being executed or reproduced as written.","major_comments":[{"comment":"The masked language modeling loss is written as L_MLM = -log sum_{i in M} P(x_i | x_sim i; theta_E, theta_D). Standard masked LM uses the sum of per-token cross-entropy terms, i.e., -sum_i log P(...), because each masked position has its own softmax distribution; the negative log of the sum of these probabilities has no probabilistic interpretation and would lead to a different and incorrect training signal. Since this term is half of the joint objective in Eq. (5), the training objective as written is mathematically incorrect and must be corrected.","section":"Section 3.1, Eq. (1)"},{"comment":"The triplet loss compares E_p, E_pos_p and E_neg_p, each stated to be in R^{L_p x D}. Positive and negative proteins are sampled independently in Section 3.2, so their lengths generally differ from the anchor length L_p; the matrix differences in Eq. (3) are therefore undefined unless a length-alignment step such as pooling, padding with masking, or residue-level alignment is applied. No such step is described in Section 3.2, in the implementation details, or in the appendix, and the complexity analysis in Section A.9 implicitly assumes a single length L by reporting O(3Ld). This is not a cosmetic issue, because L_PTL is the only global-structure term in Eq. (5) and its gradient is the mechanism that injects global structural similarity into the representation. The authors should specify the exact pooling/alignment procedure and revise the complexity analysis accordingly.","section":"Section 3.2, Eq. (3)"}],"minor_comments":[{"comment":"The citation for ESM-2 points to Beal (2015), which is a paper about experience sampling methods in organizational research, not the protein language model ESM-2; this reference must be corrected.","section":"References"},{"comment":"The conclusion contains the typo 'GLProtien' and should read 'GLProtein'.","section":"Section 5"},{"comment":"'Breath-First Search' should be 'Breadth-First Search'.","section":"Table 2 caption"},{"comment":"The claim that GLProtein is 'the first framework' to incorporate global structural similarity and local amino acid details is too strong without a more complete comparison against prior structure-aware methods; the paper's own Table 3 and Table 4 show ties and losses on several tasks, so the outperformance claim should be qualified.","section":"Abstract and Introduction"},{"comment":"Appendix A.7 says it studies 'the number of protein samples in the protein local structure information modelling component', but the sampled positives and negatives are used by the global triplet loss, so the wording should be corrected.","section":"Section 3.4 / Appendix A.7"},{"comment":"The Gaussian basis kernel is written with a negative prefactor -1/(sqrt(2 pi) |sigma_k|), which is unusual for a radial basis function; if this sign is intentional, the authors should explain it, and the role of |sigma_k| as a learnable scaling factor should be clarified.","section":"Eq. (4)"},{"comment":"The global structure component is described as leveraging 'self-supervised signals', but the triplet labels are produced by TM-Vec, a separately trained model, and the structures come from AlphaFoldDB; the authors should describe this as external supervision or knowledge-based sampling rather than self-supervision.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The two objective errors in Eqs. (1) and (3) are fixable in principle, but they are central to the method, so the paper should not be accepted in its current form. The empirical evaluation is extensive and the limitations are honestly reported, which supports giving the authors a chance to revise. The ESM-2 reference error suggests the bibliography should be checked carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GLProtein is a sensible integration of known ingredients—masked protein LM, triplet contrastive learning with TM-Vec-selected positives/negatives, Gaussian 3D distance encoding, and mol2vec substructure embeddings—and that combination appears to be new. The paper runs a broad baseline set, reports ablations where each component helps, and honestly flags tasks where it loses (binding affinity, fluorescence). The central idea is worth taking seriously. But there is a load-bearing gap. Eq. (3) defines the triplet loss on whole-sequence matrices E_p, E_pos, E_neg in R^{L_p x D}; once the three proteins in a triplet have different lengths, both the subtraction and the L2 norm are undefined. The paper specifies no pooling, padding, or length-restricted sampling, and the appendix complexity O(3Ld) silently assumes a single length. Since this loss is the only global-structure term in the joint objective, the pre-training pipeline as written cannot be run on variable-length proteins. It is very likely fixable with a mean-pooling sentence, but until that is specified, the global component is not reproducible. Eq. (1) is also wrong as written—a -log of a sum instead of a sum of -log terms—though that one reads as a typo. The SaProt comparison is handled in a way that may understate SaProt: replacing its structural tokens with '#' is not how SaProt was designed to be used, so the outperformance claim against SaProt should be read cautiously. No code is released. Tuning alpha and the number of samples on validation is standard, so I treat that as minor. The citation list covers the relevant work; self-citations are not load-bearing. Net: the empirical story is plausible but not yet trustworthy as written. I would send this to peer review: a serious referee should require the pooling step, corrected objectives, and ideally code plus an honest SaProt comparison. I would bring it to a reading group in the meantime—it is a good case study in how structural similarity is injected into protein LMs—but I would not cite it yet.","headline":"A promising integration of global structural similarity and local substructure signals into a protein LM, but the triplet loss as written is undefined for variable-length proteins and needs a pooling specification plus corrected objectives before the empirical claims can be trusted.","tokens_in":21159,"tokens_out":2899,"would_cite":false,"duration_ms":29714,"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":"GLProtein claims that pre-training on global structure similarity plus local amino-acid geometry improves protein contact and interaction prediction.","keywords":["protein representation learning","protein language model","structure-aware pre-training","triplet contrastive learning","TM-score","3D distance encoding","amino acid substructure encoding","protein-protein interaction prediction"],"falsifier":"Compute the triplet loss exactly as written in Eq. (3) for a protein pair of different lengths, e.g. 100 and 200 residues, and observe that the subtraction is undefined; then re-run pre-training with a specified pooling (mean pooling, a learned pooler, or a CLS token) and check whether the reported contact and PPI gains persist. If the gains disappear for some pooling choices, the claimed advantage depends on the hidden alignment rather than on structure similarity.","tokens_in":20139,"feed_emoji":"🧬","tokens_out":11962,"duration_ms":103004,"temperature":0.7,"pith_summary":"GLProtein sets out to show that protein language models improve when pre-training uses structure at two scales: the global similarity between whole proteins and the local geometry of amino acids. It combines masked amino-acid reconstruction with a triplet contrastive loss that pulls a protein's representation toward structurally similar proteins and pushes it away from structurally dissimilar ones, using TM-score to define similarity. Locally, its decoder attends to molecular substructure embeddings of amino acids and injects 3D distance information as attention bias. The paper reports that this combined objective outperforms prior sequence-only and structure-aware models on amino acid contact prediction and protein-protein interaction prediction after fine-tuning. If the claim holds, structure-aware pre-training can transfer functional knowledge that sequence alone hides, such as similarity between proteins with no detectable sequence identity.","feed_headline":"Global plus local structure pre-training beats prior protein models","feed_subtitle":"Combining TM-score triplet similarity with amino-acid 3D encodings improves contact and interaction prediction after fine-tuning.","key_machinery":"The load-bearing mechanism is the joint objective $L = L_{MLM} + \\alpha L_{PTL}$, coupling three components: masked protein language modelling, global triplet structure modelling, and local decoder injection. The triplet loss $L_{PTL}$ is the carrier of global information: positives and negatives are chosen by TM-Vec using TM-score, a standard measure of whole-structure similarity, and the loss aligns representations of structurally similar proteins while repelling dissimilar ones. Local information enters through the decoder, where Gaussian-basis 3D distance encoding $\\Phi^{distance}$ biases attention and mol2vec-generated amino acid substructure embeddings $E_a$ act as keys and values. Ablations in the paper show that removing either the triplet component, the distance encoding, or the molecule embeddings lowers contact and interaction performance, which identifies these three pieces as the machinery carrying the claimed gains.","core_discovery":"The central claim is that protein structure information should be encoded at both global and local scales during pre-training, and that a model trained this way transfers better to downstream prediction than sequence-only or single-scale structure models. Globally, the framework samples a triplet $(P, P^{pos}, P^{neg})$ using TM-Vec's predicted structure similarity, then optimizes a triplet loss $$L_{PTL}(P,$P^{{pos}}$,$P^{{neg}}$) = \\max\\big(\\|E_p - $E_p^{{pos}}$\\|_2 - \\|E_p - $E_p^{{neg}}$\\|_2 + \\epsilon,\\,0\\big)$$ so that a protein's representation approaches its structural neighbors and moves away from decoys with TM-score below $0.2$. Locally, the decoder computes Gaussian-basis 3D distance encoding $\\Phi^{distance}$ from $\\alpha$-carbon coordinates and uses it as attention bias, while mol2vec substructure embeddings $E_a$ serve as keys and values, so masked amino-acid reconstruction is conditioned on atomic and geometric detail. The full objective is $L = L_{MLM} + \\alpha L_{PTL}$, and on contact prediction and protein-protein interaction benchmarks the paper reports the best results among the compared baselines.","pith_inferences":["Our inference: if the global triplet signal drives the gains, scaling the structure database used for sampling should improve remote-homology and fold-classification performance even with the encoder held fixed; this is directly testable.","Our inference: because the triplet loss as written subtracts whole-sequence representations of different lengths, the empirical implementation must rely on an unstated pooling or alignment; making that choice explicit and varying it would reveal how much of the gain comes from structural similarity versus the length reduction itself.","Our inference: the same global-plus-local recipe could transfer to RNA structures, protein-ligand complexes, or protein-nucleic acid interfaces, where whole-molecule similarity and local chemical detail play analogous roles.","Our inference: the paper's own limitation on binding-affinity prediction suggests that mutation or MSA-style variation, not additional geometric structure, is the missing signal for that task."],"forward_implications":["Pre-training on global structure similarity should let sequence-only fine-tuning inherit functional relationships that sequence identity misses, such as the similar beta-helical fold of FfIBP and CaTrailin_4.","Contact prediction and protein-protein interaction models initialized from GLProtein can be fine-tuned without structural input and still beat structure-aware baselines on the reported benchmarks.","The ablations imply that each of the three components — triplet similarity, 3D distance encoding, and amino-acid substructure encoding — contributes an independent performance gain, so strengthening any one of them should yield further improvement.","The same representation also transfers to stability prediction, where GLProtein reports the highest Spearman correlation among compared models on the TAPE benchmark, pointing to use in protein engineering."],"supporting_citations":[{"why":"It supplies the ProtBert BERT-style encoder that GLProtein initializes and continues pre-training from.","marker":"(Elnaggar et al., 2021)"},{"why":"It provides TM-Vec, which selects structurally similar positive proteins and confirms dissimilar negatives for the triplet loss.","marker":"(Hamamsy et al., 2023)"},{"why":"It defines the TM-score formula used to quantify global structure similarity between proteins.","marker":"(Zhang and Skolnick, 2004)"},{"why":"It supplies the TM-score threshold below which proteins are treated as structurally dissimilar negatives.","marker":"(Xu and Zhang, 2010)"},{"why":"It supplies mol2vec, the substructure-based molecular embeddings used to encode amino acid molecules.","marker":"(Jaeger et al., 2018)"},{"why":"It is the source of the Gaussian-basis 3D distance encoding used as attention bias in the protein decoder.","marker":"(Ying et al., 2021)"},{"why":"SaProt is the structure-aware protein language model baseline that GLProtein compares against in contact and PPI experiments.","marker":"(Su et al., 2023)"},{"why":"TAPE provides the contact prediction, stability, fluorescence, and secondary structure benchmarks used for transfer evaluation.","marker":"(Rao et al., 2019)"},{"why":"ProteinNet supplies the contact prediction test set and the TBM-hard example used in the visualization.","marker":"(AlQuraishi, 2019)"},{"why":"It provides the Swiss-Prot sequence database used as the pre-training corpus.","marker":"(Boeckmann et al., 2003)"}],"fun_headline_variants":["Global and local protein structure pre-training boosts prediction","First protein pre-training with global and local structure cues","Protein pre-training with triplet similarity and atom-level encodings","New protein pre-training combines global topology and local chemistry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The contrastive loss subtracts the representation of one whole protein from another as though they were equal-sized vectors, but protein sequences differ in length and the paper never states how they are pooled to a common size; the global structural signal therefore rests on an unspecified length-reduction step.","fun_headline_variants_meta":{"raw":{"variants":["Global and local protein structure pre-training boosts prediction","First protein pre-training with global and local structure cues","Protein pre-training with triplet similarity and atom-level encodings","New protein pre-training combines global topology and local chemistry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001124,"raw_usage":{"total_tokens":4676,"prompt_tokens":945,"completion_tokens":3731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":3676}},"tokens_in":561,"tokens_out":3731,"duration_ms":23947,"temperature":1.0,"reasoning_tokens":3676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:41:57.952365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the triplet loss exactly as written in Eq. (3) for a protein pair of different lengths, e.g. 100 and 200 residues, and observe that the subtraction is undefined; then re-run pre-training with a specified pooling (mean pooling, a learned pooler, or a CLS token) and check whether the reported contact and PPI gains persist. If the gains disappear for some pooling choices, the claimed advantage depends on the hidden alignment rather than on structure similarity.","supporting_citations":[{"cited_title":"Prottrans: Toward understanding the language of life through self-supervised learning","cited_arxiv_id":null,"evidence_quote":"It supplies the ProtBert BERT-style encoder that GLProtein initializes and continues pre-training from."},{"cited_title":"Protein remote homology detection and structural alignment using deep learning","cited_arxiv_id":null,"evidence_quote":"It provides TM-Vec, which selects structurally similar positive proteins and confirms dissimilar negatives for the triplet loss."},{"cited_title":"How significant is a protein structure similarity with tm-score= 0.5? Bioinformatics , 26(7):889--895, 2010","cited_arxiv_id":null,"evidence_quote":"It supplies the TM-score threshold below which proteins are treated as structurally dissimilar negatives."},{"cited_title":"Mol2vec: unsupervised machine learning approach with chemical intuition","cited_arxiv_id":null,"evidence_quote":"It supplies mol2vec, the substructure-based molecular embeddings used to encode amino acid molecules."},{"cited_title":"Evaluating protein transfer learning with tape","cited_arxiv_id":null,"evidence_quote":"TAPE provides the contact prediction, stability, fluorescence, and secondary structure benchmarks used for transfer evaluation."}],"review_version":1}