acceptodds
Under review as a conference paper at ICLR 2027

PepSeek: Receptor-Conditioned Representation Learning for Peptide Recognition

Abstract

Modeling peptide recognition requires learning receptor-dependent interactions from limited affinity measurements while accounting for structural constraints that are expensive to evaluate at scale. We introduce PepSeek, a receptor-conditioned representation learning framework that couples sequence-derived structural priors with few-shot adaptation and coarse-to-fine structural evaluation. The central design principle is to separate reusable molecular representations from target-specific interaction learning, reserving detailed complex modeling for prioritized candidates. PepSeek integrates pretrained residue features, predicted contact topology, and atom-level connectivity into graph representations, and adapts cross-entity interaction layers using limited receptor-specific supervision. This enables structure-informed ranking without requiring an experimentally determined receptor structure at the initial stage. A subsequent retrieval-conditioned structural module shares receptor context across candidates, reducing redundant computation during complex evaluation. Experiments across a panel of G protein-coupled receptors show that target-specific adaptation improves within-receptor affinity ranking over the zero-shot model. In a food-derived peptide library, prioritized candidates exhibit 4.4-fold enrichment for high-confidence complex predictions among the top 50, while structural benchmarking demonstrates reduced receptor-side computation. Together, these findings support integrating receptor-conditioned learning with structural inference to translate sparse affinity measurements into structurally informed prioritization of peptide candidates. PepSeek thus provides a computational framework for translating sparse biochemical supervision into focused, structurally informed hypotheses for experimental testing.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Receptor-conditioned representation learning, Geometric deep learning, Protein language models, Peptide–receptor interactions, Structure-guided virtual screening

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QsoCOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents PepSeek, a coarse-to-fine framework for peptide–receptor recognition and candidate screening. The proposed workflow combines hierarchical molecular graph representation learning, receptor-specific adaptation, and accelerated complex structure prediction to rank and prioritize candidate peptides. The authors construct P1, a benchmark based on cold-chemical-series splitting, and evaluate the method across five activity prediction tasks and 16 GPCRs. The workflow is further assessed through structural enrichment analysis, AF-turbo runtime comparisons, and cellular functional assays.

Critical strengths

The paper is generally well organized, with a clear motivation and a practically relevant screening workflow. The authors integrate molecular activity prediction, structural evaluation, and cellular experiments, providing a relatively broad evaluation that connects computational predictions with biological validation.

The P1 benchmark adopts a chemical-series-based splitting strategy that also accounts for repeated measurements and peptide–receptor pair identities, offering a more rigorous evaluation protocol than conventional random splitting. In addition, AF-turbo substantially reduces computational time while maintaining comparable structural prediction quality. The cellular assays also identify peptides with concentration-dependent responses, demonstrating the potential practical value of the workflow.

Critical weaknesses

The methodological novelty is not sufficiently established. The core graph architecture of PepSeek closely resembles the existing PSICHIC framework, including molecular and receptor graph representations, hierarchical graph processing, and cross-entity attention. The model is also initialized from the pretrained PSICHIC checkpoint. Although the authors introduce task-specific adaptation and a new screening workflow, the manuscript does not clearly distinguish the proposed interaction mechanism from the original PSICHIC architecture. Moreover, the absence of controlled ablation studies makes it difficult to determine the independent contributions of the proposed components.

The main performance comparison has a significant limitation. The best results in Table 1 are obtained from a metric-specific oracle whose ensemble weights are fitted using the test labels, rather than from a predictor selected independently of the test set. Although the authors explicitly acknowledge that this is a diagnostic upper bound, the main comparison does not report the actual test performance of the validation-selected eight-seed ensemble. Consequently, the reported results do not sufficiently establish that PepSeek outperforms existing methods under an independent evaluation setting.

The baseline comparisons and ablation studies are insufficient. In Table S2, a simple sequence-nearest-neighbor baseline reaches a median Spearman ρ of 0.3805, which is higher than the best adapted PepSeek setting (selected initialization) at 0.3155, but this result receives limited discussion. Furthermore, the reference models largely retain their frozen pretrained weights, whereas PepSeek undergoes task-specific adaptation. This makes it difficult to distinguish improvements arising from architectural design from those attributable to additional training. The paper also lacks a direct comparison with the original PSICHIC under matched training conditions and systematic ablations of the key architectural components and training strategies.

Questions for authors
  1. Table 1 reports oracle ensembles optimized using test labels. Could the authors provide the actual RMSE and Spearman correlation of the eight-seed ensemble, selected entirely using training and validation data, on the P1 test sets, and compare these results against the existing baselines?
  2. The core architecture of PepSeek appears highly similar to PSICHIC. Could the authors clarify the specific differences between the proposed interaction mechanism and the original PSICHIC architecture? Could they also provide a controlled comparison with PSICHIC under identical training and evaluation conditions, together with ablations of the key architectural components?
  3. In Table S2 the sequence-nearest-neighbor baseline ranks better than the adapted model on both the median (0.3805 vs. 0.3155) and the mean (0.3741 vs. 0.3611) Spearman ρ. How do the authors explain this observation? Additionally, could the main baselines be evaluated under the same data splits and task-specific adaptation settings to provide a more direct assessment of the proposed method's improvements?
  4. In section 6.4, the 6.22-fold enrichment is based on AlphaFold's ipTM rather than experimentally validated binding. Could the authors quantify the correlation between PepSeek scores and ipTM, and provide independent evidence that this structural enrichment translates into improved biological hit rates?
3 comments
AuthorsOpenReview ↗

Response to Reviewer QsoC

Thank you for your review. To prioritize likely binders from large peptide libraries for a protein, we systematically refined PepSeek's data, architecture, training, and evaluation through protein multi-region and peptide multi-conformation representations.

  1. Weakness 1 — Biological motivation and novelty Protein–peptide binding depends on local protein chemistry and spatial compatibility with a flexible peptide. This motivates two coupled representations. Up to 32 soft regions pool sequence-distant residues into motifs. Each peptide residue reads these motifs, conditioning a variational generator of multiple backbones. Their internal distances guide residue communication. Shared region–residue features support contact reconstruction and partner ranking, followed by state-weighted scoring. This allows the same peptide's region reads, geometry and matching evidence to change with target. PSICHIC learns protein–ligand graph fingerprints. PepSeek matches motifs to generated peptide geometry without PSICHIC architecture or checkpoints (Figure 1, Sections 4.1–4.4).
  2. Weakness 2 — Independent evaluation We selected and froze the checkpoints by mean validation AUROC before testing, and averaged independently computed test metrics across the two seeds. Prediction weights were not fitted on test labels. All methods score the same 989 candidates per target. The top 1% contains ten peptides. PepSeek full achieves 0.91018 AUROC, 44.481-fold EF1% and 13.332-fold EF5%, best in Table 1. The same checkpoints achieve 0.845849 AUROC on 98 Test167 pairs, highest in Table S7. PepPrCLIP is reported separately because its checkpoint's training set is not fully verified (Appendix C.3). Our 499,161 source records span Propedia26, PepBDB, PPIKB, PDB, AlphaFold DB, PLINDER, CPSea, TED and GlycoShape. Records are reconciled, deduplicated and split, providing pair-ranking and structural-supervision layers (Section 5.1, Appendix A, Table S1). Training comprises region learning, backbone pretraining and joint conformation-branch optimization (Appendix C.1).
  3. Weakness 3 and Question 1 — Results and baselines Table S12 reports validation-selected RMSE and Spearman results for the five original endpoints. Our main comparisons now evaluate PepSeek on test and external benchmarks. Our PepSeek leads all listed methods in Table 1 on AUROC, EF1% and EF5%, and in Table S7 on AUROC. TRAIN-only nearest neighbors test sequence-based transfer on the same candidates. PepSeek outperforms both NN baselines on both test panels (Tables S16–S17). On the fixed 48-query panel, peptide/protein–peptide nearest neighbors achieve 0.326045/0.461800 AUROC, 0.644/11.635-fold EF1% and 0.644/3.054-fold EF5% (Table S16). PepSeek full improves AUROC over paired NN by 0.448381 (Appendix G.2). Averaged over seeds, full retrieves a net 18 more partners in the top 1% and 32 in the top 5% (Table S18). On the external 98 Test167 pairs, AUROC is 0.845849 versus 0.824781 for paired NN. For the 42/48 targets (87.5%) whose closest training protein has less than 30% full-length identity, PepSeek full achieves 0.91635 AUROC, 41.919-fold EF1% and 13.244-fold EF5%, versus 0.44675, 10.231-fold and 2.899-fold for paired NN (Table S19).
  4. Question 2 — Component evidence We examine the representations through a conformation ablation, an overall representation comparison, and held-out geometry evaluation. Our conformation ablation keeps the region scorer fixed and adding conformations raises AUROC from 0.91000 to 0.91018 and EF1% from 43.754 to 44.481. The overall representation comparison gives 0.77106 AUROC, 21.816 EF1% and 8.314 EF5% without both representations (Table 2, Figure S5, Appendix F.2). On held-out protein groups, pretraining reduces intrapeptide distance error from 11.95 to 5.37 Å and triplets containing near-duplicates from 9.52% to 3.85% (Section 6.1, Figure S3). This measures learned geometry and candidate diversity.
  5. Question 3 — Sequence transfer and generalization Thank you for highlighting sequence transfer. Table S8 compares within-receptor activity using leave-one-out NN and five-fold adaptation. Screening a shared library against new proteins makes target similarity central. Against 415,475 training proteins, per-query maximum global identity ranges from 19.22% to 68.52%, with median 25.45%. Per-query medians have median 13.13% and exact full-length overlap is zero (Appendix H, Table S20).
  6. Question 4 — Structural and functional evidence Thank you for raising biological relevance. Top-25 enrichment is 6.22-fold, 5/25 meet ipTM ≥0.60 versus 243/7,561 overall (Figure S6f). We added functional tests across six GPCRs, including MC1R, MC3R, MC4R, MC5R, GLP1R and GIPR (Section 6.3, Figures S7–S9, S11–S13). Seven of 32 evaluable melanocortin candidates show robust activity, a 21.9% observed hit rate, with five submicromolar hits. Figure 5 presents the resolved MC1R–M4-18 complex. These results link computational prioritization to receptor-dependent activity.

Please take a moment to respond the authors. Thank you!

Thank you for facilitating the discussion. We appreciate your time and coordination.

Reviewer_uVvxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PepSeek, a coarse-to-fine workflow for GPCR-targeted peptide discovery. The ranking stage represents peptides as atom–bond graphs and receptors as residue graphs with sequence and predicted-contact edges, then uses hierarchical message passing and cross-entity attention between peptide motifs and learned receptor clusters to produce a receptor-conditioned interaction score. Only terminal interaction/regression parameters are adapted per receptor task, while shared molecular encoders remain frozen. The structural stage applies AF-turbo to the shortlist, reusing receptor-side preparation context across peptides to reduce complex-prediction time while preserving interface quality.

The authors introduce a leakage-controlled cold-chemical-series benchmark, P1, covering five activity endpoints, and evaluate on 16 GPCRs, MC4R structural enrichment, 59 paired complexes for AF-turbo, and cellular assays for 32 evaluable synthesized peptides. Main reported results include improved RMSE/Spearman on P1 under a metric-specific oracle envelope, receptor-specific ranking gains for 13/16 GPCRs, 6.22-fold MC4R top-25 structural enrichment, 7.26-fold median runtime reduction with AF-turbo, and seven concentration-dependent cellular responses among 32 evaluable peptides.

Critical strengths
  1. Well-motivated and coherent end-to-end workflow. The paper addresses a real bottleneck in peptide discovery: ranking huge peptide–receptor spaces before expensive structure prediction and assays. The two-stage design—receptor-conditioned ranking followed by cached structural refinement—is practically sensible. The hierarchical peptide–receptor interaction design, especially motif–pocket cross-attention with receptor clustering, is a reasonable inductive bias beyond simple graph concatenation.

  2. Careful leakage-controlled benchmark construction and broad evaluation axes. The P1 cold-chemical-series split is a meaningful attempt to avoid near-duplicate chemotypes and repeated measurements inflating performance. The evaluation spans activity prediction, receptor-specific ranking, structural enrichment, runtime/DockQ, and cellular responses. This is more end-to-end than many purely computational peptide–receptor papers.

Critical weaknesses
  1. The headline P1 result is an oracle envelope using test labels, which weakens the central claim. Table 1’s “PepSeek (oracle)” uses metric-specific non-negative convex weights fitted on the displayed P1 test labels. The authors acknowledge this, but the abstract and results still risk conflating a diagnostic upper envelope with a deployable model. This is a significant issue because the main activity-prediction claim (“best evaluated RMSE and Spearman for every endpoint”) is based on test-label-optimized weights. The paper should clearly separate the sealed deployable eight-seed predictor from the oracle envelope and report the former as the primary result. Without this, the comparison to baselines is not fully convincing.

  2. Baseline comparisons and ablation evidence are insufficient to isolate the proposed contributions. The reference systems are frozen checkpoints trained on other datasets/endpoints, with some RMSE values unavailable due to incompatible output scales. It is unclear whether these baselines were retrained or tuned on P1 under comparable protocols. In addition, the paper lacks systematic ablations for the key architectural choices: motif–pocket cross-attention, receptor clustering (5/10/20), junction-tree pooling, receptor-contact edges, and the choice of adapting only terminal layers versus more of the network. Without these, it is hard to attribute gains to receptor-conditioned hierarchical interaction rather than adaptation or ensemble effects.

  3. Structural enrichment is evaluated only on MC4R within a shared 7,561-peptide pool, yet the paper discusses receptor-conditioned ranking across 16 GPCRs and structural prioritization more broadly. The 16-receptor ranking cohort has only 16–44 observations per training fold, and several P1 endpoints have small test partitions (e.g., pKd test = 66, functional pIC50 test = 75). The cellular assay has 32 evaluable peptides and seven responses, without detailed dose–response, subtype selectivity, or cytotoxicity data in the main text. These are promising but not yet sufficient to support strong generalized claims about peptide discovery across receptors.

Questions for authors
  1. Baseline fairness: Were any reference systems retrained or fine-tuned on P1 with comparable data access and hyperparameter selection? For GraphDTA KIBA and MGraphDTA ToxCast, native outputs are not on the same p-activity scale; how should readers interpret the Spearman-only comparisons?

  2. Can the authors report the sealed eight-seed calibrated PepSeek predictor’s RMSE and Spearman on P1 separately from the test-label-fitted oracle envelope? If the oracle uses displayed test labels, what is the main non-oracle evidence that PepSeek outperforms the strongest baseline?

4 comments
AuthorsOpenReview ↗

Response to Reviewer uVvx

We thank the reviewer for the questions on model selection, fair comparisons and experiments. PepSeek ranks peptides for each target to prioritize candidates for structural screening and functional assays.

  1. Data, training and model selection (Weakness 1 and Question 2)

Table 1 replaces the test-label-fitted oracle with fixed predictors. Our 499,161 records span curated pairs (Propedia26, PepBDB, PPIKB), experimental and predicted structures (PDB, AlphaFold DB), structural collections (PLINDER, CPSea, TED), glycosylation context (GlycoShape) and predicted antibody–antigen and TCR–peptide–MHC complexes. After deduplication and splitting, 455,994 training pairs yielded 41,958 ranking associations and 451,406 contact records in overlapping layers (Sec. 5.1, Appendix A, Table S1). For 48 test targets and 415,475 distinct training proteins, including contact inputs, 19,942,800 Needleman–Wunsch global alignments (BLOSUM62) gave no exact matches (Table S13). Identity counts matching canonical residues over all alignment columns. Identity to the closest training protein ranged from 19.22% to 68.52% (median 25.45%, Appendix F.4, Eq. 33, Tables S14, S20). On 42 targets below 30% maximum identity, PepSeek had 0.91635 AUROC and 41.919-fold EF1%, versus 0.44675 and 10.231-fold for the paired nearest neighbor (Table S19).

We learned regions from partners and contacts, then fitted the generator to observed backbones. With the region scorer fixed, ranking, contact and geometry losses trained the conformation branch and readout (Appendix C.1). Step 94 had the highest two-seed mean validation AUROC among trained snapshots. Both were fixed before testing. Table S12 in Appendix F.1 gives RMSE and Spearman correlation by P1 endpoint for the fixed eight-seed predictor.

  1. Baseline fairness and enrichment (Weakness 2 and Questions 1–2)

We excluded exact full-length matches to disclosed supervised training targets. All methods ranked the same 989 peptides per target. External models were frozen for screening and P1 (Appendix C.2, Table S5). KIBA and ToxCast scores lack p-activity units, so Spearman assesses rank agreement. Across 48 queries and 34 equally weighted groups, PepSeek led Table 1 with 0.91018 AUROC, 44.481-fold EF1% and 13.332-fold EF5%. PepPrCLIP achieved 0.76500, 18.665-fold and 5.818-fold (Fig. 2).

Appendix G adds peptide-only and protein–peptide nearest neighbors using known training pairs and the same test candidates. At the 1% cutoff, the stronger paired neighbor achieved 11.635-fold enrichment versus 44.481-fold for PepSeek (Table S16). Across the 48 queries, PepSeek recovered a net 18 additional partners on average across seeds (Table S18).

Test167 used the same exact-match rule. All methods scored the same 98 pairs across 35 targets (Appendix C.3). PepSeek averaged 0.845849 pooled AUROC across seeds versus 0.824781 for the paired neighbor (Table S17). PepSeek ranked first among the 17 entries in Table S7. All the disclosures are shown in Fig. 2 and Apps. A–C.

  1. What the two representations contribute (Weakness 2)

PepSeek learns up to 32 recognition motifs per protein. Each peptide residue reads its own combination of target motifs. The same peptide can therefore acquire a different representation when the target changes. A conditional variational model generates three candidate backbones. Within each, residue distances guide information exchange. The scorer combines all three matches (Fig. 1, Eqs. 1–6).

EF1% was 21.816-fold without either representation, 43.754-fold with protein regions, and 44.481-fold for PepSeek full (Table 2, Fig. S5). Adding conformations to the fixed region scorer raised AUROC from 0.91000 to 0.91018 and preserved EF5% at 13.332-fold. At a ten-candidate budget, seed 223 recovered 24 known partners versus 23, and seed 227 retained the same 20 (Appendix F.2). On 337 structures held out by protein group, geometry fitting reduced intrapeptide distance error from 11.95 to 5.37 Å (Fig. S3).

  1. Structural screening and functional evidence across GPCRs (Weakness 3)

Receptor-specific ranking covered 16 GPCRs. In the MC4R library, 5 of the top 25 peptides met the structural criterion versus 243 of 7,561 overall, a 6.22-fold enrichment (Fig. S6). Across 59 paired complexes, AF-turbo reduced median runtime from 23.85 to 3.283 minutes with similar mean DockQ. Figures 3–4 show the paired results and representative poses. Appendix D gives the threshold crossings.

Functional assays covered MC1R, MC3R, MC4R, MC5R, GLP1R and GIPR. Of 40 melanocortin candidates synthesized, 32 were evaluable and seven showed concentration-dependent activity (21.9%), including five with submicromolar potency. M4-18 suppressed stimulated signaling at all four melanocortin subtypes, and cryo-EM resolved its complex with MC1R (Fig. 5). The subtype screens, metabolic-receptor assays and dose–response curves show which shortlisted candidates were active and at which receptors (Figs. S7–S9, S11 and S13, Table S11).

Please take a moment to respond the authors. Thank you!

Thank you for facilitating the discussion. We appreciate your time and coordination.

Thank you to the authors for the response, which has clarified most of my questions regarding the experimental details. Therefore, I will maintain my current score. However, I noticed that the newly updated manuscript contains substantial changes compared to the original version. Given that the response was posted very close to the end of the discussion period, I unfortunately do not have sufficient time to thoroughly review the revised manuscript. Consequently, my confidence in the final assessment is not very high. Nevertheless, I appreciate the authors' efforts.

Reviewer_CvoEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

PepSeek proposes a receptor-conditioned peptide screening pipeline that combines graph-based activity prediction with structure prediction and experimental validation. It adapts a pretrained interaction model to rank peptide candidates, then uses AF-turbo to reuse receptor-side features for faster complex prediction.

Critical strengths

The proposed pipeline addresses a practical screening scenario where structure prediction and experimental validation are limited by available resources. Using receptor-conditioned ranking to prioritize candidates for these more expensive stages is a useful direction.

Critical weaknesses
  • Table 1 compares the PepSeek oracle, where the oracle ensemble weights are optimized on the same test labels used for evaluation. This is not a fair comparison as the performance of baselines are drawn from pretrained weights without an access to the test labels. Moreover, its performance does not guarantee how well PepSeek generalizes when test labels are unavailable.
  • In Table S2, sequence-nearest-neighbor baseline outperforms even the adapted PepSeek model, which does not demonstrate an advantage over a simple similarity-based approach.
Questions for authors
  • For each test peptide, what is its maximum/median sequence similarity to a training peptide? Reporting performance across similarity ranges would also help clarify the generalization of PepSeek.
  • What performance does PepSeek achieve when ensemble weights are selected using validation data only and then fixed before test evaluation?
3 comments
AuthorsOpenReview ↗

Response to Reviewer CvoE

Thank you for asking how test-label use and sequence similarity affect our conclusions. We redesigned PepSeek to identify, for a given protein target, which peptides in a large library merit structural assessment and experimental testing.

  1. Weakness 1 — Data, training and fairness of the comparison

In revised Table 1, the validation-selected PepSeek model replaces the test-fitted oracle. All methods score the same candidates with fixed predictors, and test labels are used only to compute evaluation metrics. Our dataset draws on 499,161 source records. Propedia26, PepBDB and PPIKB provide curated protein–peptide records. PDB and AlphaFold DB provide experimental and predicted structures, respectively. PLINDER, CPSea and TED contribute structural collections, and GlycoShape provides glycosylation context. Two additional predicted-complex collections cover antibody–antigen and TCR–peptide–MHC recognition.

We organized the data into two supervision layers. The structural layer teaches the model where a peptide contacts a protein and how its backbone is arranged. The pair-ranking layer uses known protein–peptide partners to teach the model which candidates to rank highly for each target (Section 5.1, Appendix A, Table S1).

We first learned protein recognition regions from known partners and contact maps, then trained the peptide geometry generator on observed backbones. Finally, we fixed the region scorer and trained the target-conditioned conformation branch and its readout jointly on ranking, contacts and geometry (Appendix C.1).

  1. Weakness 2 — What the revised screening model learns

The nearest-neighbor result in Table S8 concerns activity prediction for peptides tested on the same receptor. We redesigned PepSeek to rank a peptide library for a given target across protein families. Table 1 compares methods on that screening task, and Table 2 examines the contribution of the learned representations. On the 48-target panel, PepSeek full achieved 0.91018 AUROC versus 0.32605 for peptide-only and 0.46180 for protein–peptide nearest neighbors. It also led both in top-ten partner recovery and Test167 AUROC (Tables S16–S18).

PepSeek groups protein residues, including residues distant in sequence, into up to 32 learned regions. Each peptide residue reads from these regions, so its representation can change with the target. A conditional variational model captures peptide flexibility through three candidate backbones. Distances within each backbone guide information exchange between peptide residues. Ranking and contact reconstruction share the resulting region–residue features. The model combines evidence across conformations into one score (Figure 1, Sections 4.1–4.3).

With the protein-region scorer held fixed, adding peptide conformations raised EF1% from 43.754 to 44.481 and AUROC from 0.91000 to 0.91018. EF5% stayed at 13.332. The version without either representation reached 0.77106 AUROC under its own encoder and training protocol (Table 2, Figure S5, Appendix F.2). On held-out protein groups, geometry pretraining reduced intrapeptide distance error from 11.95 to 5.37 Å and the fraction of candidate triplets containing near-duplicates from 9.52% to 3.85% (Sec. 6.1). On the screening panel, the strongest external baseline in Table 1 reached 18.665-fold EF1%, compared with 44.481-fold for PepSeek full.

  1. Question 1 — Generalization to new protein targets

We screen a shared peptide library against new protein targets, so we assess target similarity to training proteins.

Removing five conflicting training records left zero exact full-length target matches. We globally aligned 48 queries to 415,475 distinct training proteins, counting gap columns. One 160-residue query occurs within a longer training protein. Per-query maximum identity ranged from 19.22% to 68.52%, with a median of 25.45%. The median of the 48 per-query medians was 13.13%. For the 42 targets whose closest training match had less than 30% full-length identity, PepSeek achieved 0.91635 AUROC, 41.919-fold EF1% and 13.244-fold EF5% (Tables S13–15, App. H: S19–S20).

  1. Question 2 — Performance with validation-only model selection

Among trained snapshots available at selection, step 94 had the highest two-seed mean validation AUROC. We locked both checkpoints before test scoring, evaluated each seed independently and report the arithmetic mean of their metrics. With 989 candidates per target, PepSeek full achieved 0.91018 AUROC, 44.481-fold EF1% and 13.332-fold EF5%, the best values among methods listed in Table 1. On the 98 shared Test167 pairs retained after excluding exact matches to disclosed training targets, the same checkpoints achieved 0.845849 AUROC, the highest among models listed in Table S7. Table S12 reports test performance on the original activity tasks with ensemble weights selected on validation data and fixed before testing.

We have resubmitted the manuscript with revisions highlighted and the anonymized repository ZIP.

Please take a moment to respond the authors. Thank you!

Thank you for facilitating the discussion. We appreciate your time and coordination.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.