acceptodds
Under review as a conference paper at ICLR 2027

Episodic Relational Context Learning for Protein-Protein Interaction Prediction

Abstract

Protein-protein interaction (PPI) prediction is essential for understanding cellular processes and disease mechanisms. However, existing learning-based methods struggle in the Neither-Seen (NS) setting, where neither protein in a test pair appears in the training data. To understand this limitation, we analyze the failure modes of existing baselines under the NS setting. Our analysis shows that, beyond the well-recognized effect of sequence similarity, the absence of PPI network topology for unseen proteins constitutes another key source of performance degradation. In this paper, our core idea is to induce relational context for unseen proteins by linking them to the training PPI network rather than leaving them isolated. To this end, we propose Episodic Relational Context Learning (ERCL) to improve PPI prediction generalization, particularly under the NS setting. For each query pair, ERCL connects the retrieved anchors from the PPI network to the query proteins via virtual edges, forming a relational context graph. This graph is fed into a relation-aware graph encoder and classified at the graph level, with the resulting label serving as the PPI prediction. To learn relational context from isolation, we construct each training episode by masking the observed interactions of query proteins, turning seen proteins into pseudo-unseen ones. On the cross-species benchmark with exclusively NS testing pairs, our method achieves state-of-the-art performance on all five testing species, with particularly pronounced gains on species more phylogenetically distant from the human training data. Moreover, our model achieves these gains with 585 fewer parameters than strong baselines such as PLM-interact, while being up to 29 faster in both training and inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein-protein interaction prediction, graph neural networks, protein representation learning, inductive learning, relational cold start, computational biology

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CgDtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a new form of protein-protein interaction prediction that fuses sequence-based information (for each query protein) with network information (the pattern of known PPIs connecting the two query proteins). Specifically, it tackles the case where neither of the query proteins appears in the training PPI network (i.e., the "neither seen/NS" or "cold start" problem that many others have investigated). In that NS case, network-based features are unavailable for the test pair. Instead, they use sequence similarity (MMseqs2 matching score) to identify "anchor" (~homologous) proteins within the training PPI network, create virtual links from the test proteins to those anchor proteins and also to some internal nodes along short PPI paths connecting the anchor proteins. A "relation aware graph encoder" then encodes the proteins/nodes and the subgraph connecting the query, anchor, and internal nodes.

Experiments demonstrate increases in accuracy (AUPR) and throughput/efficiency on 5 non-human test PPI datasets (D-SCRIPT) when human data are used for training. Several methods are included as SOTA baselines, including three network-based methods that are modified to account for the NS case. Results are convincing.

Critical strengths

The authors spend considerable time establishing the case for adding surrogate network topology information in the NS use case, which motivates their method.

The paper is clearly written and easy to follow. The method is explained well, but some details feel unspoken (see below).

The comparison with the SOTA is convincing and the selection of baseline methods appears to cover the SOTA.

Critical weaknesses

The authors equate cross-species PPI prediction (where the training and test data arise from different species) with the neither-seen (NS) use case. I would argue that this is not representative of interesting NS cases... Since so many protein complexes and key protein pathways are preserved between species, we expect good matches for many non-human (test) protein pairs within the human (training) PPI network (e.g., interlogs). So, we can expect that many of the NS non-human test pairs are likely represented by direct equivalent pairs in the human training data. To me, the more interesting case is where there is no direct homolog available for the test sequences (e.g., engineering a novel peptide in virgin sequence space). Cross-species PPI prediction is a well-established use case and does not feel equivalent to the NS use case...

Several hyperparameters are required in the proposed method. For example, Section 5.2.2 includes "L_path" and "B", while Section 5.2.3 mentions adding virtual edges between the query and internal nodes when their matching score exceeds "a fixed threshold" (an unnamed hyperparameter?). How are these chosen/optimized? Is a sensitivity analysis completed for these hyperparameters?

In Section 4.2, the authors show that train-test sequence similarity does not explain PPI prediction accuracy, whereas others (Bernett 2024 and others) have shown the opposite: that sequence similarity between training and testing proteins explains much of the performance of PPI predictors. How do the authors explain the difference in findings here?

Similarly, Bernett and others have shown that node degree explains much of the remainder of PPI prediction accuracy (where a node with high degree in the training data would be expected to also participate in test PPIs, when compared with a training protein of low degree). Using the authors' approach, wouldn't a hub protein in a non-human test set, when found to be sequence-similar to a human hub protein in the training set, thus be predicted to interact with many test proteins? Isn't this a form of short-cut learning and data leakage?

Line 340 says "with a relation-aware graph encoder". What is its structure? Is it trained from scratch on the human training data? Is it entirely frozen during non-human tests?

Minor: the tone of the early sections of the paper makes it feel like the authors are discovering/demonstrating the challenge of the NS use case for the first time (e.g., Introduction and Section 4.1). The challenge associated with the NS use case was described at least as early as 2012 by Park & Marcotte (in which they defined the C1/C2/C3 test classes, where C3 pairs share neither protein with the training set, so no known interactions of either test protein can be used.) More recently, in the "protein disjoint" Gold PPI dataset (Bernett 2024) test proteins are excluded from the training dataset to avoid such data leakage.

Questions for authors

I encourage the authors to respond to the Weaknesses listed above.

8 comments
AuthorsOpenReview ↗

Response to the minor comment: NS definition and prior work

We thank the reviewer for pointing out prior work on the NS setting. At the level of train/test partitioning, NS means that neither protein in a test pair appears in any training pair. We agree that this criterion is not new: it corresponds to the C3 class of Park and Marcotte (2012) and is also reflected in the protein-disjoint evaluation of Bernett et al. (2024). We will tighten wording that might suggest we were the first to identify this split or its associated challenge, and acknowledge these studies more explicitly.

While we adopt this established split-level definition of NS, the setting studied here also specifies what PPI information is available at inference: neither query protein has observed PPI edges in the reference network accessible to the model. A protein-disjoint training split alone does not guarantee this, because a separate reference network may still contain test proteins and their known neighbors. For example, the original GNN-PPI reports an NS test subset but predicts the types of known interactions; in its original graph-based evaluation, test-pair edges can enter the graph structure while their interaction-type labels remain withheld. LLaPA removes test edges during training and inference, but uses a reference PPI network assembled from multiple sources. A test protein absent from the training pairs may therefore still have other known PPI neighbors in that network. Thus, methods using an NS split can differ in the network information available at inference.

In our cross-species evaluation, the only observed PPI edges available to the predictor come from the human training network. Target-species proteins are absent from that reference network. At inference, the model is given neither observed PPI edges incident to either query protein nor any positive or negative pair labels involving them; test-pair labels are used to compute evaluation metrics, not as model inputs. Their sequences are used to retrieve anchors in the human network, while their precomputed sequence representations serve as node features. The query–anchor links are virtual similarity edges rather than observed PPIs. ERCL builds relational context from PPI relationships within the human training network. Our contribution is therefore not a new NS split, but a way to construct relational context under this explicit information constraint.

AuthorsOpenReview ↗

Response: Cross-species NS evaluation and direct human PPI templates

We thank the reviewer for raising this question. Our cross-species evaluation satisfies the NS split: neither protein in a test pair appears in the training set. At inference, PPI information is restricted to the human training network; no known interactions, non-interaction labels, or network neighbors of the test-species proteins are provided to the model. NS does not, however, imply an absence of sequence similarity. We therefore asked two questions: does ERCL retain an advantage on weakly matched pairs, and does it retain an advantage on pairs without detected direct human PPI templates?

Sequence-matching strata. We compared test proteins against all 15,816 human training proteins using CD-HIT-2D, requiring ≥40% identity and ≥80% coverage of both sequences. Pairs with no qualifying hit for either endpoint account for 94.8%, 76.3%, 56.0%, and 73.1% of the E. coli, yeast, fly, and worm tests, respectively. These test sets thus contain substantial subsets without qualifying matches, rather than only highly matched pairs. CD-HIT is used only for this stratification; “no qualifying hit” does not mean that ERCL has no anchor.

Within each stratum, ERCL and our internal sequence-only control are evaluated on identical test pairs. Their AUPRs below are three-seed means; differences are computed before rounding.

SpeciesQualifying hitPairs (positives)ERCLSequence-onlyΔAUPR
E. coliNeither endpoint20,858 (1,739)0.7450.736+0.009
E. coli≥1 endpoint1,142 (261)0.8940.851+0.042
MouseNeither endpoint1,139 (8)———
Mouse≥1 endpoint53,861 (4,992)0.9310.887+0.044
YeastNeither endpoint41,969 (2,285)0.6200.594+0.026
Yeast≥1 endpoint13,031 (2,715)0.8330.806+0.028
FlyNeither endpoint30,777 (946)0.8700.767+0.103
Fly≥1 endpoint24,223 (4,054)0.9560.916+0.041
WormNeither endpoint40,178 (1,321)0.7680.680+0.088
Worm≥1 endpoint14,822 (3,679)0.9590.933+0.026

Note: — denotes fewer than 100 positives; AUPR is not reported. “No qualifying hit” means no match detected under the stated identity and coverage criteria, not an absence of sequence similarity or homology.

ERCL outperforms sequence-only in all four evaluable no-hit groups. In fly and worm, the gains are +0.103 and +0.088, compared with +0.041 and +0.026 in their respective groups with at least one qualifying hit.

Direct human PPI templates. To address the reviewer’s concern about interolog-based transfer, we used two complementary checks for direct human PPI templates: (1) a training PPI edge between the two sets of selected Top-5 anchors; and (2) a training PPI edge connecting a human protein matched to one query with a human protein matched to the other, considering all qualifying matches rather than only the selected anchors. Across species, 90.2%–94.8% of test pairs have no direct PPI edge detected by either check. We evaluated all three models on these same pairs:

SpeciesPairs (positives)Share of species’ positivesERCLSequence-onlyPLM-Interact
E. coli20,862 (1,422)71.1%0.6490.6600.630
Mouse49,924 (1,027)20.5%0.6330.5670.572
Yeast50,849 (2,609)52.2%0.4740.4670.427
Fly49,628 (1,012)20.2%0.7080.6130.649
Worm49,946 (1,374)27.5%0.6340.5920.580

Note: The second check requires ≥40% sequence identity and ≥80% coverage of both query and target. PLM-Interact uses its existing frozen per-sample predictions.

On these pairs, ERCL outperforms PLM-Interact in all five species and sequence-only in four. ERCL’s predictive advantage is therefore not confined to pairs with a detected direct human PPI template.

As supporting evidence, we counted three-hop PPI paths between the two anchor sets in the context graphs of these same pairs: using path count alone to distinguish positive from negative pairs yields ROC-AUCs of 0.596, 0.862, 0.692, 0.886, and 0.810 for E. coli, mouse, yeast, fly, and worm, respectively. These results show that multi-hop PPI connectivity between the two anchor sets is informative for interaction prediction, even when no direct human PPI template is detected.

Together, these analyses show that ERCL’s predictive gains extend to weakly matched pairs and to pairs without detected direct human PPI templates.

Our cross-species benchmark evaluates one NS setting; de novo designed peptides without homologous counterparts were not directly evaluated in this study.

AuthorsOpenReview ↗

Response: Sequence proximity and the findings of Bernett et al

We thank the reviewer for highlighting this point. We agree that sequence similarity between training and test proteins affects PPI prediction performance. Section 4.2 does not reject this relationship; it explicitly notes that sequence proximity broadly aligns with AUPR across the five test sets. Our narrower observation is that the measured test-set-level sequence proximity does not fully account for the differences in PLM-Interact’s reported performance across these species.

Bernett et al. changed the training/test partitions to examine how existing models perform under stricter generalization conditions. They partitioned each of the human and yeast proteomes into two non-overlapping subsets while minimizing overall sequence similarity between the subsets. When trained on PPIs within one subset and tested on PPIs within the other, the tested models performed at or near chance level. Their findings highlight the importance of protein overlap, sequence similarity, and node-degree information that models can exploit under conventional evaluation settings. We agree with these findings. Bernett et al., 2024

Section 4.2 asks a different, concrete question: with the human training set and five non-human NS test sets fixed, are the species whose proteins more closely resemble the training proteins also those with higher prediction performance? We compare each test set’s sequence proximity, summarized by the median protein-level score, with PLM-Interact’s reported AUPR. For example, worm and yeast have similar sequence-proximity values, but their AUPRs are 0.888 and 0.706, respectively. Mouse has substantially higher sequence proximity than fly, yet its AUPR is not higher: 0.904 versus 0.913. Thus, across these five test sets, the relative sequence-proximity values alone do not adequately explain why prediction performance differs between species.

These observations are compatible with Bernett et al.’s findings: sequence similarity can strongly affect performance without a single test-set-level proximity measure fully explaining cross-species performance differences. This motivated us to investigate whether PPI relational context provides complementary predictive information, with its contribution examined through the subsequent topology diagnostics and ablation experiments. Our aim is to extend the understanding of NS prediction difficulties, not to discount the importance of sequence similarity.

Reviewer_MgTKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a topology-based method to tackle the unseen PPI challenge, achieving state-of-the-art performance on all five testing species

Critical strengths

The proposed method is simple yet effective, and the writing is pretty clear.

Critical weaknesses
  1. The method appears heavily dependent on the query having existing relationships with the training graph and fails to handle genuinely semantic-level unseen cases. It would be better for the authors to provide experimental evaluations when query proteins bear low to the training graph.

  2. If query proteins must share similarity with the training graph, this essentially reduces the setting to semantic-level IID (rather than true unseen generalization). Under that case, I do not think protein-centered methods perform poorly under because the deep learning is good at treating IID. Thus, why does the proposed method suppress the protein-centered method? The answer may be that the proposed complex context graph being an inherently better input representation for learning than two isolated endpoints, rather than from claimed unseen generalization. To disentangle this, the authors should stratify test set samples by their similarity to the training graph (e.g., distance between queries and their nearest training neighbors) and report the corresponding performance gains across different similarity bins.

  3. Does the claimed 29× speedup include the context graph construction time? If not, it would be better for the authors to provide the full end-to-end latency.

  4. Please clarify the exact calculation behind the 585× parameter reduction. Given that the method strictly relies on an external reference graph at test time, completely excluding this graph's footprint constitutes an overclaim.

Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

Response to Q1: Performance under weak sequence matching

We thank the reviewer for the thoughtful and constructive assessment of our work. We appreciate the questions about generalization under weak sequence similarity, the role of relational context, and computational efficiency, and address them point by point in separate comments.

Performance under weak sequence matching. We tested whether ERCL’s advantage depends on test proteins having close matches among the 15,816 human training proteins. Using CD-HIT-2D, we required at least 40% reported identity and 80% coverage of both sequences. We assigned each pair to an identity band using the higher reported identity of its two endpoints’ selected hits; pairs with no qualifying hit for either endpoint formed the “no-hit” group. This group accounts for 94.8% of E. coli, 76.3% of yeast, 73.1% of worm, and 56.0% of fly test pairs. Thus, these test sets contain substantial weak-match subsets; their overall results are not based solely on highly matched pairs.

The following AUPRs are evaluated on the same test pairs within each stratum. ERCL values are three-seed means; PLM-Interact uses its frozen per-sample predictions.

SpeciesStratumPairs (positives)ERCLPLM-InteractΔAUPR
E. coliNo qualifying hit20,858 (1,739)0.7450.715+0.030
YeastNo qualifying hit41,969 (2,285)0.6200.575+0.046
WormNo qualifying hit40,178 (1,321)0.7680.699+0.069
FlyNo qualifying hit30,777 (946)0.8700.829+0.041
Mouse60–<80% identity6,726 (280)0.9450.900+0.045

All three ERCL seeds outperformed PLM-Interact in each of the four no-hit groups. Mouse had only 8 and 28 positives in its no-hit and 40–<60% groups, below the 100-positive reporting threshold applied to all strata; we therefore report its first evaluable lower-identity group. The observed advantage was not merely preserved under weaker matching: compared with the ≥80% group, the gap was larger in the no-hit groups of yeast (+0.046 vs. +0.005), fly (+0.041 vs. +0.015), and worm (+0.069 vs. +0.008), and in the 60–<80% group of mouse (+0.045 vs. +0.024). E. coli also gained +0.074 in its 40–<60% group, although its ≥80% group was not evaluable.

We further stratified pairs by the anchors ERCL actually uses. For pairs with usable LCC anchors for both proteins, we took the higher of their actual Top-1 anchor scores, normalized as MMseqs2 bits/self-bits, and formed lower-half, middle-quartile, and upper-quartile groups within each species. The table reports ERCL minus PLM-Interact ΔAUPR in these groups:

SpeciesLower halfMiddle quartileUpper quartile
E. coli+0.052+0.012+0.031
Mouse+0.033+0.039+0.018
Yeast+0.049+0.057+0.032
Fly+0.051+0.037+0.023
Worm+0.143+0.059+0.016

All 15 mean differences are positive. In the lower-half groups, all three ERCL seeds outperformed PLM-Interact in every species, and the observed gap exceeded that of the corresponding upper-quartile group in all five species. Because positive rates differ between bins, we interpret the between-bin pattern descriptively; the within-bin comparisons use identical samples. Together, both stratifications show that ERCL’s advantage over PLM-Interact persists, and is often more pronounced, among relatively weakly matched pairs rather than being confined to high-scoring training-network anchors. “No qualifying hit” refers to the stated CD-HIT search criterion, not to an absence of usable anchors.

AuthorsOpenReview ↗

Response to Q2: Sequence similarity and the contribution of PPI context

We address two related explanations: whether ERCL’s advantage is confined to queries that closely resemble training proteins, and whether it arises simply because the context graph contains more protein nodes. In our NS setting, neither query protein appears in the training PPI network, but this does not imply an absence of sequence similarity to training proteins. We therefore stratified the NS test pairs by similarity to their selected anchors. As detailed in our response to Q1, ERCL’s mean AUPR exceeded that of the protein-centered PLM-Interact baseline in the low-, middle-, and high-similarity strata of all five species. The advantage also persisted in the CD-HIT no-hit groups of the four species with sufficient positive samples. Thus, the gain is not confined to queries that closely resemble training proteins.

We agree that a richer relational representation may help ERCL. To distinguish the contribution of PPI relationships from simply providing more protein nodes, we conducted a PPI-edge ablation. Both arms contain the same query proteins, anchors, and intermediate nodes, and retain the same sequence-similarity edges, edge weights, and other input features. The full-graph arm uses the existing ERCL checkpoints. For the ablated arm, we trained the same architecture with matched seeds and training protocol, removing PPI edges from both its training and test graphs. Mean test AUPR over three seeds was:

Test setERCL, full graphPPI edges removedDifference (computed from unrounded values)
E. coli0.7630.753+0.011
Mouse0.9300.916+0.014
Yeast0.7400.723+0.017
Fly0.9410.916+0.025
Worm0.9140.897+0.017

Full-graph values follow the submitted manuscript; differences are calculated from full-precision results. The full graph outperformed the edge-removed graph for every species in all three seeds. This benefit also persisted among test pairs for which neither query had a detected match to a training protein at ≥40% sequence identity and ≥80% coverage of both aligned sequences: the AUPR gains were +0.012, +0.024, +0.043, and +0.034 for E. coli, yeast, fly, and worm, respectively, again with the full graph ahead in all three seeds.

Together, the similarity-stratified comparison and fixed-node ablation show that ERCL’s advantage is not restricted to highly similar queries, and that PPI relationships in its context graph provide predictive information beyond the inclusion of additional protein nodes.

AuthorsOpenReview ↗

Response to Q3–Q4: Latency and resource accounting

Online latency including context-graph construction

The reported 29.3× speedup measures model forward passes with the context graphs and model inputs already prepared; it does not include graph construction. To quantify that cost, we conducted a separate benchmark that constructs a graph for each query pair and measures the time through to its prediction score.

We tested the methods sequentially on the same server and single GPU, with batch size 1, using 1,000 pairs from each of the five test species and three timing repeats over the same pairs. ERCL’s measured time includes anchor selection, PPI path search between anchors, context graph assembly, feature preparation, device transfer, and prediction. PLM-Interact’s includes tokenization, device transfer, and prediction. Under the paper’s default settings of K=5K=5, Lpath=4L_{\mathrm{path}}=4, and B=200B=200, the mean online latencies across species are 73.06 ms/pair for ERCL and 88.72 ms/pair for PLM-Interact, corresponding to a 1.21× speedup.

Because the training PPI network is fixed across queries, we also tested an index that precomputes shortest-path distances among its existing nodes up to the allowed hop limit. During online path search, these distances rule out candidate nodes that cannot reach the target anchor within the remaining hops, avoiding repeated distance calculations. The index stores no test proteins or query-specific anchor sets, paths, or intermediate-node lists; path enumeration, query-specific scoring, and graph assembly remain online. With this index, ERCL’s mean online latency is 60.77 ms/pair, corresponding to a 1.46× speedup over PLM-Interact.

Thus, the original 29.3× figure describes forward-pass speed, whereas the new 1.21× and 1.46× figures describe online latency per test pair including graph construction. ERCL uses precomputed ESM features and MMseqs2 search results in the new benchmark; those offline steps and the one-time index construction are excluded from per-pair online latency.

Scope of the 585× parameter comparison and reference-graph footprint

The approximately 585× figure compares the parameter counts of the two prediction-stage models, not the resources required by their complete workflows. We counted the elements in each predictor’s parameter tensors and calculated the ratio from the unrounded counts; displayed to two decimal places, PLM-Interact has 651.05M parameters and ERCL has 1.11M.

The external reference network and its associated inputs are not model parameters, but they do incur storage costs. The human PPI adjacency occupies 0.30 MiB on disk; the human similarity lookup and precomputed protein features occupy 20.39 MiB and 77.23 MiB, respectively. These reference-side files total approximately 97.93 MiB and are shared across all five test species. For E. coli, the target-specific lookup and features add 43.68 MiB, and the ERCL checkpoint occupies 12.79 MiB, bringing the selected cached-input and model files to 154.39 MiB. Separately, PLM-Interact’s fine-tuned checkpoint and configuration/tokenizer files occupy 2488.78 MiB. These file sizes are not measurements of total system memory and are not used to derive the parameter ratio.

ERCL’s protein features are generated offline using a frozen ESM-2 650M encoder. This encoder is not loaded during prediction from precomputed features and is therefore excluded from the prediction-stage parameter count. If the full workflow begins with raw sequences, the encoder must be included as an additional model component; the 585× figure does not describe a full-pipeline parameter reduction.

AuthorsOpenReview ↗

Follow-up on our responses to your concerns

Dear Reviewer MgTK,

Thank you for your careful review. We have posted responses addressing all four of your points, including additional evaluations under weak sequence matching, a PPI-edge ablation, online latency including context-graph construction, and clarification of the parameter-count comparison and reference-side storage costs.

If you have an opportunity, we would appreciate your taking a look at these responses. We would be glad to address any remaining questions.

Best regards,

The authors

Reviewer_aE92OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors developed ERCL, a method for predicting PPI when neither protein appears in the training data (NS situation). They first analyze and demonstrate why this setting is hard, showing that sequence similarity to the training proteins does not fully explain performance across species, and that a topology-only model with privileged access to the test network performs strongly without any sequence information. To this end, they build a context for each query pair: homologous anchors are retrieved from the training network, interaction paths between anchors are collected, the query proteins are attached to their anchors by virtual edges, and the resulting graph is classified. During training, each pair's interactions are masked so that seen proteins behave as if they were unseen. ERCL is evaluated on five non-human test sets and improves on the reported baselines across all of them.

Critical strengths

The main strength, to my understanding, is that the method follows from a diagnosis rather than being assumed. They establish that the difficulty is partly due to a lack of relational context rather than solely to a sequence generalization problem, and the topology-only analysis is carefully set up, with the queried edge removed during both training and evaluation, and the result explicitly described as an oracle rather than a usable predictor. Substituting a homology-derived context for a protein's missing interaction neighborhood makes biological sense. The efficiency comparison is clearly reported.

Critical weaknesses

My background is in computational biology, but I have limited experience in the PPI field. Therefore, I have concentrated on the evaluation design and the biological reasoning. I have two questions: one is technical, and the other is about how the results should be interpreted. Neither of which affects my positive assessment.

Questions for authors
  1. MMseqs2 is used in three places (the proximity measure in Section 4.2, the anchor retrieval in Section 5.2.1, and the virtual-edge weights in Section 5.2.3). Appendix A.1 states that the self-alignment check uses the default search, but the anchor retrieval settings are not specified, and no version is reported (defaults can differ between releases). Could the authors state the parameters and version used throughout? Since the retrieved anchors determine the context graph, some indication of how sensitive the results are to these settings would also be valuable, even a single comparison at a higher sensitivity setting for one of the more distant species.

  2. Anchors are selected by raw sequence-matching score, while Section 2 cites an orthologue-informed method among the related work. Did the authors consider using curated orthology to select anchors instead, and would they expect this to change the results for more distant species such as yeast and E. coli?

2 comments
AuthorsOpenReview ↗

Response to Q1: MMseqs2 settings and anchor-retrieval sensitivity

We thank the reviewer for requesting the MMseqs2 settings and a sensitivity check. The three uses in the paper are based on MMseqs2 alignment bit scores, with different downstream processing. Section 4.2 uses each test protein’s best match among the 15,816 human training proteins to calculate sequence proximity: the matching bit score is divided by that test protein’s self-alignment bit score and weighted by query coverage. As stated in Appendix A.1, the self-alignment is obtained through a separate default-parameter search; it is used for normalization, not anchor selection. Section 5.2.1 selects anchors from the 7,097 proteins in the largest connected component of the human training PPI network. Section 5.2.3 uses the corresponding precomputed bit scores as virtual-edge weights, without running MMseqs2 during graph construction.

For the cross-protein matching searches, we used MMseqs2 version 18.8cc5c with easy-search, -s 7.5, -e 1000, --max-seqs 1000, and --threads 32; we did not impose an additional minimum sequence-identity or coverage threshold. Matches were ranked by bit score, and up to 100 candidates were cached per query protein. Graph construction then selected up to five anchors per endpoint (K = 5). With the cache and selection rule fixed, reconstructing the same query pair does not randomly change its anchors.

To examine retrieval sensitivity, we compared the paper’s -s 7.5 setting with a lower-sensitivity -s 5.7 setting for E. coli and yeast. These independent searches used the same query sequences, 7,097-protein reference set, MMseqs2 version, other search options, and hit-processing rules; only -s changed. For any query–target pair returned by both searches, its bit score was identical. What changed was which candidate pairs were returned.

SpeciesQueries with no hit: 7.5 → 5.7Different Top-1, excluding first-place tiesDifferent Top-5 sets, excluding rank-5/6 ties
E. coli10 → 2382,580/4,572 (56.4%)973/1,134 (85.8%)
Yeast29 → 2521,545/4,028 (38.4%)824/1,238 (66.6%)

For Top-1, the denominator includes only queries with a hit under both settings and no tie between the first and second bit scores in either search. For Top-5, it includes only queries with at least five hits under both settings and no tie at the fifth/sixth-rank boundary in either search. These conditional proportions do not represent all queries.

This comparison shows that changing search sensitivity changes the candidates available for anchor selection, even though retrieval is reproducible under a fixed setting and cache. It is a retrieval-level analysis, not a measurement of the resulting AUPR change. We plan a controlled downstream comparison for E. coli and will report it in a follow-up comment if completed during the discussion period.

AuthorsOpenReview ↗

Response to Q2: Orthology-informed anchor selection

We thank the reviewer for suggesting curated orthology as an alternative for anchor selection. We have considered this possibility, but did not use or experimentally evaluate an orthology-based selection rule in the current work. We therefore cannot claim that it would improve or reduce AUPR for yeast or E. coli.

We use MMseqs2 sequence-matching scores so that query proteins absent from the training PPI network can follow the same retrieval rule when their sequences are available. Making curated orthology a requirement would introduce an additional annotation or inference dependency. An orthologous candidate would also need to belong to the human training PPI network used for graph construction. This could reduce the number of usable anchors and the fraction of queries for which anchors can be retrieved, a concern particularly relevant to distant species such as yeast and E. coli. It would also add a prerequisite when applying the method to new proteins or species.

For queries with reliable orthology information, orthology could provide complementary evolutionary evidence and change the selected anchors and their relational context. However, ERCL uses PPI relationships between the two sets of anchors; identifying an ortholog for each query protein does not by itself establish that the resulting context is more informative for predicting the query pair’s interaction. We therefore view orthology as an optional source of evidence worth testing, rather than an assumed improvement. A controlled comparison would need to report both anchor coverage across the full test set and predictive performance on a clearly defined comparable subset.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.