acceptodds
Under review as a conference paper at ICLR 2027

From Pairs to Networks: Motif Graph Prompt Tuning for Inductive Protein–Protein Interaction Prediction

Abstract

Predicting protein–protein interactions (PPIs) is essential for reconstructing the networks that organize cellular function. Predictors trained on individual pairs can score pairs accurately yet assemble networks with distorted topology, and supplying network context is hardest when both queried proteins are unseen and have no recorded interactions. Our key idea is that topology, although unavailable for the queried proteins, is fully observed during training, so it can serve as supervision rather than as input, at two levels. We propose GraSPPI (GRAph Structural Prompting for PPI), a motif graph prompt tuning method inspired by prompt tuning in language, vision-language, and graph models. At the pair level, it infers from the two sequences a virtual motif graph of the pair's triadic closure, the shared partners that interacting proteins tend to have; this graph is learned against the true motif graphs of the training network and read as a prompt for the interaction decision. At the network level, structural losses regularise training: the decisions on sampled training subgraphs are scored jointly, so that the network they form follows the structure of the training network, while every test pair is still scored independently. Extensive experiments demonstrate that GraSPPI is effective in reconstructing held-out interaction networks, with the lowest clustering discrepancy among the compared methods and the highest complex-pathway recall.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein–protein interaction, inductive link prediction, graph prompt tuning, AI for biology

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XWRKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes GraSPPI, which uses training-network topology in two ways. First, a sequence-conditioned module predicts a small virtual graph representing shared-neighbor structure, which supplies attention prefixes to a pairwise predictor. Second, ranking, degree, and motif losses supervise predictions jointly on sampled training subgraphs. A two-stage procedure first teaches the predictor to use observed motif graphs and then adapts it to inferred graphs. Test-time prediction requires only the two protein sequences.

Critical strengths
  1. The problem formulation is useful, and the method respects the intended inference constraint. Predicting individual interactions and reconstructing a useful network are different objectives. The paper directly addresses this distinction while allowing both queried proteins to be unseen and have no recorded interactions. The combination of topology supervision during training and sequence-only inference is practically meaningful.
  2. The reconstruction results are promising, and the appendix provides informative diagnostics. GraSPPI achieves the best average rank in Table 1 and the lowest clustering discrepancy across all three PRING splits. The inferred path-product statistic remains associated with its target on unseen proteins, including after controlling for endpoint degrees: the reported test correlation changes only from 0.400 to 0.384. The density-matched experiment also shows improved edge recovery. The authors deserve credit for reporting limitations that qualify their own results.
Critical weaknesses
  1. The ablations do not isolate the contributions of the proposed components. Removing the motif prompt also changes the training streams and optimization schedule, while removing all structural losses eliminates the entire structural stream, including its cross-entropy supervision. Moreover, removing motif-graph supervision leaves prompt inference untrained because of the stop-gradients. These comparisons cannot establish whether the improvements arise from the proposed mechanisms, additional training examples, or optimization differences, weakening the evidence for the method’s central contributions.
  2. The evidence for improved topology beyond density is limited. Section 6.3.4 shows that the topology-discrepancy advantage largely disappears when predicted densities are matched; at relative density one, GraSPPI’s aggregate MMD discrepancy is slightly worse than that of the variant without structural losses (6.34 versus 5.83). Although improved edge recovery and threshold transfer remain valuable, these findings do not establish that the structural losses improve network topology beyond their effect on density. This distinction is consequential because structurally faithful reconstruction is a central claim of the paper.
Questions for authors
  1. Clarification of the ablations: Section 6.1.4 indicates that several ablations also change the training streams or optimization schedule. Do you have controlled comparisons that retain both streams, their cross-entropy supervision, and comparable training budgets while removing only the motif prompt or structural regularizers? These results would help establish which components account for the improvements.
  2. Evidence for motif-graph supervision: Have you compared against a prompt trained through the interaction objective, with the relevant stop-gradient removed, rather than a prompt frozen at initialization? Does explicit motif supervision outperform this alternative?
  3. Interpretation of the topology gains: At matched density, Section 6.3.4 reports that the aggregate MMD advantage disappears, although edge recovery improves. What evidence supports a topology benefit beyond improved density transfer? Do separate degree, clustering, and spectral comparisons at matched density support this interpretation against the strongest baselines?
2 comments

1. Controlled ablations that keep both streams and the training budget.

We agree that the original ablations changed the training setup. The revision adds two controlled ablations on PRING BFS. Both keep the 1:5 task-pair stream, the structural subgraph stream, binary cross-entropy on both streams, and the training schedule of GraSPPI. Each one removes a single component, and each row of Table R5 uses its own Vval\mathcal V_{\mathrm{val}}-frozen threshold.

  • Without the three regularizers (λrank=λdeg=λmot=0\lambda_{\mathrm{rank}}=\lambda_{\mathrm{deg}}=\lambda_{\mathrm{mot}}=0; the prompt and Lgraph\mathcal L_{\mathrm{graph}} stay). All five metrics get worse: GS falls from 0.413 to 0.354, RD falls from 0.894 to 0.665, and the three MMD ratios rise by a factor of 1.16 to 1.26. The regularizers thus improve edge recovery, density and network shape together.
  • Without the motif prompt (the prompt and Lgraph\mathcal L_{\mathrm{graph}} are removed; the three regularizers stay). RD falls from 0.894 to 0.522 and the three MMD ratios rise by a factor of 1.8 to 2.1, while GS is 0.009 higher (0.422 against 0.413). The prompt thus gives a large reconstruction benefit: without it, the network has about half the true density and about twice the discrepancy, for a small difference in GS.
  • Single losses. The Lrank\mathcal L_{\mathrm{rank}} row of Table 4 completes the three single-loss comparisons. Each removal moves RD away from 1 (0.614, 0.791 and 0.727) and raises all three MMD ratios, so each loss contributes. GS stays within 0.015 of GraSPPI in each of these rows and falls to 0.354 only when all three losses are removed, so the three losses complement each other.

Table R5. Controlled ablations, BFS test.

VariantGS ↑RD → 1Degree ↓Clustering ↓Spectral ↓
GraSPPI0.4130.8946.786.028.97
Without rank, degree, and motif losses0.3540.6657.857.3011.31
Without motif prompt0.4220.52213.0610.5918.77
Without Lrank\mathcal L_{\mathrm{rank}}0.4280.6147.526.4911.81
Without Ldeg\mathcal L_{\mathrm{deg}}0.4040.7917.446.9510.26
Without Lmot\mathcal L_{\mathrm{mot}}0.4090.7278.117.5011.28

Section 4.5 and Table 4 of the revision report these comparisons, and Section 6.1.4 ("Ablation variants") states the streams and losses that each run keeps.

2. Prompt trained through the interaction objective.

The revision adds a task-trained prompt, in which the gradients of the predictor's losses reach prompt inference PϕP_\phi. The cross-entropy on both streams and the three network-level regularizers (Lrank\mathcal L_{\mathrm{rank}}, Ldeg\mathcal L_{\mathrm{deg}}, Lmot\mathcal L_{\mathrm{mot}}), which together are LI\mathcal L_{\mathrm{I}}, update PϕP_\phi through the reader and the predictor. Table R6 compares it with explicit motif-graph supervision (GraSPPI) and with the frozen prompt. The three runs share the Stage I checkpoint, both streams, Ew=4E_w=4 and the learning rates of GraSPPI, and differ in two settings only:

  • Loss. The task-trained run and the frozen run disable the full Lgraph\mathcal L_{\mathrm{graph}} of Eq. (9): the slot term ℓslot(B^uv,Buv∗)\ell_{\mathrm{slot}}(\widehat B_{uv},B^{\ast}_{uv}) and the presence term λπBCE⁡(πuv,auv)\lambda_\pi\operatorname{BCE}(\pi_{uv},a_{uv}). Eq. (10) then becomes LII=LI\mathcal L_{\mathrm{II}}=\mathcal L_{\mathrm{I}}, with LI\mathcal L_{\mathrm{I}} unchanged. These two runs have no direct motif supervision.
  • Stop-gradient. The task-trained run also removes sg⁡(B^uv)\operatorname{sg}(\widehat B_{uv}) at the input of the predictor (Eq. 10), so LI\mathcal L_{\mathrm{I}} trains PϕP_\phi: through the frozen predictor in phase A, and together with it in phase B. sg⁡(Hu),sg⁡(Hv)\operatorname{sg}(H_u),\operatorname{sg}(H_v) stays in all runs. In the frozen run, no loss trains PϕP_\phi.

Table R6. Signal that trains the prompt, BFS test.

VariantLgraph\mathcal L_{\mathrm{graph}} (Eq. 9)sg⁡(B^uv)\operatorname{sg}(\widehat B_{uv}) (Eq. 10)GS ↑RD → 1Degree ↓Clustering ↓Spectral ↓
GraSPPI (explicit motif supervision)onkept0.4130.8946.786.028.97
Task-trained promptoffremoved0.4290.6069.678.2413.89
Frozen promptoffkept0.4280.6129.558.1613.67

The task-trained prompt and the frozen prompt give nearly the same result on all five metrics. Explicit motif supervision gives much better density transfer (RD 0.894 against 0.606) and MMD ratios 1.4 to 1.5 times lower, with GS within 0.016 of the two alternatives. Explicit motif supervision thus outperforms the task-trained prompt in the network that the model reconstructs at the deployed threshold. Table 4 and Section 4.5 of the revision report this comparison, and Section 6.1.4 states these settings.

3. Topology gain beyond density.

First, we respectfully clarify the claim. The paper does not claim that the structural losses improve the shape of the network independently of its density. The claim is about reconstruction at inference: each test pair is scored independently, one threshold is frozen on Vval\mathcal V_{\mathrm{val}}, and the network that these decisions form is closer to the true test network. The structural losses serve this claim during training, where they make the decisions on sampled training subgraphs form a network that follows the structure of the training network (abstract and the second contribution). Density is part of the structure of a network, so a comparison at matched density does not test this claim, for three reasons:

  • Density error is the distortion that the paper sets out to correct. At their published thresholds, TUnA and PLM-interact (35M) predict 2.2 to 4.4 times the true density, and every discrepancy rises with it (Section 4.3).
  • The losses supervise density by design. Ldeg\mathcal L_{\mathrm{deg}} matches the expected degree of each protein to its true degree, and the degrees fix the edge count. A comparison at matched density thus removes the quantity that the losses are trained to correct.
  • A matched-density comparison is not available at inference. The test density is unknown, and the comparison needs the test labels.

Second, we give the comparison that the reviewer asks for, as a diagnostic to determine which gains persist when density is held fixed. For the regularizers, Section 6.3.4 of the revision now uses the controlled variant without the three regularizers, in place of the earlier variant that removed the structural stream. On BFS at relative density 1, GraSPPI has higher graph similarity than this variant (0.415 against 0.371), and all three MMD ratios are numerically higher for GraSPPI (6.05, 5.71 and 7.36 against 5.44, 5.53 and 6.71). For the baselines, Table R7, which is Table 12 in Section 6.3.4 of the revision, gives the separate comparisons on the three splits.

Table R7. Baselines at test relative density 1, three splits. Columns per split: GS ↑, degree ↓, clustering ↓, spectral ↓. †: matched on the arithmetic mean of RD over the test subgraphs, because one of the 500 subgraphs has no predicted edge and the geometric mean used for the other cells is then 0. L3-PPI has tied scores, so its RD is 1.012, 0.968 and 0.963.

MethodBFS GSDeg.Clus.Spec.DFS GSDeg.Clus.Spec.RW GSDeg.Clus.Spec.
GraSPPI0.4156.055.717.360.51084.1416.9718.260.51717.295.637.91
TUnA0.4113.874.606.890.498107.5925.8423.260.53416.125.218.55
LLP0.4263.894.246.530.455125.4828.0925.580.56022.478.3211.44
L3-PPI0.3358.5010.8120.410.30094.4458.3614.350.44116.167.818.38
PLM-interact (35M)0.42213.938.9611.180.448179.2244.2827.570.496†45.62†12.85†15.27†
UniPrompt0.3858.657.2013.170.346151.5529.9631.630.43643.2318.7221.78
SPRINT0.2634.437.906.420.260109.3327.2724.810.22810.879.6510.96

The matched-density results vary by split. GraSPPI has the highest graph similarity and the lowest degree and clustering discrepancies on DFS. On BFS, TUnA and LLP have lower discrepancies on all three measures. On RW, GraSPPI's spectral discrepancy is lower than either baseline's, while TUnA has lower degree and clustering discrepancies. The average ranks across the 12 columns are 2.33 for TUnA, 2.42 for GraSPPI, and 3.17 for LLP.

Under the deployed protocol, one threshold is selected on validation and applied unchanged to test. GraSPPI reaches relative densities of 0.894, 0.942, and 1.058 across the three splits. On BFS, removing the three regularizers moves relative density to 0.665. Also on BFS, TUnA and LLP reach relative densities of 0.569 and 0.369 with their validation-frozen thresholds, and both have higher MMD ratios than GraSPPI (Table 1). These results support two specific benefits of the regularizers: the validation-selected threshold transfers to a density closer to the true one, and edge overlap improves even when test density is matched.

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

To address the challenges of topological distortion in predicted PPI networks and the lack of network context for unseen proteins, this paper proposes GraSPPI, a motif graph prompt tuning framework for inductive PPI prediction. At the pair level, the authors design a fixed motif template based on triadic closure and adopt a two-stage training strategy. At the network level, the authors further introduce three structural losses, including neighbor ranking, degree matching, and motif matching, to jointly regularize the predictions of multiple protein pairs in sampled training subgraphs, encouraging the resulting networks to better preserve the topological characteristics of the training network.

Critical strengths
  1. The paper combines local motif graph supervision with network-level structural regularization, allowing the model to leverage network topology during training while relying solely on protein sequences at inference time. The overall design is reasonable.

  2. The paper adopts a two-stage training strategy, first training the graph reader and pair predictor using true motif graphs, and then learning to infer virtual motif graphs from protein sequences. This design provides a clear separation between learning to interpret structural information and learning to infer it from sequences.

Critical weaknesses
  1. The necessity of the explicit virtual motif graph representation is not sufficiently justified. The idea of transferring training-network topology into a sequence-only predictor is closely related to existing graph knowledge distillation approaches. Although GraSPPI introduces an explicit virtual motif graph, it is unclear whether this representation provides additional benefits over directly predicting structural statistics, such as shared-neighbor counts or closure strengths. The paper lacks controlled comparisons to establish the necessity of the proposed motif graph prompting mechanism.

  2. The fixed motif template design lacks sufficient justification. The model uses a fixed template with k=8k=8 closure slots and selects shared neighbors in increasing order of degree. Although the paper demonstrates the predictive value of triadic closure, it does not provide a sensitivity analysis of kk or evaluate alternative neighbor selection strategies. It remains unclear how the template capacity affects structural information preservation and prediction performance.

Questions for authors
  1. What are the advantages of explicitly constructing virtual motif graphs over directly predicting structural statistics? Since the motif-graph supervision primarily matches sorted closure-path strengths and shared-neighbor presence, could a simpler sequence-based model directly predict these statistics and use them as auxiliary features for PPI prediction? A controlled comparison would help clarify the contribution of the explicit motif graph representation.

  2. How sensitive is GraSPPI to the number of closure slots kk? What motivates the choice of k=8k=8, and what proportion of training protein pairs have more than eight shared neighbors? Could the authors report performance with different values of kk and explain whether the degree-based neighbor selection strategy provides advantages over random selection?

1 comment

1. Explicit motif graph vs. directly predicted structural statistics.

We tested the direct-prediction control that the reviewer suggests (Section 4.5 and Table 4; details in Section 6.3.5 and Table 13).

The control. The model predicts three statistics from the two sequences: shared-neighbour presence, the fraction of filled closure slots, and the sum of the retained resource-allocation terms divided by the number of slots. The targets come from the same compiled motif graphs that supervise GraSPPI. The prompt supervision keeps the presence term and matches the other two statistics in place of the sorted closure-path strengths. A parameter-matched MLP replaces the graph reader and converts the three scalars into the same endpoint and relation tokens. The network-level structural losses, the training schedule and the model-selection rules are held fixed.

What the graph keeps beyond the three scalars. The explicit motif graph represents each shared-neighbour path with its own weight. Two pairs can have the same presence, count and total closure strength, and still distribute that strength differently across their paths. The three scalars are identical for such pairs; the motif graph is not.

Which outcomes improve. Table 13 compares the two models on PRING BFS, each at its own Vval\mathcal V_{\mathrm{val}}-frozen threshold. The explicit motif graph gives better density transfer (RD 0.894 against 0.600) and lower discrepancies: the three MMD ratios of the scalar control are 1.6 to 1.7 times those of GraSPPI. The scalar control has slightly higher edge overlap (GS 0.429 against 0.413). The comparison thus supports the explicit representation for network reconstruction.

2. Template capacity kk and witness selection.

We selected k=8k=8 on Vval\mathcal V_{\mathrm{val}} from k∈{4,8,12}k\in\{4,8,12\} (Table 6). Table R2 gives the best validation trial of each capacity in our 20-trial hyperparameter search: k=8k=8 has the best mean rank, the relative density closest to one and the lowest spectral ratio. The revision reports the coverage analysis and the retrained variants in Section 6.3.6 and Table 14, with a summary in Section 4.6.

Table R2. Best validation trial of each kk (Vval\mathcal V_{\mathrm{val}}, PRING BFS). The search also varies the other hyperparameters. Mean rank: over the five selection metrics (Section 6.2.4).

kkMean rank ↓GS ↑RD → 1Degree ↓Clustering ↓Spectral ↓
46.60.3951.13111.504.808.47
8 (selected)4.00.3831.0917.872.937.23
126.60.3681.1037.412.877.28

How often are neighbourhoods truncated? On the effective training graph of Table 7, 45.1%, 41.5% and 14.1% of the training edges have more than eight shared neighbours on BFS, DFS and RW (Table R3). Over all pairs of the 1:5 task stream (Section 6.1.4), interacting and non-interacting, the shares are 7.0%, 7.1% and 2.0%, because non-interacting pairs rarely have more than eight (0.5%, 1.1% and 0.03%).

Table R3. Share of training edges by number of shared neighbours (queried edge removed, self-interactions not counted). Last column: all pairs of the task stream.

Split> 4> 8> 12Median> 8, all pairs
BFS56.2%45.1%37.3%67.0%
DFS51.1%41.5%34.1%57.1%
RW26.7%14.1%7.5%12.0%

Why select neighbours by degree? A shared neighbour of degree dd contributes 1/d1/d to the resource allocation index. Selection in increasing order of degree therefore keeps the kk largest terms. At k=8k=8, the template retains on average 69%, 71% and 93% of the index on BFS, DFS and RW, against 63%, 65% and 91% under random selection (Table R4).

Table R4. Mean share of the resource allocation index retained (degree-based / random), over training edges with a shared neighbour.

Splitk=4k=4k=6k=6k=8k=8k=10k=10k=12k=12k=16k=16
BFS0.56 / 0.480.64 / 0.570.69 / 0.630.74 / 0.680.77 / 0.720.81 / 0.78
DFS0.59 / 0.510.66 / 0.590.71 / 0.650.75 / 0.690.78 / 0.730.82 / 0.77
RW0.83 / 0.780.89 / 0.860.93 / 0.910.95 / 0.940.96 / 0.950.98 / 0.97

How do these choices affect reconstruction? On the BFS test split, k=4k=4 and k=8k=8 have similar graph similarity and degree discrepancy. The four-slot model gives better density and spectral results, while the eight-slot model gives the highest graph similarity and the lowest clustering discrepancy. The evidence thus supports k=8k=8 as the validation-selected setting. A capacity of 12 degrades all five metrics relative to k=8k=8. Random selection changes which neighbours enter the template: a random subset of the shared neighbours, fixed for each pair, fills the slots. At k=8k=8, degree-based selection is better than random selection on all five metrics (Table 14).

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

The paper proposes a two-stage training method that first trains a model to estimate PPI conditioned on a motif graph and then adapts the model to generate a virtual graph for inference without a provided motif graph. Except that, several techniques have been proposed and achieved effective results.

Critical strengths

The proposed method is clear and seems easy to follow.

Critical weaknesses
  1. The selection of the threshold τ\tau has certain effects. It would be better for the author to provide the influence of the proposed threshold selection scheme in ablation study.

  2. The paper evaluates the virtual graph module only via indirect downstream ablation, without quantitatively assessing whether the inferred virtual graphs match true local subgraphs. A direct quantitative evaluation is needed to verify that the model captures biologically meaningful micro-topology rather than degenerate regularizations.

Questions for authors

See weakness

1 comment

1. Influence of the threshold selection scheme.

We evaluated GraSPPI under both threshold rules in Table 16 (Section 6.3.8). The validation-based rule selects, for each model, the threshold that brings relative density closest to one on VvalV_{\mathrm{val}}, then applies it unchanged to the test split. We also evaluated GraSPPI at PRING’s fixed probability threshold of 0.5.

Changing from 0.5 to the validation-selected threshold brings GraSPPI’s test relative density closer to one on all three splits. On BFS, where the effect is largest, relative density changes from 1.84 to 0.894, while graph similarity changes from 0.407 to 0.413. On DFS and RW, relative density changes from 1.06 to 0.942 and from 1.10 to 1.058, respectively. The discrepancy metrics respond differently to the threshold. For example, on BFS the validation-selected threshold reduces the degree and clustering discrepancies, while the spectral discrepancy is lower at 0.5.

GraSPPI ranks first overall under both comparisons. Among the eight methods evaluated with the common validation-based protocol, its average rank is 1.87, followed by TUnA at 3.00 and L3-PPI at 3.40. At probability 0.5, GraSPPI ranks first in 14 of the 15 metric columns against the published PRING results, with an average rank of 1.27. Thus the threshold affects the reconstructed networks, particularly their density, and GraSPPI retains the best overall rank in both comparison settings.

2. Direct evaluation of the inferred graphs against the true neighbourhoods.

Table 10 (Section 6.3.2) gives the direct evaluation that the reviewer asks for. The revision adds a decoder to its last block, and Section 4.6 of the main text now summarises these results.

Agreement with the observed motifs. The sum of the inferred closure-path products correlates with the corresponding sum in the observed motif graphs on test pairs (Spearman ρ=0.400\rho=0.400). The correlation remains at 0.384 after controlling for both endpoint degrees. It is also positive within both interaction classes, at 0.357 for interacting pairs and 0.169 for non-interacting pairs. The inferred shared-neighbour presence score achieves a test AUROC of 0.693.

Closure information in the prompt read by the classifier. We froze GraSPPI and trained a small MLP to recover the compiled motif from the three graph-reader tokens alone. The targets are ordered by closure-path strength. We score the decoded graph on slot occupancy, shared-neighbour presence, the number of occupied slots, which is capped at the template capacity, and the summed closure-path product (Table R1).

Table R1. Closure quantities decoded from the prompt of the frozen model, PRING BFS. This is the last block of Table 10 of the revision.

Decoded quantityTrainingValidationTest
Closure-slot occupancy, AUROC0.8270.7830.722
Shared-neighbour presence, AUROC0.8500.7890.681
Number of occupied slots, Spearman ρ\rho0.6020.5380.409
Summed closure-path product, Spearman ρ\rho0.6060.5320.403

A constant prompt provides no pair-specific information for these predictions. The results show that the learned prompt retains information about the presence, extent, and total strength of triadic closure for pairs of unseen proteins.

Effect on reconstruction. At the same checkpoint and threshold, permuting the inferred graphs across test pairs preserves the distribution of prompts but breaks their correspondence to the queries. Graph similarity drops from 0.413 to 0.372 and relative density from 0.894 to 0.540. Replacing all inferred graphs with the training-mean graph further lowers relative density to 0.200 (Section 6.3.3, Table 11).

Comments on the paper

4 comments

We thank the three reviewers for their careful reading and constructive comments. We answer each reviewer's points below, in one reply.

Response to Reviewer NTPB

1. Influence of the threshold selection scheme.

We evaluated GraSPPI under both threshold rules in Table 16 (Section 6.3.8). The validation-based rule selects, for each model, the threshold that brings relative density closest to one on VvalV_{\mathrm{val}}, then applies it unchanged to the test split. We also evaluated GraSPPI at PRING’s fixed probability threshold of 0.5.

Changing from 0.5 to the validation-selected threshold brings GraSPPI’s test relative density closer to one on all three splits. On BFS, where the effect is largest, relative density changes from 1.84 to 0.894, while graph similarity changes from 0.407 to 0.413. On DFS and RW, relative density changes from 1.06 to 0.942 and from 1.10 to 1.058, respectively. The discrepancy metrics respond differently to the threshold. For example, on BFS the validation-selected threshold reduces the degree and clustering discrepancies, while the spectral discrepancy is lower at 0.5.

GraSPPI ranks first overall under both comparisons. Among the eight methods evaluated with the common validation-based protocol, its average rank is 1.87, followed by TUnA at 3.00 and L3-PPI at 3.40. At probability 0.5, GraSPPI ranks first in 14 of the 15 metric columns against the published PRING results, with an average rank of 1.27. Thus the threshold affects the reconstructed networks, particularly their density, and GraSPPI retains the best overall rank in both comparison settings.

2. Direct evaluation of the inferred graphs against the true neighbourhoods.

Table 10 (Section 6.3.2) gives the direct evaluation that the reviewer asks for. The revision adds a decoder to its last block, and Section 4.6 of the main text now summarises these results.

Agreement with the observed motifs. The sum of the inferred closure-path products correlates with the corresponding sum in the observed motif graphs on test pairs (Spearman ρ=0.400\rho=0.400). The correlation remains at 0.384 after controlling for both endpoint degrees. It is also positive within both interaction classes, at 0.357 for interacting pairs and 0.169 for non-interacting pairs. The inferred shared-neighbour presence score achieves a test AUROC of 0.693.

Closure information in the prompt read by the classifier. We froze GraSPPI and trained a small MLP to recover the compiled motif from the three graph-reader tokens alone. The targets are ordered by closure-path strength. We score the decoded graph on slot occupancy, shared-neighbour presence, the number of occupied slots, which is capped at the template capacity, and the summed closure-path product (Table R1).

Table R1. Closure quantities decoded from the prompt of the frozen model, PRING BFS. This is the last block of Table 10 of the revision.

Decoded quantityTrainingValidationTest
Closure-slot occupancy, AUROC0.8270.7830.722
Shared-neighbour presence, AUROC0.8500.7890.681
Number of occupied slots, Spearman ρ\rho0.6020.5380.409
Summed closure-path product, Spearman ρ\rho0.6060.5320.403

A constant prompt provides no pair-specific information for these predictions. The results show that the learned prompt retains information about the presence, extent, and total strength of triadic closure for pairs of unseen proteins.

Effect on reconstruction. At the same checkpoint and threshold, permuting the inferred graphs across test pairs preserves the distribution of prompts but breaks their correspondence to the queries. Graph similarity drops from 0.413 to 0.372 and relative density from 0.894 to 0.540. Replacing all inferred graphs with the training-mean graph further lowers relative density to 0.200 (Section 6.3.3, Table 11).

Response to Reviewer YYEN

1. Explicit motif graph vs. directly predicted structural statistics.

We tested the direct-prediction control that the reviewer suggests (Section 4.5 and Table 4; details in Section 6.3.5 and Table 13).

The control. The model predicts three statistics from the two sequences: shared-neighbour presence, the fraction of filled closure slots, and the sum of the retained resource-allocation terms divided by the number of slots. The targets come from the same compiled motif graphs that supervise GraSPPI. The prompt supervision keeps the presence term and matches the other two statistics in place of the sorted closure-path strengths. A parameter-matched MLP replaces the graph reader and converts the three scalars into the same endpoint and relation tokens. The network-level structural losses, the training schedule and the model-selection rules are held fixed.

What the graph keeps beyond the three scalars. The explicit motif graph represents each shared-neighbour path with its own weight. Two pairs can have the same presence, count and total closure strength, and still distribute that strength differently across their paths. The three scalars are identical for such pairs; the motif graph is not.

Which outcomes improve. Table 13 compares the two models on PRING BFS, each at its own Vval\mathcal V_{\mathrm{val}}-frozen threshold. The explicit motif graph gives better density transfer (RD 0.894 against 0.600) and lower discrepancies: the three MMD ratios of the scalar control are 1.6 to 1.7 times those of GraSPPI. The scalar control has slightly higher edge overlap (GS 0.429 against 0.413). The comparison thus supports the explicit representation for network reconstruction.

2. Template capacity kk and witness selection.

We selected k=8k=8 on Vval\mathcal V_{\mathrm{val}} from k∈{4,8,12}k\in\{4,8,12\} (Table 6). Table R2 gives the best validation trial of each capacity in our 20-trial hyperparameter search: k=8k=8 has the best mean rank, the relative density closest to one and the lowest spectral ratio. The revision reports the coverage analysis and the retrained variants in Section 6.3.6 and Table 14, with a summary in Section 4.6.

Table R2. Best validation trial of each kk (Vval\mathcal V_{\mathrm{val}}, PRING BFS). The search also varies the other hyperparameters. Mean rank: over the five selection metrics (Section 6.2.4).

kkMean rank ↓GS ↑RD → 1Degree ↓Clustering ↓Spectral ↓
46.60.3951.13111.504.808.47
8 (selected)4.00.3831.0917.872.937.23
126.60.3681.1037.412.877.28

How often are neighbourhoods truncated? On the effective training graph of Table 7, 45.1%, 41.5% and 14.1% of the training edges have more than eight shared neighbours on BFS, DFS and RW (Table R3). Over all pairs of the 1:5 task stream (Section 6.1.4), interacting and non-interacting, the shares are 7.0%, 7.1% and 2.0%, because non-interacting pairs rarely have more than eight (0.5%, 1.1% and 0.03%).

Table R3. Share of training edges by number of shared neighbours (queried edge removed, self-interactions not counted). Last column: all pairs of the task stream.

Split> 4> 8> 12Median> 8, all pairs
BFS56.2%45.1%37.3%67.0%
DFS51.1%41.5%34.1%57.1%
RW26.7%14.1%7.5%12.0%

Why select neighbours by degree? A shared neighbour of degree dd contributes 1/d1/d to the resource allocation index. Selection in increasing order of degree therefore keeps the kk largest terms. At k=8k=8, the template retains on average 69%, 71% and 93% of the index on BFS, DFS and RW, against 63%, 65% and 91% under random selection (Table R4).

Table R4. Mean share of the resource allocation index retained (degree-based / random), over training edges with a shared neighbour.

Splitk=4k=4k=6k=6k=8k=8k=10k=10k=12k=12k=16k=16
BFS0.56 / 0.480.64 / 0.570.69 / 0.630.74 / 0.680.77 / 0.720.81 / 0.78
DFS0.59 / 0.510.66 / 0.590.71 / 0.650.75 / 0.690.78 / 0.730.82 / 0.77
RW0.83 / 0.780.89 / 0.860.93 / 0.910.95 / 0.940.96 / 0.950.98 / 0.97

How do these choices affect reconstruction? On the BFS test split, k=4k=4 and k=8k=8 have similar graph similarity and degree discrepancy. The four-slot model gives better density and spectral results, while the eight-slot model gives the highest graph similarity and the lowest clustering discrepancy. The evidence thus supports k=8k=8 as the validation-selected setting. A capacity of 12 degrades all five metrics relative to k=8k=8. Random selection changes which neighbours enter the template: a random subset of the shared neighbours, fixed for each pair, fills the slots. At k=8k=8, degree-based selection is better than random selection on all five metrics (Table 14).

Response to Reviewer XWRK

1. Controlled ablations that keep both streams and the training budget.

We agree that the original ablations changed the training setup. The revision adds two controlled ablations on PRING BFS. Both keep the 1:5 task-pair stream, the structural subgraph stream, binary cross-entropy on both streams, and the training schedule of GraSPPI. Each one removes a single component, and each row of Table R5 uses its own Vval\mathcal V_{\mathrm{val}}-frozen threshold.

  • Without the three regularizers (λrank=λdeg=λmot=0\lambda_{\mathrm{rank}}=\lambda_{\mathrm{deg}}=\lambda_{\mathrm{mot}}=0; the prompt and Lgraph\mathcal L_{\mathrm{graph}} stay). All five metrics get worse: GS falls from 0.413 to 0.354, RD falls from 0.894 to 0.665, and the three MMD ratios rise by a factor of 1.16 to 1.26. The regularizers thus improve edge recovery, density and network shape together.
  • Without the motif prompt (the prompt and Lgraph\mathcal L_{\mathrm{graph}} are removed; the three regularizers stay). RD falls from 0.894 to 0.522 and the three MMD ratios rise by a factor of 1.8 to 2.1, while GS is 0.009 higher (0.422 against 0.413). The prompt thus gives a large reconstruction benefit: without it, the network has about half the true density and about twice the discrepancy, for a small difference in GS.
  • Single losses. The Lrank\mathcal L_{\mathrm{rank}} row of Table 4 completes the three single-loss comparisons. Each removal moves RD away from 1 (0.614, 0.791 and 0.727) and raises all three MMD ratios, so each loss contributes. GS stays within 0.015 of GraSPPI in each of these rows and falls to 0.354 only when all three losses are removed, so the three losses complement each other.

Table R5. Controlled ablations, BFS test.

VariantGS ↑RD → 1Degree ↓Clustering ↓Spectral ↓
GraSPPI0.4130.8946.786.028.97
Without rank, degree, and motif losses0.3540.6657.857.3011.31
Without motif prompt0.4220.52213.0610.5918.77
Without Lrank\mathcal L_{\mathrm{rank}}0.4280.6147.526.4911.81
Without Ldeg\mathcal L_{\mathrm{deg}}0.4040.7917.446.9510.26
Without Lmot\mathcal L_{\mathrm{mot}}0.4090.7278.117.5011.28

Section 4.5 and Table 4 of the revision report these comparisons, and Section 6.1.4 ("Ablation variants") states the streams and losses that each run keeps.

2. Prompt trained through the interaction objective.

The revision adds a task-trained prompt, in which the gradients of the predictor's losses reach prompt inference PϕP_\phi. The cross-entropy on both streams and the three network-level regularizers (Lrank\mathcal L_{\mathrm{rank}}, Ldeg\mathcal L_{\mathrm{deg}}, Lmot\mathcal L_{\mathrm{mot}}), which together are LI\mathcal L_{\mathrm{I}}, update PϕP_\phi through the reader and the predictor. Table R6 compares it with explicit motif-graph supervision (GraSPPI) and with the frozen prompt. The three runs share the Stage I checkpoint, both streams, Ew=4E_w=4 and the learning rates of GraSPPI, and differ in two settings only:

  • Loss. The task-trained run and the frozen run disable the full Lgraph\mathcal L_{\mathrm{graph}} of Eq. (9): the slot term ℓslot(B^uv,Buv∗)\ell_{\mathrm{slot}}(\widehat B_{uv},B^{\ast}_{uv}) and the presence term λπBCE⁡(πuv,auv)\lambda_\pi\operatorname{BCE}(\pi_{uv},a_{uv}). Eq. (10) then becomes LII=LI\mathcal L_{\mathrm{II}}=\mathcal L_{\mathrm{I}}, with LI\mathcal L_{\mathrm{I}} unchanged. These two runs have no direct motif supervision.
  • Stop-gradient. The task-trained run also removes sg⁡(B^uv)\operatorname{sg}(\widehat B_{uv}) at the input of the predictor (Eq. 10), so LI\mathcal L_{\mathrm{I}} trains PϕP_\phi: through the frozen predictor in phase A, and together with it in phase B. sg⁡(Hu),sg⁡(Hv)\operatorname{sg}(H_u),\operatorname{sg}(H_v) stays in all runs. In the frozen run, no loss trains PϕP_\phi.

Table R6. Signal that trains the prompt, BFS test.

VariantLgraph\mathcal L_{\mathrm{graph}} (Eq. 9)sg⁡(B^uv)\operatorname{sg}(\widehat B_{uv}) (Eq. 10)GS ↑RD → 1Degree ↓Clustering ↓Spectral ↓
GraSPPI (explicit motif supervision)onkept0.4130.8946.786.028.97
Task-trained promptoffremoved0.4290.6069.678.2413.89
Frozen promptoffkept0.4280.6129.558.1613.67

The task-trained prompt and the frozen prompt give nearly the same result on all five metrics. Explicit motif supervision gives much better density transfer (RD 0.894 against 0.606) and MMD ratios 1.4 to 1.5 times lower, with GS within 0.016 of the two alternatives. Explicit motif supervision thus outperforms the task-trained prompt in the network that the model reconstructs at the deployed threshold. Table 4 and Section 4.5 of the revision report this comparison, and Section 6.1.4 states these settings.

Response to Reviewer XWRK

3. Topology gain beyond density.

First, we respectfully clarify the claim. The paper does not claim that the structural losses improve the shape of the network independently of its density. The claim is about reconstruction at inference: each test pair is scored independently, one threshold is frozen on Vval\mathcal V_{\mathrm{val}}, and the network that these decisions form is closer to the true test network. The structural losses serve this claim during training, where they make the decisions on sampled training subgraphs form a network that follows the structure of the training network (abstract and the second contribution). Density is part of the structure of a network, so a comparison at matched density does not test this claim, for three reasons:

  • Density error is the distortion that the paper sets out to correct. At their published thresholds, TUnA and PLM-interact (35M) predict 2.2 to 4.4 times the true density, and every discrepancy rises with it (Section 4.3).
  • The losses supervise density by design. Ldeg\mathcal L_{\mathrm{deg}} matches the expected degree of each protein to its true degree, and the degrees fix the edge count. A comparison at matched density thus removes the quantity that the losses are trained to correct.
  • A matched-density comparison is not available at inference. The test density is unknown, and the comparison needs the test labels.

Second, we give the comparison that the reviewer asks for, as a diagnostic to determine which gains persist when density is held fixed. For the regularizers, Section 6.3.4 of the revision now uses the controlled variant without the three regularizers, in place of the earlier variant that removed the structural stream. On BFS at relative density 1, GraSPPI has higher graph similarity than this variant (0.415 against 0.371), and all three MMD ratios are numerically higher for GraSPPI (6.05, 5.71 and 7.36 against 5.44, 5.53 and 6.71). For the baselines, Table R7, which is Table 12 in Section 6.3.4 of the revision, gives the separate comparisons on the three splits.

Table R7. Baselines at test relative density 1, three splits. Columns per split: GS ↑, degree ↓, clustering ↓, spectral ↓. †: matched on the arithmetic mean of RD over the test subgraphs, because one of the 500 subgraphs has no predicted edge and the geometric mean used for the other cells is then 0. L3-PPI has tied scores, so its RD is 1.012, 0.968 and 0.963.

MethodBFS GSDeg.Clus.Spec.DFS GSDeg.Clus.Spec.RW GSDeg.Clus.Spec.
GraSPPI0.4156.055.717.360.51084.1416.9718.260.51717.295.637.91
TUnA0.4113.874.606.890.498107.5925.8423.260.53416.125.218.55
LLP0.4263.894.246.530.455125.4828.0925.580.56022.478.3211.44
L3-PPI0.3358.5010.8120.410.30094.4458.3614.350.44116.167.818.38
PLM-interact (35M)0.42213.938.9611.180.448179.2244.2827.570.496†45.62†12.85†15.27†
UniPrompt0.3858.657.2013.170.346151.5529.9631.630.43643.2318.7221.78
SPRINT0.2634.437.906.420.260109.3327.2724.810.22810.879.6510.96

The matched-density results vary by split. GraSPPI has the highest graph similarity and the lowest degree and clustering discrepancies on DFS. On BFS, TUnA and LLP have lower discrepancies on all three measures. On RW, GraSPPI's spectral discrepancy is lower than either baseline's, while TUnA has lower degree and clustering discrepancies. The average ranks across the 12 columns are 2.33 for TUnA, 2.42 for GraSPPI, and 3.17 for LLP.

Under the deployed protocol, one threshold is selected on validation and applied unchanged to test. GraSPPI reaches relative densities of 0.894, 0.942, and 1.058 across the three splits. On BFS, removing the three regularizers moves relative density to 0.665. Also on BFS, TUnA and LLP reach relative densities of 0.569 and 0.369 with their validation-frozen thresholds, and both have higher MMD ratios than GraSPPI (Table 1). These results support two specific benefits of the regularizers: the validation-selected threshold transfers to a density closer to the true one, and edge overlap improves even when test density is matched.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.