acceptodds
Withdrawn from ICLR 2027

LOCUS: Label-Gated Region Proposals for Protein Function Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Where function resides in a protein is tightly coupled to what that function is. Recent predictors pair protein language model representations with structural encoders and reduce them to a single protein-level vector. The design is effective but leaves no way to express how one region bears on one label. Catalytic residues and binding pockets are sharply localized, while many biological process and cellular component terms have no compact structural correlate at all. Routing every label through one pooled vector and routing every label through a site bottleneck therefore fail for opposite reasons. We introduce Locus, a framework that gives localized evidence its own pathway and keeps the global one. A class-agnostic region proposal network inspired by object detection emits multi-scale soft 3D regions over the residue graph, supervised only by residue-level annotations of active sites, binding sites and motifs. Graph non-maximum suppression selects among the proposals, and each survivor is pooled into a region-of-interest token. A classifier adds this regional evidence to global sequence and structure features as label-specific gated residuals. Every label then decides how much local evidence to admit, and no functional label enters the model as an input. Since the fusion is additive and decomposes exactly, the model that uses the proposal pathway can also measure its own contribution. On the standard Gene Ontology and enzyme classification benchmarks Locus exceeds published sequence- and structure-based predictors under the official evaluator. The regional pathway contributes in a strongly label-dependent way: it is worth a margin clear of zero on all three Gene Ontology branches and limited contributions on enzyme classification.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein function, Region proposal network

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ytV1OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes LOCUS framework, which adds a supervised regional pathway to global sequence and structure predictors. A class-agnostic proposal network learns functional regions from residue annotations; their pooled representations contribute through label-specific gated residuals. The paper evaluates GO and EC prediction and uses pathway removal, fixed-gate decomposition, and proposal diagnostics to examine when regional information helps.

Critical strengths

Giving local evidence its own predictive pathway while retaining global information is a sensible design. I also appreciated the effort to understand the EC result rather than simply report the overall score: the ROI route contributes little there, and Appendix C checks its normalized geometry instead of interpreting tiny raw mask values as evidence of collapse. This distinction between learning regions and benefiting from them makes the paper more informative than a performance comparison alone.

Critical weaknesses

The comparison in Section 4.5 does not isolate the gate effect. Its fixed-gate gain of 0.020 is seq+ROI minus seq. Removing ROI from the complete model instead calls for full minus seq+str, approximately 0.394−0.378=0.0160.394-0.378=0.016 using Table 5’s rounded values. These are different comparisons, so the difference between 0.020 and the reported removal effect of 0.017 cannot be attributed to gate adaptation. This needs correcting before drawing the proposed distinction between pathway value and reliance.
Moreover, the source-recovery issue is my other concern. The reproducibility statement says that two central modules and the source for the shared site-pretraining run still require reconstruction. I appreciate the disclosure, but this is more than an unreleased repository. Since every reported model uses that warm start, I would need evidence that the restored implementation agrees with the saved checkpoints and outputs. Provenance hashes alone do not establish that agreement.

Questions for authors
  • Using the same complete-model background, how do fixed-gate and recomputed-gate ROI removal compare? Does the Section 4.5 interpretation survive?
  • Has the missing source been restored, and what output-level checks establish agreement with the original implementation?
  • Could the authors provide an anonymized codebase for review, including the restored pretraining scripts and code verification to ensure the reproducibility of the warm-start checkpoints?
Reviewer_QpQEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes LOCUS, a protein function predictor that adds a region-proposal pathway, adapted from Faster R-CNN, to a frozen ESM-2 sequence encoder and a geometric residue-graph encoder. The RPN is pretrained on VenusX residue-level annotations (active sites, binding sites, motifs) and proposes up to 16 soft spherical 3D regions. These are pooled into ROI tokens and enter the final logits as a label-specific gated residual alongside sequence and structure logits. Because the fusion is additive, the authors decompose each pathway’s contribution exactly. They report state-of-the-art results on the DeepFRI/HEAL GO benchmark and competitive EC results. The ROI pathway yields small but statistically significant gains on GO (+0.005 to +0.010 Fmax) and none on EC.

Critical strengths
  1. The additive fusion permits exact pathway read-outs from a single forward pass, and the paper reports per-pathway contributions on all label spaces.

  2. A matched ESM-2 probe on the same training data and evaluator is included, which is more informative than the published baseline rows.

  3. Null and negative results are reported: the ROI pathway is inert on EC, co-training the site loss degrades EC performance, and macro checkpoint selection trades off Act/BindB against BindI/Motif.

Critical weaknesses
  1. Reproducibility. By the authors' own statement:

    • two core modules (functional_rpn.py, geometric_encoder.py) exist only as compiled bytecode;
    • the working tree is not under version control;
    • the source that produced the site checkpoint, which initializes every reported model, is no longer on disk.

    This contradicts Appendix E's claim of "a single frozen code revision." Section 3.5 refers to Appendix E for how the pipeline was consolidated, but Appendix E contains only the assertion. As it stands, the results cannot be reproduced, even by the authors.

  2. Test-set threshold selection. Fmax and Smin are taken as the max/min over a threshold sweep on the test set (Sec. 4.2). This is inherited from HEAL, but it inflates all reported numbers.

  3. Contradictory threshold protocol. Sections 4.5 and 4.7 state that ablation arms use validation-selected thresholds, while Sec. 4.2 defines Fmax as a test-set sweep. Either the ablation tables report a different metric than Table 1, or the validation-threshold statement has no effect.

  4. Unfair baseline comparison.

    • Table 1 compares against copied numbers on a different test set (3,416 vs. 3,410 chains).
    • LOCUS uses additional VenusX residue supervision that the baselines do not.
    • The EC rows in Table 2 come from different splits.
    • The published ESM-2 row (BP AUPR 0.062) is far weaker than the authors' own ESM-2 probe (0.292), which overstates the margin. Appendix A then compares against that published row anyway, despite the paper's own caution about cross-split comparisons.
  5. Weakly specified probe. The matched ESM-2 probe is described in one sentence, with no hyperparameters or tuning effort reported, so it may be undertuned. It also does not share "everything but the read-out" as claimed: LOCUS additionally uses structure and VenusX pretraining.

  6. Non-unique attribution of the ROI contribution. The value of the ROI term depends on which partial sum it is added to. On BP AUPR it is +0.020 relative to the sequence pathway but +0.016 relative to seq+str. The latter is below the removal estimate (+0.017). The claim that the ROI term is "worth more than an inference-time ablation would suggest" holds only for the more favorable baseline. On MF Fmax, full − (seq+str) equals the removal cost exactly (+0.006).

  7. Small effects, single seed.

    • The ROI pathway adds +0.005 to +0.010 Fmax on GO.
    • The bootstrap intervals capture test-set variance only, not training variance, and all models appear to be single runs.
    • Tables 4 and 8 report no uncertainty.
    • The EC row of Table 3 places the point estimate at the upper CI bound (+0.001 [−0.001, +0.001]).
  8. Internal inconsistencies.

    • The introduction argues that BP and CC terms lack compact structural correlates, yet the ROI pathway helps most on BP and CC.
    • Section 4.3 states that compact regions carry no CC information, while Section 4.5 reports that removing the ROI route costs as much on CC as on MF.
    • Section 4.7 calls the MF warm-start gain (+0.017 Fmax) "close to" the ROI-removal cost on MF, but Table 3 gives +0.006.
    • Appendix A reports the structure residual as +0.013 AUPR on both BP and CC, whereas Table 5 gives +0.012 and +0.014.
    • The Table 8 control (0.881 / 0.885) exceeds the headline EC results (0.880 / 0.881) without explanation.
  9. Limited evidence for localization in the function models. Figure 2 is a single selected protein. No aggregate measure is reported of how often function-checkpoint proposals overlap annotated sites. Appendix C shows that function-checkpoint regions differ substantially from the site checkpoint's (23.6 vs. 6.0 effective residues), so the VenusX localization results do not describe the regions used for prediction.

  10. Overstated framing.

    • The null ROI result on EC is presented as "a central finding."
    • The co-training conclusion rests on a single continuation run on one label space.
    • The Discussion's claim that "the component to change is therefore the pooling operator" follows from that one run.
  11. Presentation. The manuscript contains what read as unedited internal notes (e.g., "must be reconstructed before artifact release" in the method section) and cross-references to content that does not exist.

Questions for authors
  1. Can you release the reconstructed source for the RPN, the geometric encoder and the site pretraining stage, and show that it reproduces Tables 1, 3 and 5 within reported uncertainty?
  2. How is the "single frozen code revision" claim in Appendix E reconciled with an unversioned working tree and missing source?
  3. Which metric do the ablation tables report: test-swept Fmax, or F at a validation-selected threshold? Please report main results with validation-selected thresholds.
  4. What is the variance across training seeds for the main results and for the Table 4 warm-start ablation?
  5. What are the probe's hyperparameters, and how was it tuned?
  6. Can you report an aggregate overlap measure between function-checkpoint proposals and annotated sites, rather than a single case study
Reviewer_trNAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LOCUS predicts protein function while keeping a separate route for localized evidence. A class-agnostic region proposal network, supervised only by residue-level active-site, binding-site and motif annotations, emits multi-scale soft 3D regions; graph non-maximum suppression keeps at most sixteen, each is pooled into an ROI token, and a classifier adds this evidence to global sequence and structure features through label-specific gated residuals. Because the fusion is additive, the ROI term is readable in one forward pass with both gates frozen. On the HEAL adaptation of the DeepFRI benchmark the model reaches F_max 0.616, 0.684 and 0.774 on BP, CC and MF and 0.880 on 538-class EC; removing the ROI residual costs 0.005 to 0.017 function AUPR on the three GO branches and nothing measurable on EC.

Critical strengths
  1. The paper is built around a real asymmetry rather than a generic motivation, and it measures that asymmetry with an instrument instead of an argument. BP and CC labels often have no compact structural correlate while MF labels often do, and the architecture lets each label choose how much local evidence to admit. Section 4.5 then separates two quantities that are usually conflated: the removal ablation drops beta times the ROI residual and zeroes z_roi inside the gate input, which perturbs the structure gate, whereas Appendix A holds both gates at the values the full forward pass produced and reads the partial sums. Reporting +0.020 gate-fixed against +0.017 removal on BP, and naming the gap as the structure gate adapting to a route that is then cut, is a genuine methodological contribution, and the same section reports that a matched sequence-only probe already matches the full model on CC yet removing the ROI residual still costs 0.011 AUPR.

  2. The evidence is reported in a form a reader can check, including the parts that cut against the paper. Negative results appear as results: the ROI residual is inert on EC with intervals spanning zero, co-training the site loss is worse on both metrics (Table 8), and Appendix C shows the site checkpoint pools 6.0 effective residues while the EC function checkpoint pools 23.6, and states that localization quality measured on the former should not be attributed to the latter. The engineering likewise supports verification rather than trust: sixteen site-training arms differing only in proposal radii and the region Dice flag, residue AUPR on four annotation tasks with a matched probe, paired bootstrap intervals, and a frozen-partial-sum self-check that reproduces the full arm to 2e-7.

Critical weaknesses
  1. The related-work subsection titled "Functional residue learning and region proposals" cites only VenusX and the object-detection original, and omits the two closest precursors. ProteinRPN (Mitra, Huang & Kellis, arXiv:2409.00610) already builds a protein region proposal network over residue graphs, pools the selected regions, predicts GO terms and reports localizing functional residues in the same benchmark lineage. Magneton (Calef, Liang, Kellis & Zitnik, arXiv:2512.18114) already states that domains and functional sites are "rarely used as training signals or representation units" and shows that substructure supervision improves function prediction. Neither appears among the twenty-six references. Without them the increment over prior region-proposal work on proteins cannot be assessed, and the subsection heading promises exactly the comparison that is missing.

  2. The comparability argument for Table 1 is contradicted by the authors' own probe. Section 4.3 states that metric, threshold rule and training set all agree with the cited rows "so the comparison turns on membership alone", and puts that membership difference at 3,416 versus 3,410 chains. On the same frozen ESM-2 650M layer-33 backbone, the published row of Table 1 scores F_max 0.351, 0.633 and 0.531, while the probe trained on the authors' own proteins, embeddings, label maps and evaluator scores 0.557, 0.682 and 0.671: gaps of 0.206, 0.049 and 0.140. Six chains cannot produce a gap that large, and cannot produce one that varies fourfold across branches. Either the published rows differ in read-out or label map as well, or membership is not the only difference. On either reading the abstract's claim of exceeding published predictors is not established by Table 1.

  3. Only the loosest identity cut is reported. Section 1 motivates the task as the remote-homology regime "where test proteins share little sequence identity with the training set", and Section 4.1 then writes "we report the 95% cut only" while noting that stricter subsets are distributed with the benchmark. At 95% identity a test chain can have a near-identical training neighbour, and near neighbours share catalytic residues and often the source annotation itself. That is the regime in which a proposal pathway trained on site annotations looks most useful, so the central claim is tested where it is easiest to satisfy rather than where the introduction says it matters.

  4. Site supervision is filtered against GO but not against EC. Appendix F keeps a record only after "homology filtering against the GO validation and test proteins" and reports removing 5,964 training records with close homology to "protected GO validation or test sequences". The 1,919 EC test proteins are not in that protected set. VenusX site annotations and EC labels are drawn from the same structural and functional databases, so a site-pretraining protein can be a close relative of an EC test protein. EC is also the branch on which the negative result rests, so leakage there would affect both the 0.880 and the account of why the ROI residual is inert.

  5. Label specificity is asserted more strongly than it is measured. The abstract says every label decides how much local evidence to admit and calls the effect strongly label-dependent, but the evidence in Section 4.8 is four scalars: mean ROI gate 0.24, 0.25, 0.24 and 0.12 for BP, CC, MF and EC. No distribution over labels, over label frequency or over ontology depth is given. The gates of Equations 5 and 6 also take only [z_seq; z_str; z_roi] as input, with no label identity, so label specificity is implemented solely by assigning one output unit per label. Four per-branch means cannot separate per-label differentiation from a per-branch constant.

  6. Independence from functional labels holds literally, not informationally. The paper supports it with "they are not converted into GO labels" (Section 3.3) and "BioLiP's catalytic-residue field is not the supervision the proposal network saw" (Section 4.6). Both say the annotations are not used as targets; neither says the annotations are uninformative about the label. Many motifs and binding sites are strongly associated with particular enzyme families, so a class-agnostic head can still carry label information through its supervision. No permuted-annotation control and no label-balanced retraining are reported.

  7. The numbers and the prose disagree in nine places a reader can check against the paper's own tables. Table 1 bolds S_CC 0.463 for LOCUS although HEAL's 0.462 is better on a downward metric, and Section 4.3 says LOCUS leads "every column but the cellular-component semantic distance". Table 3 prints an EC F_max delta of +0.001 where 0.880 minus 0.880 is zero, and Appendix A says adding the ROI term "moves the fourth decimal the wrong way" on EC F_max although the column reads 0.878 then 0.878. Appendix A puts the structure residual at "+0.013 on both" BP and CC where the table gives 0.012 and 0.014. Appendix B opens by saying the site checkpoint "trails on Act and on BindB" while Table 6 has it ahead on both and Section 4.6 says it leads, and says the Act+BindB selection "halves BindI and Motif" where the drops are 37% and 23%. Table 4 supports +0.017 AUPR where the columns give 0.016; Section 4.3 reports +0.060 F_max over the probe where the difference is 0.059; and Section 4.8 points to "Appendix C records the two ROI metadata channels it leaves inert" where Appendix C contains no such passage.

Revisions recommended: add a short comparison table against ProteinRPN and Magneton covering architecture, supervision source, evaluation protocol and whether the regional contribution is separable; reconcile the threshold protocol, since Section 4.2 defines F_max as the maximum over the t/100 grid evaluated on the test set while Section 4.5 re-selects the ablated model's threshold on validation, and state which rule applies to which table or report both under one rule; report at least two identity cuts and place the main comparison on a stricter one so that the abstract matches the regime the introduction motivates; add a per-label gate figure with labels binned by training frequency, and a control in which site annotations are permuted across proteins; and correct the nine items in Weakness 7, making the first sentence of Appendix B agree with Table 6.

Questions for authors
  1. How does LOCUS differ from ProteinRPN and from Magneton in architecture, supervision and evaluation, and why are neither cited nor compared?

  2. What explains the 0.049 to 0.206 gap between the published ESM-2 rows of Table 1 and your own probe on the same backbone? If the difference is not membership alone, what supports the abstract's comparison against published predictors?

  3. What are F_max and AUPR on the stricter identity subsets the benchmark distributes, and does the ROI contribution survive there with its intervals still clear of zero?

  4. Were the VenusX site records homology-filtered against the EC test proteins? If not, does the EC result hold when such filtering is applied?

  5. What does the gate distribution look like per label and per label-frequency bin, and what happens to the ROI contribution when site annotations are permuted across proteins?

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

Related papers

Discussion (0)

Sign in to comment.