Pith. sign in

REVIEW 2 major objections 2 minor 2 cited by

BioBlobs compresses proteins into small cohesive substructures that predict function and recover catalytic sites.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-05-21 21:58 UTC pith:GS5TNC2J

load-bearing objection BioBlobs compresses proteins into task-adaptive blobs that match baselines and recover some M-CSA sites from protein labels alone, but the claims rest on untested assumptions about what the blobs actually capture. the 2 major comments →

arxiv 2510.01632 v3 pith:GS5TNC2J submitted 2025-10-02 q-bio.BM cs.AI

BioBlobs: Unsupervised Discovery of Functional Substructures for Protein Function Prediction

Xin Wang , Kaiwen Shi , Carlos Oliver This is my paper
classification q-bio.BM cs.AI
keywords protein function predictionfunctional substructuresunsupervised discoverycatalytic sitesmachine learningbioinformatics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that proteins can be reduced to a handful of substructures whose properties alone are enough to determine overall function. It does this by training a model to compress the full protein representation into these substructures, called blobs, while still solving function-prediction tasks. A sympathetic reader would care because biological functions such as catalysis and binding are known to reside in small localized regions, yet most current predictors treat every residue equally and therefore cannot point to the responsible parts. If the central claim holds, the approach supplies a practical route to identifying functional sites across the many proteins that lack detailed experimental annotations.

Core claim

BioBlobs is an encoder-agnostic, end-to-end differentiable framework that compresses a protein into a small set of cohesive substructures (blobs) and predicts function from these blobs alone, so that each blob corresponds to a candidate functional region. Across diverse protein function prediction tasks and multiple sequence- and structure-based encoders, BioBlobs matches or exceeds strong baselines while operating on only a small fraction of residues. The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains. Trained only on protein-level labels, BioBlobs recovers experimentally annotated catalytic sites in the M-CSA database

What carries the argument

The differentiable compression step that partitions a protein encoding into a compact set of blobs from which all downstream function predictions are made.

Load-bearing premise

The learned blobs correspond to biologically meaningful functional regions rather than artifacts produced by the compression objective or encoder choice.

What would settle it

A statistical test showing that the discovered blobs overlap known M-CSA catalytic sites no more than randomly chosen residue sets of the same total size would falsify the claim of unsupervised functional substructure discovery.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Accurate function prediction is possible while ignoring the great majority of a protein's residues.
  • The spatial extent of each discovered substructure changes automatically according to the function being predicted.
  • Large-scale functional-site annotation becomes feasible for proteins that have only global labels.
  • Existing protein encoders can be used unchanged inside the framework to obtain substructure-level explanations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same compression idea could be tested on other macromolecules to locate functional motifs without site-level supervision.
  • Protein design or variant interpretation pipelines could restrict attention to the identified blobs to increase precision.
  • If the blobs prove stable across related proteins, they might serve as a new unit for evolutionary comparison.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces BioBlobs, an encoder-agnostic end-to-end differentiable framework that compresses a protein representation into a small set of cohesive substructures (blobs) and performs function prediction from these blobs alone. It claims that, across diverse function-prediction tasks and both sequence- and structure-based encoders, BioBlobs matches or exceeds strong baselines while using only a small fraction of residues; additionally, when trained solely on protein-level labels it recovers experimentally annotated catalytic sites from the M-CSA database, thereby demonstrating unsupervised functional substructure discovery.

Significance. If the central claims are substantiated, the work would represent a meaningful step toward interpretable, substructure-level protein function modeling that is independent of any particular encoder. The reported adaptability of blob scale (local sites to domains) and the ability to operate with extreme sparsity are potentially valuable for downstream biological applications such as large-scale functional-site annotation.

major comments (2)
  1. [Results section] Results section (performance tables and ablation studies): the manuscript does not report controls that replace the learned blob selection with (a) random residue subsets of matched cardinality or (b) contiguous segments chosen by a non-learned heuristic. This comparison is load-bearing for the central claim that the differentiable compression produces functionally meaningful substructures rather than merely reflecting encoder bias or the sparsity objective itself.
  2. [M-CSA evaluation] M-CSA site-recovery evaluation: the overlap between discovered blobs and experimentally annotated catalytic sites is presented without an explicit null model (e.g., random placement of blobs of the same size distribution) or statistical test against chance overlap. This detail is required to support the unsupervised-discovery claim when only protein-level supervision is used.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'small fraction of residues' should be accompanied by a concrete range or average percentage to allow readers to assess the degree of compression.
  2. [Methods] Notation: the precise definition of 'cohesive' (spatial, sequence, or embedding-space) and how it is enforced in the compression objective should be stated explicitly in the methods.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and for highlighting these important controls. We address each major comment below and have revised the manuscript to incorporate the suggested analyses.

read point-by-point responses

  1. Referee: [Results section] Results section (performance tables and ablation studies): the manuscript does not report controls that replace the learned blob selection with (a) random residue subsets of matched cardinality or (b) contiguous segments chosen by a non-learned heuristic. This comparison is load-bearing for the central claim that the differentiable compression produces functionally meaningful substructures rather than merely reflecting encoder bias or the sparsity objective itself.

    Authors: We agree that direct comparisons to random residue subsets of matched cardinality and to non-learned contiguous segments are necessary to isolate the contribution of the learned, differentiable selection. In the revised manuscript we have added these controls to the Results section (new panels in the main performance tables and an expanded ablation study). The updated results show that BioBlobs consistently outperforms both random selection and heuristic contiguous segments across encoders and tasks, thereby strengthening the claim that the compression identifies functionally relevant substructures rather than merely exploiting sparsity or encoder bias. revision: yes

  2. Referee: [M-CSA evaluation] M-CSA site-recovery evaluation: the overlap between discovered blobs and experimentally annotated catalytic sites is presented without an explicit null model (e.g., random placement of blobs of the same size distribution) or statistical test against chance overlap. This detail is required to support the unsupervised-discovery claim when only protein-level supervision is used.

    Authors: We acknowledge that an explicit null model and statistical test are required to rigorously support the unsupervised-discovery claim. In the revised manuscript we have added a permutation-based null model that randomly places blobs while preserving the observed size distribution. We now report the resulting p-values (or equivalent significance measures) for the overlap with M-CSA catalytic sites, confirming that the observed recovery is statistically significant above chance under protein-level supervision only. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical framework with independent validation

full rationale

The paper introduces BioBlobs as an encoder-agnostic differentiable compression method that selects sparse substructures (blobs) and predicts protein function from them alone. All central claims—matching baselines on diverse tasks, adapting spatial scale, and recovering M-CSA catalytic sites—are presented as empirical outcomes from training on protein-level labels without site supervision. No equations, derivations, or self-citations appear in the provided text that would reduce any result to a fitted parameter or prior author work by construction. The unsupervised discovery claim is a description of the training regime (protein labels only) rather than a definitional loop, and the method's performance is benchmarked against external baselines and databases, keeping the derivation chain self-contained and falsifiable.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The framework implicitly assumes that cohesive substructures can be discovered differentiably from sequence or structure encoders and that protein-level supervision suffices for site-level discovery.

pith-pipeline@v0.9.0 · 5727 in / 1083 out tokens · 25196 ms · 2026-05-21T21:58:45.335860+00:00 · methodology

0 comments
read the original abstract

Protein function is driven by cohesive substructures, such as catalytic triads, binding pockets, and structural motifs, that occupy only a small fraction of a protein's residues. Yet existing pipelines built on protein encoders do not model proteins at the substructure level, leaving the central biological question unanswered: which substructure of a protein is responsible for its function? We introduce BioBlobs, an encoder-agnostic, end-to-end differentiable framework that compresses a protein into a small set of cohesive substructures (blobs) and predicts function from these blobs alone, so that each blob corresponds to a candidate functional region. Across diverse protein function prediction tasks and multiple sequence- and structure-based encoders, BioBlobs matches or exceeds strong baselines while operating on only a small fraction of residues. The discovered blobs adapt their spatial scale to the task, ranging from local catalytic sites to entire structural domains. Trained only on protein-level labels, BioBlobs recovers experimentally annotated catalytic sites in the M-CSA database, demonstrating unsupervised functional substructure discovery and opening a path to large-scale functional site discovery across the unannotated proteome.

Figures

Figures reproduced from arXiv: 2510.01632 by Carlos Oliver, Kaiwen Shi, Xin Wang.

Figure 1
Figure 1. Figure 1: Overview of the BIOBLOBS pipeline. The framework consists of four main components: a protein encoder, a neural partitioner, a blob codebook, and a global-blob attention fusion module. The GVP encoder first processes the protein graph and produces residue embeddings. (a) Neural Blob Partitioner. A seed residue is first selected with Gumbel–Softmax. Its k-hop neighborhood is then identified to restrict the c… view at source ↗
Figure 2
Figure 2. Figure 2: UMAP projection of the blob and code embeddings for the EC(structure) test set, where [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Neural partitioner case study: tuning maximum cluster size [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: UMAP projection of code and cluster embeddings for EC dataset, random split [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: UMAP projection of code and cluster embeddings for EC dataset, structure split [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Structural Interpretations of Protein Language Model Representations via Differentiable Graph Partitioning

    cs.LG 2026-05 unverdicted novelty 7.0

    SoftBlobGIN combines ESM-2 representations with protein contact graphs via a lightweight GNN and differentiable substructure pooling to achieve 92.8% accuracy on enzyme classification, raise binding-site AUROC to 0.98...

  2. PaperMentor: A Human-Centered Multi-Agent Writing Tutor for AI Research Papers on Overleaf

    cs.CL 2026-06 unverdicted novelty 4.0

    A multi-agent writing tutor for Overleaf that uses 12 agents and an expert skill library to generate inline comments, with a 14-user study reporting 90.6% actionable and 67.5% valid comments that outperform a GPT-5.2 ...

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · cited by 2 Pith papers · 2 internal anchors

  1. [1]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

    Yoshua Bengio, Nicholas L ´eonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation.arXiv preprint arXiv:1308.3432,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2401.14819 , year=

    Dexiong Chen, Philip Hartout, Paolo Pellizzoni, Carlos Oliver, and Karsten Borgwardt. Endowing protein language models with structural knowledge.arXiv preprint arXiv:2401.14819,

    work page arXiv

  3. [3]

    Romanos Fasoulis, Georgios Paliouras, and Lydia E Kavraki

    doi: 10.1016/S0968-0004(98)01254-7. Romanos Fasoulis, Georgios Paliouras, and Lydia E Kavraki. Graph representation learning for structural proteomics.Emerging Topics in Life Sciences, 5(6):789–802,

    work page doi:10.1016/s0968-0004(98)01254-7

  4. [4]

    Simplicial attention networks.arXiv preprint arXiv:2204.09455,

    Christopher Wei Jin Goh, Cristian Bodnar, and Pietro Lio. Simplicial attention networks.arXiv preprint arXiv:2204.09455,

    work page arXiv

  5. [5]

    Categorical Reparameterization with Gumbel-Softmax

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2009.01411 , year=

    URLhttp://arxiv. org/abs/2009.01411. arXiv:2009.01411 [q-bio]. 10 John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin ˇZ´ıdek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold.nature, 596(7873):583–589,

    work page arXiv 2009

  7. [7]

    Georg Kustatscher, Tom Collins, Anne-Claude Gingras, Tiannan Guo, Henning Hermjakob, Trey Ideker, Kathryn S Lilley, Emma Lundberg, Edward M Marcotte, Markus Ralser, et al

    URLhttps://proceedings.neurips.cc/paper_files/paper/2023/hash/ b6167294ed3d6fc61e11e1592ce5cb77-Paper-Datasets_and_Benchmarks.pdf. Georg Kustatscher, Tom Collins, Anne-Claude Gingras, Tiannan Guo, Henning Hermjakob, Trey Ideker, Kathryn S Lilley, Emma Lundberg, Edward M Marcotte, Markus Ralser, et al. Under- studied proteins: opportunities and challenges ...

    work page 2023

  8. [8]

    doi: 10.1016/j.jmb.2003.08

    work page doi:10.1016/j.jmb.2003.08 2003

  9. [9]

    Graph pooling for graph neural networks: Progress, challenges, and opportunities.arXiv preprint arXiv:2204.07321,

    Chuang Liu, Yibing Zhan, Jia Wu, Chang Li, Bo Du, Wenbin Hu, Tongliang Liu, and Dacheng Tao. Graph pooling for graph neural networks: Progress, challenges, and opportunities.arXiv preprint arXiv:2204.07321,

    work page arXiv

  10. [10]

    Aaron Van Den Oord, Oriol Vinyals, et al

    doi: 10.1038/250194a0. Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.Advances in neural information processing systems, 30,

    work page doi:10.1038/250194a0

  11. [11]

    arXiv preprint arXiv:2509.03885 , year=

    Zhiyu Wang, Arian Jamasb, Mustafa Hajij, Alex Morehead, Luke Braithwaite, and Pietro Li `o. Topotein: Topological deep learning for protein representation learning.arXiv preprint arXiv:2509.03885,

    work page arXiv

  12. [12]

    A survey on protein representation learning: Retrospect and prospect.arXiv preprint arXiv:2301.00813,

    11 Lirong Wu, Yufei Huang, Haitao Lin, and Stan Z Li. A survey on protein representation learning: Retrospect and prospect.arXiv preprint arXiv:2301.00813,

    work page arXiv

  13. [13]

    arXiv preprint arXiv:2203.06125 , year=

    Zuobai Zhang, Minghao Xu, Arian Jamasb, Vijil Chenthamarakshan, Aurelie Lozano, Payel Das, and Jian Tang. Protein representation learning by geometric structure pretraining.arXiv preprint arXiv:2203.06125,

    work page arXiv

  14. [14]

    All LLM-produced code and text was thoroughly double-checked

    12 A Appendix Use of Large Language Models (LLMs) LLMs were used to assist in coding, writing, and producing figures. All LLM-produced code and text was thoroughly double-checked. B Reproducibility All code, data, and weights necessary to reproduce results and use our models on new data are avail- able onhttps://github.com/OliverLaboratory/BioBlobs. Bench...

    work page 2023