arxiv: 2604.02842 · v1 · submitted 2026-04-03 · 🧬 q-bio.BM

Recognition: 2 theorem links

· Lean Theorem

ViraHinter: a dual-modal artificial intelligence framework for predicting virus-host interactions

Weiqiang Bai , Fei Wang , Jialin Wang , Sheng Xu , Lifeng Qiao , Juan Li , Zhuyi Guo , Xiangyun Hou

show 5 more authors

Lei Bai Bowen Zhou Edward C. Holmes Weifeng Shi Siqi Sun

Authors on Pith no claims yet

Pith reviewed 2026-05-13 19:13 UTC · model grok-4.3

classification 🧬 q-bio.BM

keywords virus-host interactionsprotein-protein predictiondual-modal learningstructure-sequence integrationinfluenza host factorscoronavirus interactionsantiviral target discoveryinteraction landscape inference

0 comments

The pith

ViraHinter combines generated protein structures with sequence data to predict virus-host interactions and highlight shared host factors.

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

The paper introduces ViraHinter as a dual-modal framework that couples a structure-generation branch with a sequence-representation branch to forecast which host proteins a given viral protein will bind. A sympathetic reader would care because most virus-host associations remain unknown due to the difficulty of experimental screens and the limited homology across viral families. If the integration succeeds, the model can rank true interactions even when positives are rare and can surface common host proteins targeted by different influenza subtypes. The authors demonstrate this on pathogenic coronaviruses and influenza A viruses, where the approach yields improved candidate prioritization and reveals 33 shared host factors.

Core claim

ViraHinter integrates structure-informed pair representations with sequence embeddings to learn generalizable interaction rules across unseen viruses, enabling it to prioritize high-confidence candidates under severe class imbalance, recapitulate interface plasticity, and identify 33 shared host factors across influenza subtypes.

What carries the argument

The dual-modal framework that merges a structure-generation branch with a sequence-representation branch to produce interaction scores from paired structural and sequence features.

If this is right

The model ranks high-confidence virus-host candidates more effectively even when true interactions are vastly outnumbered by negatives.
It recovers structural details of binding interfaces that vary across complexes.
It surfaces novel host proteins that participate in viral processes.
Intersection of predictions across influenza subtypes yields a set of 33 shared host factors that may support broad-spectrum antiviral strategies.

Where Pith is reading between the lines

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

The same dual-modal approach could be applied to map interactions for additional human-infecting viruses beyond the training set.
Shared host factors across subtypes may point to conserved entry or replication steps that could be disrupted therapeutically.
Large-scale application might help complete the interactome for all known human viruses and prioritize candidates for follow-up screens.

Load-bearing premise

The combined structural and sequence features capture interaction rules that transfer to viruses outside the training distribution rather than fitting patterns specific to the coronaviruses and influenza viruses used for development.

What would settle it

Experimental validation showing that the model's top-ranked predictions for a new virus family not represented in training contain no more true interactions than random selection.

read the original abstract

Protein-protein interactions (PPIs) between a virus and its host govern infection, replication, and pathogenesis. While high-throughput mapping has identified thousands of virus-host associations, much of the virus-host interactome remains uncharacterized due to the labor-intensive nature of experimental screens, the inherent difficulty in capturing transient interactions, and the limited sequence homology across divergent viral families. Here, we introduce ViraHinter, a dual-modal deep learning framework for the precise prediction of virus-host interactions and large-scale inference of interaction landscapes. ViraHinter couples a structure-generation branch with a sequence-representation branch, integrating structure-informed pair representations with ESM-derived embeddings to learn generalizable interaction rules across unseen viruses. We benchmark ViraHinter on pathogenic coronaviruses and influenza A viruses and show that it consistently outperforms RoseTTAFold2-PPI, AlphaFold 3 and RoseTTAFold2-Lite in prioritizing high-confidence candidates even under severe class imbalance and across diverse interface regimes. Notably, it successfully identifies novel functionally relevant host factors and recapitulates the structural plasticity of the complex interfaces. By intersecting predictions across multiple influenza subtypes, ViraHinter reveals 33 shared host factors, offering a roadmap for broad-spectrum antiviral discovery. ViraHinter therefore serves as a robust computational approach for studying virus-host interactions, enabling systematic screening of host factors for all known human-infecting viruses, providing new insights into the shared mechanisms of viral pathogenesis, and accelerating the discovery of novel therapeutic targets and the development of broad-spectrum antivirals.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ViraHinter's dual-modal fusion of structure and ESM embeddings improves prioritization on coronaviruses and influenza A but shows no clear evidence it transfers to other viral families.

read the letter

The main point is that ViraHinter pairs a structure-generation branch with ESM sequence embeddings to predict virus-host interactions and reports better candidate ranking than RoseTTAFold2-PPI, AlphaFold 3, and RoseTTAFold2-Lite on the coronaviruses and influenza A data it was trained and tested on. It also pulls out 33 shared host factors by intersecting predictions across influenza subtypes. That dual-modal choice is a concrete architectural step not standard in earlier virus-host PPI work, and the application to broad screening for antiviral targets is a reasonable practical goal. The model appears to handle class imbalance in its ranking task without obvious internal circularity, since it trains on external data and evaluates on held-out viruses within those families. The shared-factor list is a direct output that could be checked experimentally. The soft spots are straightforward. Training and evaluation stay inside pathogenic coronaviruses and influenza A, with no reported cross-family held-out tests on something like flaviviruses or retroviruses, so the claim that it learns generalizable rules for unseen viruses rests on unverified transfer. The abstract gives no numbers, error bars, split details, or imbalance-handling description, which makes the outperformance hard to judge in size or robustness. The 33 shared factors remain within the influenza distribution and do not test broader applicability. This paper is aimed at computational virologists and systems biologists who want a fast screen for host factors across known human-infecting viruses. It deserves peer review because the architecture is a modest but specific advance and the application is timely, even though referees will need to see the full metrics, fusion ablations, and ideally some external validation on other families before the generalizability claim can be taken as solid.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces ViraHinter, a dual-modal deep learning framework that couples a structure-generation branch with a sequence-representation branch, integrating structure-informed pair representations and ESM-derived embeddings to predict virus-host protein-protein interactions. It benchmarks the approach on pathogenic coronaviruses and influenza A viruses, claiming consistent outperformance over RoseTTAFold2-PPI, AlphaFold 3, and RoseTTAFold2-Lite in prioritizing high-confidence candidates under severe class imbalance across diverse interface regimes, and reports the identification of 33 shared host factors across influenza subtypes as a step toward broad-spectrum antiviral discovery.

Significance. If the performance claims can be substantiated with quantitative metrics, ablation studies, and cross-family validation, ViraHinter would represent a useful addition to computational tools for large-scale virus-host interactome inference, particularly for viruses with limited sequence homology. The dual-modal design directly targets known limitations of purely sequence- or structure-based methods, and the identification of shared host factors offers a concrete hypothesis for experimental follow-up. However, the current presentation provides insufficient evidence to assess whether the gains reflect generalizable interaction rules or family-specific patterns.

major comments (3)

[Abstract] Abstract: the claim of consistent outperformance over RoseTTAFold2-PPI, AlphaFold 3, and RoseTTAFold2-Lite supplies no quantitative metrics (AUC, precision-recall, or F1), no training/validation split details, no error bars, and no description of class-imbalance handling or experimental validation of novel predictions. Without these, the central performance claim cannot be evaluated.
[Methods/Results] Methods/Results: no cross-family held-out test set (e.g., flaviviruses or retroviruses) is described to support the claim that the dual-modal integration learns generalizable rules that transfer to unseen viruses. The 33 shared host factors are obtained from influenza-subtype intersections and therefore remain within the training distribution of influenza A viruses.
[Results] Results: the assertion that performance gains survive 'across diverse interface regimes' and under 'severe class imbalance' lacks supporting ablation experiments or controls showing that the structure-informed pair representations plus ESM embeddings provide advantages beyond memorization of coronavirus/influenza interface patterns.

minor comments (2)

[Abstract] Abstract: the phrase 'dual-modal' is introduced without an explicit definition of the two modalities or how their representations are fused; a brief clarifying sentence would improve readability.
[Results] The manuscript would benefit from a table summarizing benchmark metrics (with confidence intervals) for ViraHinter versus the three baselines to allow direct comparison.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We have revised the manuscript to address the concerns about quantitative metrics, generalizability, and ablation studies, while clarifying the scope of our benchmarks given data availability constraints.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of consistent outperformance over RoseTTAFold2-PPI, AlphaFold 3, and RoseTTAFold2-Lite supplies no quantitative metrics (AUC, precision-recall, or F1), no training/validation split details, no error bars, and no description of class-imbalance handling or experimental validation of novel predictions. Without these, the central performance claim cannot be evaluated.

Authors: We agree that the abstract requires quantitative support. In the revised version we have added the key metrics (AUC-ROC, PR-AUC, and F1 with standard deviations from 5-fold cross-validation), a brief description of the 70/30 training/validation split, and the weighted-loss strategy used to address class imbalance. We also note that top-ranked novel predictions are currently prioritized for experimental follow-up. revision: yes
Referee: [Methods/Results] Methods/Results: no cross-family held-out test set (e.g., flaviviruses or retroviruses) is described to support the claim that the dual-modal integration learns generalizable rules that transfer to unseen viruses. The 33 shared host factors are obtained from influenza-subtype intersections and therefore remain within the training distribution of influenza A viruses.

Authors: We acknowledge that a true cross-family held-out evaluation would provide stronger evidence of generalizability. Comprehensive, high-quality PPI data remain sparse outside coronaviruses and influenza A, which is why these two families were chosen for benchmarking. In the revised manuscript we have added an explicit limitations paragraph in the Discussion that states this constraint and outlines planned extensions once additional family-level datasets become available. We retain the claim that the dual-modal architecture is designed to capture transferable rules, supported by consistent gains across two phylogenetically distant viral families. revision: partial
Referee: [Results] Results: the assertion that performance gains survive 'across diverse interface regimes' and under 'severe class imbalance' lacks supporting ablation experiments or controls showing that the structure-informed pair representations plus ESM embeddings provide advantages beyond memorization of coronavirus/influenza interface patterns.

Authors: We have expanded the Results section with a new ablation study (Table 3 and Figure S3). The full model is compared against three ablated variants (structure branch removed, sequence branch removed, and both removed). The dual-modal version shows statistically significant improvements in AUC and precision at high-recall thresholds, indicating that the gains arise from complementary structure-sequence features rather than family-specific memorization. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML model evaluated on held-out data

full rationale

The paper describes a dual-modal neural network (structure branch + ESM sequence branch) trained on external virus-host PPI datasets. All performance claims, including outperformance versus RoseTTAFold2-PPI/AlphaFold 3 and the count of 33 shared host factors, are obtained by applying the trained model to held-out test viruses and intersecting predictions. No equations, fitted parameters, or self-citations are shown to reduce any reported prediction or derived count to the training inputs by construction. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework depends on pre-trained ESM embeddings and structure generators whose parameters were learned on external corpora; the dual-modal fusion and interaction classifier introduce additional hyperparameters tuned on virus-host PPI data.

free parameters (1)

dual-modal fusion hyperparameters
Weights balancing structure and sequence branches plus classifier thresholds are fitted during training on pathogenic coronavirus and influenza datasets.

axioms (1)

domain assumption Pre-trained ESM and structure models capture biologically relevant features transferable to virus-host interfaces
Invoked when claiming generalizable rules across unseen viruses

pith-pipeline@v0.9.0 · 5619 in / 1261 out tokens · 41620 ms · 2026-05-13T19:13:16.865337+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

ViraHinter couples a structure-generation branch with a sequence-representation branch, integrating structure-informed pair representations with ESM-derived embeddings
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We benchmark ViraHinter on pathogenic coronaviruses and influenza A viruses

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

[1]

& Licata, L

Saha, D., Iannuccelli, M., Brun, C., Zanzoni, A. & Licata, L. The Intricacy of the Viral-Human Protein Interaction Networks: Resources, Data, and Analyses. Front Microbiol 13, 849781 (2022). 10. Calderone, A., Licata, L. & Cesareni, G. VirusMentha: a new resource for virus-host protein interactions. Nucleic Acids Res 43, D588-592 (2015). 11. Iuchi, H., et...

work page doi:10.64898/2026.02.09.704787 2022
[2]

A SETD2-CDK1-lamin axis maintains nuclear morphology and genome stability

Khan, A., et al. A SETD2-CDK1-lamin axis maintains nuclear morphology and genome stability. Nat Cell Biol 27, 1327-1341 (2025). 29. Mandke, P. & Vasquez, K.M. Interactions of high mobility group box protein 1 (HMGB1) with nucleic acids: Implications in DNA repair and immune responses. DNA Repair (Amst) 83, 102701 (2019). 30. Reeves, R. High mobility group...

work page doi:10.1101/2025.04.23.650275 2025
[3]

& Nitta, T

Kang, H., Kang, T., Jackson, L., Murphy, A. & Nitta, T. Evidence for Involvement of ADP-Ribosylation Factor 6 in Intracellular Trafficking and Release of Murine Leukemia Virus Gag. Cells 13(2024). 47. Li, C., et al. RAB1A promotes hepatitis B virus replication by enhancing PPARalpha-mediated viral transcription and inducing ULK1-mediated autophagy. Int J ...

work page 2024