Pith. sign in

REVIEW 2 major objections 1 minor 39 references

CALM aligns brain imaging and genetics data from completely disjoint populations to recover interpretable biomarker associations that generalize to paired subjects.

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

2026-07-03 17:21 UTC pith:2GWY3KLS

load-bearing objection CALM aligns unpaired imaging and genetics via linear projections and claims generalization to paired data, but the abstract gives almost no evidence that the alignment captures real biology rather than artifacts. the 2 major comments →

arxiv 2607.01656 v1 pith:2GWY3KLS submitted 2026-07-02 cs.LG

CALM: Interpretable Cross-Modal Alignment for Biomarker Discovery from Unpaired Data

Jueqi Wang , Zachary Jacokes , John Darrell Van Horn , Kevin A. Pelphrey , Michael C. Schatz , Archana Venkataraman This is my paper
classification cs.LG
keywords cross-modal alignmentunpaired databiomarker discoveryneuroimaginggeneticsautism spectrum disorderinterpretable machine learninglatent space alignment
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 introduces a method that projects unpaired neuroimaging and genetic data into a shared latent space where class-conditional distributions match and diagnostic groups remain separable. Linear projection coefficients then serve as direct, readable links between brain regions of interest and genetic pathways. This matters for disorders like autism because most available repositories contain only one modality, so recovering cross-modal signals without paired subjects could multiply the usable data volume. Experiments show the learned associations remain stable when compared to a paired baseline and transfer to an unseen paired test set with higher accuracy than competing alignment techniques.

Core claim

Linear projections that simultaneously match class-conditional latent distributions and preserve group separability produce stable, interpretable pathway-ROI associations from completely unpaired imaging and genetics populations; these associations generalize to held-out paired data and recover immune and metabolic links to cortical regions that match existing autism literature.

What carries the argument

The CALM alignment uses linear projections to match class-conditional latent distributions while enforcing group separability, with the projection weights serving as the interpretable cross-modal associations.

Load-bearing premise

That matching class-conditional distributions through linear projections on unpaired populations will recover the same associations that exist when the same populations are actually paired.

What would settle it

Train CALM on unpaired autism imaging and genetics cohorts, then compare the resulting pathway-ROI weights against weights obtained by training an identical model directly on paired subjects from the same cohorts; substantial mismatch in the recovered associations or failure to generalize on the paired test set would falsify the central claim.

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

If this is right

  • The projection weights directly yield pathway-ROI associations without post-hoc interpretation steps.
  • Associations learned on unpaired data remain stable relative to a paired baseline.
  • On autism data the method identifies immune and metabolic pathways linked to specific cortical regions, consistent with prior findings.
  • The framework outperforms several state-of-the-art alignment methods and ablation variants on generalization to an unseen paired dataset.

Where Pith is reading between the lines

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

  • If the linear-projection assumption holds, the same alignment strategy could be applied to other unpaired multimodal biomedical repositories such as proteomics and structural imaging.
  • Combining multiple independent unimodal collections could increase effective sample size for cross-modal biomarker searches without requiring new paired acquisitions.
  • The stability result suggests the recovered associations reflect population-level structure rather than pairing artifacts, opening the possibility of meta-analysis across many disjoint cohorts.

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 / 1 minor

Summary. The paper proposes CALM, a framework that learns interpretable cross-modal associations between brain ROIs and genetic pathways from completely unpaired unimodal imaging and genetics datasets. It aligns modalities in a shared latent space using linear projections that match class-conditional distributions and ensure group separability; the resulting projections are claimed to generalize to an unseen paired dataset, outperforming SOTA methods and ablations while yielding stable, literature-consistent associations for autism spectrum disorder.

Significance. If the generalization result holds, the work would be significant for biomarker discovery in neuropsychiatric disorders by enabling use of abundant unimodal repositories where paired data are scarce. The explicit focus on linear, interpretable projections is a constructive design choice that could facilitate biological validation.

major comments (2)
  1. [Abstract] Abstract: the central generalization claim ("generalizes to an unseen paired dataset, outperforming several state-of-the-art methods and ablation baselines") is stated without any equations, dataset sizes, statistical tests, or ablation details, rendering the soundness of the core claim impossible to evaluate from the provided text.
  2. [Methods] Methods (class-conditional alignment objective): the load-bearing assumption that linear projections matching class-conditional latent distributions learned on completely disjoint populations will recover associations that hold on paired data is not accompanied by a direct quantitative check (e.g., subject-level alignment error on the paired hold-out or ablation removing the class-conditional term), leaving open the possibility that alignment succeeds via marginal class statistics or dataset artifacts rather than shared biological signal.
minor comments (1)
  1. [Abstract] Abstract: the SOTA methods and ablation baselines are referred to generically; naming them would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We address each major comment below with point-by-point responses and indicate where revisions will be made to strengthen the presentation and validation of our claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central generalization claim ("generalizes to an unseen paired dataset, outperforming several state-of-the-art methods and ablation baselines") is stated without any equations, dataset sizes, statistical tests, or ablation details, rendering the soundness of the core claim impossible to evaluate from the provided text.

    Authors: We agree that the abstract, in its current concise form, does not provide sufficient quantitative context for readers to fully evaluate the generalization claim. In the revised manuscript, we will expand the abstract to include key supporting details such as the sizes of the unimodal training sets and the paired evaluation set, the primary performance metrics with associated statistical tests, and a brief characterization of the ablation baselines. These additions will be made while remaining within standard abstract length limits. revision: yes

  2. Referee: [Methods] Methods (class-conditional alignment objective): the load-bearing assumption that linear projections matching class-conditional latent distributions learned on completely disjoint populations will recover associations that hold on paired data is not accompanied by a direct quantitative check (e.g., subject-level alignment error on the paired hold-out or ablation removing the class-conditional term), leaving open the possibility that alignment succeeds via marginal class statistics or dataset artifacts rather than shared biological signal.

    Authors: The referee correctly identifies that a more targeted ablation isolating the class-conditional term would provide stronger direct evidence. Our current validation relies on the out-of-distribution generalization to the unseen paired dataset (where marginal or artifact-driven alignments would be unlikely to succeed) together with the reported stability of associations relative to a fully paired baseline. Nevertheless, we will add the requested ablation (removing the class-conditional term) and report subject-level alignment error on the paired hold-out set in the revised Methods and Results sections to directly quantify its contribution and rule out alternative explanations. revision: partial

Circularity Check

0 steps flagged

No circularity: generalization performance measured on independent paired hold-out dataset

full rationale

The paper trains CALM exclusively on disjoint unimodal imaging and genetics datasets using class-conditional distribution matching via linear projections. It then evaluates generalization, stability, and biomarker recovery on a completely unseen paired dataset, with comparisons to external SOTA methods and ablations. No equation or claim reduces the reported associations or performance metrics to quantities defined by the training objective itself; the paired test set supplies an external benchmark. No self-citation chains or fitted-input-as-prediction patterns appear in the derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract alone supplies no concrete free parameters, axioms, or invented entities; the method is described at a high level as using linear projections and distribution matching.

pith-pipeline@v0.9.1-grok · 5716 in / 960 out tokens · 33506 ms · 2026-07-03T17:21:41.153443+00:00 · methodology

0 comments
read the original abstract

The interaction between brain structure and genetic influences is key to understanding neuropsychiatric disorders. However, most large-scale datasets are unimodal, providing either neuroimaging or genetics data. We propose CALM, a framework that learns interpretable associations between brain ROIs and genetic pathways from completely disjoint populations. CALM aligns the two modalities in a shared latent space via linear projections that simultaneously match the class-conditional latent distributions and ensure group separability. These projections provide interpretable pathway--ROI associations. When trained on unimodal imaging and genetics datasets, CALM generalizes to an unseen paired dataset, outperforming several state-of-the-art methods and ablation baselines. We also demonstrate stability of the learned associations against a paired baseline. Our experiments on autism spectrum disorder reveal immune and metabolic pathways linked to specific cortical regions and are consistent with established literature. Thus, CALM opens the door to leveraging large unimodal repositories for studying cross-modal interactions in brain disorders across disparate datasets.

Figures

Figures reproduced from arXiv: 2607.01656 by Archana Venkataraman, John Darrell Van Horn, Jueqi Wang, Kevin A. Pelphrey, Michael C. Schatz, Zachary Jacokes.

Figure 1
Figure 1. Figure 1: Overview CALM. Given unpaired MRI and genetics data, pretrained encoders (EI , EG) output item-specific representations that are projected into a shared latent space via the learned linear maps WI and WG. The loss function simultaneously matches the within-class distributions across modalities, while separating diagnostic groups. The cross-modal association matrix A = W⊤ I WG quantifies the association bet… view at source ↗
Figure 2
Figure 2. Figure 2: Validation of learned pathway-ROI associations. (a) Correlation between elements of the mean association matrix A¯ learned by CALM using unpaired data and when learned using a paired baseline. (b) Cross-fold stability measured by the survival function of signal-to-noise ratio (SNR) across five folds. Genetics: SSC and ACE participants were genotyped on different Illumina chips. Quality control was performe… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Bipartite graph showing the top pathway-ROI associations; edge thickness encodes association strength. (b) Gradient-based attribution of each GWAS disorder channel to the shared latent space, normalized to sum to 100%. Association Validation: To validate that CALM learns pathway–ROI asso￾ciations consistent with those obtained from true paired data, we replace the class-conditional distribution matchin… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 2 internal anchors

  1. [1]

    Nature526(7571), 68 (2015)

    1000 Genomes Project Consortium: A global reference for human genetic variation. Nature526(7571), 68 (2015)

    work page 2015

  2. [2]

    Ad- vances in neural information processing systems26(2013)

    Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Ad- vances in neural information processing systems26(2013)

    work page 2013

  3. [3]

    Nature genetics 55(2), 198–208 (2023)

    Demontis, D., et al.: Genome-wide analyses of ADHD identify 27 risk loci, refine the genetic architecture and implicate several cognitive domains. Nature genetics 55(2), 198–208 (2023)

    work page 2023

  4. [4]

    Molecular psychiatry 19(6), 659–667 (2014)

    Di Martino, A., et al.: The autism brain imaging data exchange: towards a large- scale evaluation of the intrinsic brain architecture in autism. Molecular psychiatry 19(6), 659–667 (2014)

    work page 2014

  5. [5]

    Annals of human genetics51(3), 227–233 (1987)

    Falk, C., et al.: Haplotype relative risks: an easy reliable way to construct a proper control sample for risk calculations. Annals of human genetics51(3), 227–233 (1987)

    work page 1987

  6. [6]

    Neuron68(2), 192–195 (2010)

    Fischbach, G.D., et al.: The simons simplex collection: a resource for identification of autism genetic risk factors. Neuron68(2), 192–195 (2010)

    work page 2010

  7. [7]

    Neuroimage62(2), 774–781 (2012)

    Fischl, B.: FreeSurfer. Neuroimage62(2), 774–781 (2012)

    work page 2012

  8. [8]

    In: Medical Imaging 2021: Image Processing

    Ghosal, S., et al.: G-MIND: an end-to-end multimodal imaging-genetics framework for biomarker identification and disease classification. In: Medical Imaging 2021: Image Processing. vol. 11596, pp. 63–72. SPIE (2021)

    work page 2021

  9. [9]

    In: International confer- ence on learning representations (2022)

    Ghosal, S., et al.: A biologically interpretable graph convolutional network to link genetic risk pathways and imaging phenotypes of disease. In: International confer- ence on learning representations (2022)

    work page 2022

  10. [10]

    The journal of machine learning research13(1), 723–773 (2012) 10 J

    Gretton, A., et al.: A kernel two-sample test. The journal of machine learning research13(1), 723–773 (2012) 10 J. Wang et al

    work page 2012

  11. [11]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Hoshen, Y., et al.: Unsupervised correlation analysis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3319–3328 (2018)

    work page 2018

  12. [12]

    Frontiers in cellular neuroscience12, 405 (2018)

    Hughes, H.K., et al.: Immune dysfunction and autoimmunity as pathological mech- anisms in autism spectrum disorders. Frontiers in cellular neuroscience12, 405 (2018)

    work page 2018

  13. [13]

    Brain144(6), 1911–1926 (2021)

    Jack, A., et al.: A neurogenetic analysis of female autism. Brain144(6), 1911–1926 (2021)

    work page 1911

  14. [14]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Jaume, G., et al.: Modeling dense multimodal interactions between biological path- ways and histology for survival prediction. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 11579–11590 (2024)

    work page 2024

  15. [15]

    Biostatistics8(1), 118–127 (2007)

    Johnson, W.E., et al.: Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics8(1), 118–127 (2007)

    work page 2007

  16. [16]

    Nucleic acids research28(1), 27–30 (2000)

    Kanehisa, M., et al.: KEGG: kyoto encyclopedia of genes and genomes. Nucleic acids research28(1), 27–30 (2000)

    work page 2000

  17. [17]

    Advances in neural information processing systems33, 18661–18673 (2020)

    Khosla, P., et al.: Supervised contrastive learning. Advances in neural information processing systems33, 18661–18673 (2020)

    work page 2020

  18. [18]

    In: 2023 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI)

    Lee, S., et al.: Integrating genetic information for early alzheimer’s diagnosis through mri interpretation. In: 2023 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI). pp. 1–6 (2023)

    work page 2023

  19. [19]

    Cerebral Cortex33(9), 5264–5275 (2023)

    Li, W., et al.: Brainnetome atlas of preadolescent children based on anatomical connectivity profiles. Cerebral Cortex33(9), 5264–5275 (2023)

    work page 2023

  20. [20]

    Li, Y., et al.: AlignMamba: Enhancing multimodal Mamba with local and global cross-modalalignment.In:ProceedingsoftheComputerVisionandPatternRecog- nition Conference. pp. 24774–24784 (2025)

    work page 2025

  21. [21]

    Nature Genetics57(7), 1611–1619 (2025)

    Litman, A., et al.: Decomposition of phenotypic heterogeneity in autism reveals underlying genetic programs. Nature Genetics57(7), 1611–1619 (2025)

    work page 2025

  22. [22]

    International journal of methods in psy- chiatric research27(2), e1608 (2018)

    Marees, A.T., et al.: A tutorial on conducting genome-wide association studies: Quality control and statistical analysis. International journal of methods in psy- chiatric research27(2), e1608 (2018)

    work page 2018

  23. [23]

    Human brain mapping33(6), 1470–1489 (2012)

    Nickl-Jockschat, T., et al.: Brain structure anomalies in autism spectrum disor- der—a meta-analysis of VBM studies using anatomic likelihood estimation. Human brain mapping33(6), 1470–1489 (2012)

    work page 2012

  24. [24]

    Na- ture genetics54(4), 437–449 (2022)

    Okbay, A., et al.: Polygenic prediction of educational attainment within and be- tween families from genome-wide association analyses in 3 million individuals. Na- ture genetics54(4), 437–449 (2022)

    work page 2022

  25. [25]

    Representation Learning with Contrastive Predictive Coding

    Oord, A.v.d., et al.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  26. [26]

    Molecular psychiatry23(1), 6–14 (2018)

    Pedersen, C.B., et al.: The iPSYCH2012 case–cohort sample: new directions for unravelling genetic and environmental architectures of severe mental disorders. Molecular psychiatry23(1), 6–14 (2018)

    work page 2018

  27. [27]

    Journal of Child Psychology and Psychiatry52(6), 631–644 (2011)

    Pelphrey, K.A., et al.: Research review: constraining heterogeneity: the social brain and its development in autism spectrum disorder. Journal of Child Psychology and Psychiatry52(6), 631–644 (2011)

    work page 2011

  28. [28]

    The American journal of human genetics81(3), 559–575 (2007)

    Purcell, S., et al.: PLINK: a tool set for whole-genome association and population- based linkage analyses. The American journal of human genetics81(3), 559–575 (2007)

    work page 2007

  29. [29]

    Nature genetics50(7), 912–919 (2018) CALM: Interpretable Cross-Modal Alignment 11

    Savage, J.E., et al.: Genome-wide association meta-analysis in 269,867 individuals identifies new genetic and functional links to intelligence. Nature genetics50(7), 912–919 (2018) CALM: Interpretable Cross-Modal Alignment 11

    work page 2018

  30. [30]

    Neuroimage53(3), 1051–1063 (2010)

    Shen, L., et al.: Whole genome association study of brain-wide imaging phenotypes for identifying quantitative trait loci in mci and ad: A study of the adni cohort. Neuroimage53(3), 1051–1063 (2010)

    work page 2010

  31. [31]

    Not Just a Black Box: Learning Important Features Through Propagating Activation Differences

    Shrikumar, A., et al.: Not just a black box: Learning important features through propagating activation differences. arXiv preprint arXiv:1605.01713 (2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  32. [32]

    Bioinformatics36(11), 3537–3548 (2020)

    Smedley, N.F., et al.: Discovering and interpreting transcriptomic drivers of imag- ing traits using neural networks. Bioinformatics36(11), 3537–3548 (2020)

    work page 2020

  33. [33]

    Nature genetics 55(9), 1471–1482 (2023)

    Stevelink, R., et al.: GWAS meta-analysis of over 29,000 people with epilepsy identifies 26 risk loci and subtype-specific genetic architecture. Nature genetics 55(9), 1471–1482 (2023)

    work page 2023

  34. [34]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Wang, J., et al.: Learning explainable imaging-genetics associations related to a neurological disorder. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 347–357. Springer (2025)

    work page 2025

  35. [35]

    Nature genetics53(9), 1276–1282 (2021)

    Wightman, D.P., et al.: A genome-wide association study with 1,126,563 individu- als identifies new risk loci for alzheimer’s disease. Nature genetics53(9), 1276–1282 (2021)

    work page 2021

  36. [36]

    Nature genetics50(5), 668–681 (2018)

    Wray, N.R., et al.: Genome-wide association analyses identify 44 risk variants and refine the genetic architecture of major depression. Nature genetics50(5), 668–681 (2018)

    work page 2018

  37. [37]

    Advances in Neural Information Processing Systems (2025)

    Yacobi, A., et al.: Learning shared representations from unpaired data. Advances in Neural Information Processing Systems (2025)

    work page 2025

  38. [38]

    Advances in Neural Infor- mation Processing Systems36, 54716–54726 (2023)

    Zhang, Y., et al.: Learning unseen modality interaction. Advances in Neural Infor- mation Processing Systems36, 54716–54726 (2023)

    work page 2023

  39. [39]

    Nature genetics54(9), 1305–1319 (2022)

    Zhou, X., et al.: Integrating de novo and inherited variants in 42,607 autism cases identifies mutations in new moderate-risk genes. Nature genetics54(9), 1305–1319 (2022)

    work page 2022