M$^3$-Gen: Interpretable Multimodal Generation of Gene Expression Profiles Using Clinical and Imaging Data

Carlo Sgaravatti; Francesca Pia Panaccione; Marco Venere

REVIEW 4 major objections 6 minor 17 references

M3-Gen generates realistic gene expression profiles from histopathology images and clinical metadata, with attention maps tracing which tissue regions drove the synthesis.

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 · deepseek-v4-flash

2026-08-01 07:42 UTC pith:5X4P644Y

load-bearing objection A real architectural contribution for multimodal gene-expression generation, but the headline interpretability claim is not supported by the current evidence. the 4 major comments →

arxiv 2607.21343 v1 pith:5X4P644Y submitted 2026-07-23 cs.LG cs.AIcs.CV

M³-Gen: Interpretable Multimodal Generation of Gene Expression Profiles Using Clinical and Imaging Data

Francesca Pia Panaccione , Carlo Sgaravatti , Marco Venere This is my paper

classification cs.LG cs.AIcs.CV

keywords generative adversarial networksgene expression generationmultimodal learninghistopathologycontrastive learninginterpretabilityTCGAdata augmentation

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.

M3-Gen tries to show that gene expression profiles—normally costly and privacy-sensitive to obtain—can be generated from two routinely collected inputs: histopathology whole-slide images and clinical metadata. The paper's core claim is that a GAN conditioned on a contrastively aligned embedding of images and text produces synthetic transcriptomes that are statistically realistic and biologically coherent, preserving disease-associated gene and pathway signals. A second claim is that the model is interpretable by design: an attention mechanism over image patches, queried by the clinical text, yields maps of which tissue regions most influenced each generated profile. If these claims hold, synthetic expression data could augment scarce cancer cohorts and support in silico experiments without new wet-lab assays.

Core claim

On its own terms, M3-Gen's discovery is that multimodal conditioning—clinical text embeddings aligned to histopathology patch embeddings via contrastive learning, then fused by a multi-head attention layer where the text acts as query and the patches as keys/values—lets a WGAN-GP synthesize gene expression profiles that are hard to distinguish from real TCGA data and that retain disease-relevant biology. The authors report that classifiers trained only on generated profiles match those trained on real data, that 10x augmentation improves downstream disease classification, and that the top differentially expressed genes and enriched pathways largely overlap between real and generated data for

What carries the argument

The central mechanism is the attention-conditioned WGAN-GP. Patch-level image embeddings (from a pretrained histopathology vision transformer) and a clinical text embedding (from a medical language model) are first aligned in a shared space with an InfoNCE contrastive loss. In the generative model, the text embedding serves as the query attending over image patch embeddings as keys and values; the resulting attention-weighted image embedding is concatenated with Gaussian noise and fed to the generator (and similarly conditions the discriminator). The mean attention weight per patch, averaged across heads, is what the paper calls the interpretable map linking tissue regions to gene expression

Load-bearing premise

The load-bearing premise is that the averaged attention weights reflect biological importance of image patches; the paper offers no external validation (no pathologist annotations, no known-marker correlation, no perturbation test), so if attention is driven by low-level visual artifacts or by the disease label in the text, the interpretability claim loses support.

What would settle it

Correlate per-patch attention weights from the generator with independent ground truth on the same slides—e.g., pathologist tumor annotations or spatial transcriptomics expression of a known marker. If attention hotspots do not overlap true tumor regions, or if randomly permuting the attention weights leaves the generated expression profiles nearly unchanged, the interpretability claim collapses. For the generation claim, a simpler check: if a classifier distinguishing real from generated profiles is trained on the same disease-type contrast and reaches near-perfect accuracy while the DE-gene

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

Share X Bluesky LinkedIn Reddit HN

If this is right

A classifier trained only on M3-Gen-synthesized profiles reaches accuracy and F1 comparable to training on real TCGA data, so synthetic data can stand in for scarce real profiles.
Augmenting real training data with ten synthetic samples per real sample yields small but consistent accuracy gains in disease-type classification.
Disease-associated signal survives generation: for the tested contrast, 89/100 up-regulated and 86/100 down-regulated top genes are shared between real and synthetic data, with 8/10 top enriched pathways in common.
Attention maps computed from the generator's weights provide a traceable link from image patches to generated transcriptomic profiles, enabling patch-level explanation.
The model stays robust under simulated histopathology artifacts (blur, stain variation, tissue folding), with utility metrics essentially unchanged.

Where Pith is reading between the lines

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

A natural stress test the authors did not run: replace the attention weights with random or uniform weights over the same patch embeddings and measure how much generated profiles change. If profiles barely change, the attention map is not doing the explanatory work attributed to it.
Because the clinical text embedding already separates disease types cleanly in the t-SNE visualization, the text modality may be carrying most of the biological signal; an image-only vs text-only comparison on a matched disease subtype would clarify whether attention over patches adds information beyond the disease label.
The interpretability claim could be made testable by comparing attention hotspots against spatial transcriptomics or pathologist-annotated tumor regions on the same slides; that would convert 'inherent explainability' into a measured quantity.
If the robustness results generalize, the practical implication is that image conditioning can tolerate typical clinical slide degradations, which lowers the bar for deployment in retrospective cohorts where slide quality varies.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A real architectural contribution for multimodal gene-expression generation, but the headline interpretability claim is not supported by the current evidence.

read the letter

The core thing to know: this paper builds something genuinely new — a WGAN-GP conditioned on contrastively aligned histopathology and clinical-text embeddings via multi-head attention — and the downstream utility evaluation (TSTR) is a real holdout test. That part is solid. The authors also deserve credit for being upfront about their own related GEMM-GAN work and positioning this paper's contribution as interpretability-focused rather than capacity-focused. The contrastive pretraining and the attention-based fusion are clean and well motivated, and the robustness analysis under image artifacts is a nice extra that most papers in this space skip.

The soft spot is exactly what the stress-test note flags. The attention weights in Eqs. (2)-(3) are computed as a softmax over the text embedding querying image patch keys. Since the text embedding contains the disease type, high attention patches are simply those most similar to the disease label after contrastive pretraining. That is not evidence that those patches causally influenced the generated expression; it is evidence of text-image similarity. Section 3.6 just shows ten slides with hot spots, and there is no pathologist annotation, no known-marker correlation, and no perturbation test. The abstract's 'interpretable by design' claim is therefore not established. Relatedly, the biological-coherence section (§3.5) uses a disease contrast that is explicitly part of the text condition, so the high DE overlap partly reflects conditioning fidelity, not a learned morphology-transcriptome association. The stress-test is right on both counts.

Other issues are smaller but still need addressing: all metrics are means of five runs with no variances or significance tests, making it hard to tell whether the differences between M3-Gen and the mean-fusion baseline are real. The biological validation is a single contrast. Code is promised but not yet released. There is also a stray '??' reference in §3.7. None of these kill the central claim, but they all need fixing.

Who is this for? Researchers working on synthetic transcriptomics or multimodal generation in pathology. The generation and augmentation story is plausible and useful, and the architecture is worth knowing about. The interpretability framing should be dialed back until it is externally validated.

Recommendation: this deserves a proper peer review, not a desk reject. A serious referee could push for the variance analysis, a proper validation of the attention maps (marker correlation or perturbation), and a clear statement that the current maps show text-similarity weighting, not causal influence. With those changes it would be a reasonable contribution.

Referee Report

4 major / 6 minor

Summary. The paper introduces M3-Gen, a WGAN-GP conditioned on histopathology image patches and clinical text embeddings for generating gene expression profiles. Image and text encoders are aligned with an InfoNCE loss; the generator and discriminator use multi-head attention with the text embedding as query and patch embeddings as keys/values. The authors report unsupervised similarity, detectability, downstream disease classification, DE/pathway overlap for one disease contrast, and artifact robustness. The main claimed contributions are realistic generation and intrinsic interpretability via attention maps.

Significance. The generation pipeline is coherent and the utility evaluation uses a genuine TSTR holdout protocol with external real-data benchmarks, which is a strength. The DE overlap and pathway agreement for the single contrast are meaningful if replicated and extended. However, the interpretability claim is not validated beyond qualitative examples, and the comparative claims rest on small differences without variance or significance testing. The paper is a useful contribution to synthetic transcriptomic generation, but the central interpretability and biological-coherence claims need substantial additional evidence before they can be accepted.

major comments (4)

[§2.4, §3.6] The interpretability claim rests on the attention weights of Eqs. (2)–(3). Because the query is the text embedding, which contains disease type, and contrastive pretraining aligns patch embeddings to that text, these weights largely measure text–image similarity rather than an independent causal contribution to the generated profile. The claim that they 'quantify the contribution' of each patch requires external validation: correlation with pathologist annotations, correlation with known marker expression, or a perturbation test (e.g., masking high-attention patches and measuring change in output). Section 3.6 only shows ten unquantified visualizations. Without such validation, the abstract's 'interpretable by design' and 'biologically meaningful explanation' are not established.
[§3.5] Biological coherence is evaluated for a single disease-type contrast (Squamous Cell Neoplasms vs Adenomas and Adenocarcinomas), and the conditioning text explicitly includes the disease type. The observed 89/100 DE overlap may largely reflect conditioning fidelity to the disease label, not a learned morphology–transcriptome association. No ablation reports DE agreement for image-only, text-only, or full conditioning. To support 'functionally meaningful', the authors should report DE and pathway overlap across multiple tumor-type contrasts, and ideally for contrasts not directly encoded in the text condition, or compare with a label-only baseline.
[Tables 1, 3, 5] All comparative metrics are reported as means of 5 runs without standard deviations, confidence intervals, or significance tests. In Table 3, the claimed advantage of M3-Gen over the mean-fusion baseline is small (MLP accuracy 0.883 vs 0.870; RF F1 0.880 vs 0.880), and in Table 2 the TSTR/TRTR differences are 0.001–0.017. These differences may be within run-to-run noise. The authors should report per-run values, standard deviations, and paired significance tests before claiming that attention 'consistently achieves the highest' utility or 'outperforms' baselines.
[§3.4, Table 1] The WGAN-GP baseline is described in the text as 'without conditioning', but Table 1's caption says it is 'conditioned on the disease type'. This ambiguity is important because the comparison is meant to isolate the effect of multimodal attention. If the baseline is disease-type-conditional, the improvement of M3-Gen should be measured against the same class-conditional information; if it is unconditional, the comparison is not a controlled ablation. Please clarify the baseline and, if needed, add a disease-type-conditional WGAN-GP to the ablation.

minor comments (6)

[§3.3, Figure 6] The text says a 'more detailed visualization of the image embeddings' is provided in Figure 6, but Figure 6 shows simulated histopathology artifacts. The referenced figure should be Figure 3.
[Figure 1 caption / §2.2] The caption states Otsu thresholding selects 'patches with a percentage of background pixels higher than 80%', while §2.2 says tiles with more than 20% tissue are kept. The wording should be corrected to avoid implying that background-rich patches are selected.
[Eq. (1)] The cross-entropy terms are written as CE(·) without a definition; please define the operation (e.g., softmax cross-entropy over the batch) or replace it with an explicit formula.
[§3.7] The sentence '?? reports some visualizations of the simulated artifacts' contains an unresolved cross-reference; a figure number should be inserted.
[References] References [6] and [7] are duplicates of the same Lacan et al. item. Please consolidate and verify that the Correlation metric in §3.2 is attributed to the intended source.
[Table 1] The baseline name 'WGAN-GP' in Table 1 should match the description in §3.4; also consider reporting whether the baseline uses class-conditioning via the disease label (see major comment).

Circularity Check

0 steps flagged

No significant circularity — the derivation and benchmarks are self-contained against external real-data evaluations, and self-citations are contextual only.

full rationale

The paper's central generation claim is not circular: the WGAN-GP is conditioned on contrastively aligned image and text embeddings, and the evaluation is performed against external TCGA real data using TSTR classification, distributional metrics, DE overlap with real samples, and Enrichr pathway enrichment. No fitted parameter is renamed as a prediction, and no quantity in the derivation is defined in terms of the target evaluation. The contrastive loss (Eq. 1) aligns modalities to produce conditioning embeddings; the attention weights (Eqs. 2-3) are internal model weights used to condition the generator, so identifying patches with high attention as 'most influential' is true by construction of the architecture, but this is an interpretability description rather than a circularly fitted prediction. The stronger claim that these patches are 'biologically relevant' is under-validated (Section 3.6 provides only ten qualitative maps and no pathologist annotation or perturbation test), but that is an external-validity/correctness concern, not circularity. The DE overlap in Table 4 partly reflects disease-label conditioning, but the comparison is still against held-out real profiles and serves as a conditioning-fidelity check, not a definitional equivalence. Self-citations [11] and [12] are disclosed and are only contextual/parallel-work references; no load-bearing uniqueness theorem or ansatz is imported from them. The incomplete '??' figure reference in Section 3.7 is a formatting artifact, not a circular step. Overall, no significant circularity is present.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical or biological entities; its contribution is an empirical architecture built from standard components (UNI, Clinical ModernBERT, CLIP-style contrastive loss, WGAN-GP, multi-head attention). The central claims rest on domain assumptions about TCGA pairing, the informativeness of frozen encoder features, the adequacy of attention pooling as a slide summary, and the generalization of the one-contrast biological validation.

free parameters (6)

InfoNCE temperature τ = not reported (Eq. 1)
Scales logits in the contrastive loss; chosen by hand and affects alignment sharpness.
Embedding dimension d = 128
Projection dimensionality for image/text embeddings; arbitrary capacity choice (§3.1).
Generator latent dimension = 256
Dimension of Gaussian noise z; chosen in §3.4.
Number of patches N per sample = 16 (training), 32 (t-SNE analysis)
Random subset of patches fed to the encoder; affects the aggregation and conditioning (§2.3, §3.3).
Number of attention heads H = not reported
Multi-head attention in Eqs. (2)-(3); H is an architectural free choice not specified in the paper.
Top-N genes for DE agreement = 100
Threshold chosen for the DE overlap analysis (§3.5); affects the reported Jaccard values.

axioms (6)

domain assumption TCGA pairs of WSI, clinical metadata, and RNA-seq are independent samples with learnable correspondence; the LLM-generated 200-word clinical summaries preserve task-relevant information.
Invoked in §2.2 preprocessing; if the LLM summaries drop discriminative clinical information, conditioning is weakened.
domain assumption Frozen UNI patch embeddings and Clinical ModernBERT text embeddings carry sufficient morphological/clinical signal, and a linear projection suffices to align them.
§2.3; the contrastive step trains only projection layers; the rest of the claim depends on pretrained feature quality.
domain assumption A single attention-weighted sum of patch embeddings is an adequate summary of a WSI for conditioning gene expression generation.
§2.4; the generator receives one multimodal vector; no multiple-instance or spatial modeling beyond mean/attention.
ad hoc to paper Attention weights (Eqs. 2-3) can be interpreted as the contribution of image patches to the generated expression profile (attention-as-explanatoriness).
§2.4, §3.6; assumed without validation against annotations or perturbation; load-bearing for the interpretability claim.
ad hoc to paper The DE/pathway analysis for the single contrast SqCC vs AdenoCa is representative of biological coherence across the twelve tumor types.
§3.5; generalization from one contrast underlies the 'functionally meaningful' claim.
standard math Standard GAN/CLIP theory (WGAN-GP, InfoNCE) is valid and applicable at this data scale.
§2.3-2.4; relies on Gulrajani et al. and Radford et al. as background.

pith-pipeline@v1.3.0-alltime-deepseek · 9710 in / 16255 out tokens · 155232 ms · 2026-08-01T07:42:49.518491+00:00 · methodology

0 comments

Cite this review

Pith. "Pith review of M$^3$-Gen: Interpretable Multimodal Generation of Gene Expression Profiles Using Clinical and Imaging Data." pith.science (2026). https://pith.science/paper/5X4P644Y

@misc{pith2026260721343,
  author       = {Pith},
  title        = {Pith review of: M$^3$-Gen: Interpretable Multimodal Generation of Gene Expression Profiles Using Clinical and Imaging Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5X4P644Y}},
  note         = {Machine review of arXiv:2607.21343}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

Integrating heterogeneous biomedical data, including clinical metadata, histopathology images, and molecular profiles, is crucial for comprehensive disease understanding. However, gene expression data acquisition remains constrained by high costs and privacy concerns, limiting its use in multimodal research and AI-driven applications. We present MultiModal Molecular Generation (M$^3$-Gen), a novel framework for the generation of gene expression profiles by conditioning a Generative Adversarial Network on histopathology images and clinical metadata. M$^3$-Gen learns a unified latent representation from the clinical variables and the images, leveraging contrastive learning, and exploits the embeddings of the two modalities to guide a generative model in producing biologically coherent gene expression profiles. Evaluations on the TCGA dataset demonstrate that M$^3$-Gen generates realistic and functionally meaningful gene expression data. Importantly, by integrating multiple modalities in an attention-based mechanism, M$^3$-Gen provides intrinsic explainability: it allows the identification of which regions of the histopathology images most strongly influenced the generation of specific gene expression profiles, making the model's decisions interpretable by design.

Figures

Figures reproduced from arXiv: 2607.21343 by Carlo Sgaravatti, Francesca Pia Panaccione, Marco Venere.

**Figure 1.** Figure 1: The overall pipeline of M3 -Gen. In the preprocessing step (a), we extract patches from tissue slides using the Otsu thresholding algorithm to select the patches with a percentage of background pixels higher than 80%, and we extract textual descriptions from patients’ metadata with an LLM. We then pre-train an image encoder and a text encoder to align their embeddings with contrastive learning (b). Finally… view at source ↗

**Figure 2.** Figure 2: TSNE visualizations of the embeddings after pre-training with contrastive [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: t-SNE visualization of image embeddings obtained after contrastive pre [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Top enriched pathways for the contrast Squamous Cell Neoplasms vs Adenomas and Adenocarcinomas, computed from the top-100 up-regulated (top row) and down-regulated (bottom row) genes. Results are shown for real data (left) and synthetic profiles generated by M3 -Gen (right). Bar lengths represent pathway significance as − log10 adjusted p-values, highlighting a strong agreement between real and generated d… view at source ↗

**Figure 5.** Figure 5: Attention maps generated by the multi-head attention mechanism for [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

**Figure 6.** Figure 6: Examples of simulated histopathology image artifacts used in the robust [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 2 canonical work pages

[1]

https://huggingface.co/ContactDoctor/Bio-Medical-Llama-3-8B (2024)

Bio-medical: A high-performance biomedical language model. https://huggingface.co/ContactDoctor/Bio-Medical-Llama-3-8B (2024)

2024
[2]

Nature Reviews Bioengineering2(12), 991–1004 (2024)

van Breugel, B., Liu, T., Oglic, D., van der Schaar, M.: Synthetic data in biomedicine via generative artificial intelligence. Nature Reviews Bioengineering2(12), 991–1004 (2024)

2024
[3]

Nature Medicine (2024)

Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F., Jaume, G., Chen, B., Zhang, A., Shao, D., Song, A.H., Shaban, M., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine (2024)

2024
[4]

Advances in neural information processing systems30 (2017)

Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems30 (2017)

2017
[5]

Advances in neural information processing systems32(2019)

Kynkäänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in neural information processing systems32(2019)

2019
[6]

Bioinformatics39(Supplement_1), i111–i120 (2023)

Lacan, A., Sebag, M., Hanczar, B.: GAN-based data augmentation for transcrip- tomics: survey and comparative assessment. Bioinformatics39(Supplement_1), i111–i120 (2023)

2023
[7]

Bioinformatics39(Supplement_1), i111–i120 (06 2023)

Lacan, A., Sebag, M., Hanczar, B.: Gan-based data augmentation for transcrip- tomics: survey and comparative assessment. Bioinformatics39(Supplement_1), i111–i120 (06 2023). https://doi.org/10.1093/bioinformatics/btad239, https://doi. org/10.1093/bioinformatics/btad239

work page doi:10.1093/bioinformatics/btad239 2023
[8]

arXiv preprint arXiv:2504.03964 (2025)

Lee, S.A., Wu, A., Chiang, J.N.: Clinical modernbert: An efficient and long context encoder for biomedical text. arXiv preprint arXiv:2504.03964 (2025)

Pith/arXiv arXiv 2025
[9]

Computers in Biology and Medicine170, 107997 (2024)

Liu, H., Zhang, Y., Luo, J.: Contrastive learning-based histopathological features infer molecular subtypes and clinical outcomes of breast cancer from unannotated whole slide images. Computers in Biology and Medicine170, 107997 (2024)

2024
[10]

Auto- matica11(285-296), 23–27 (1975)

Otsu, N., et al.: A threshold selection method from gray-level histograms. Auto- matica11(285-296), 23–27 (1975)

1975
[11]

Bioengineering12(6), 658 (2025)

Panaccione, F.P., Mongardi, S., Masseroli, M., Pinoli, P.: Biogan: Enhancing transcriptomic data generation with biological knowledge. Bioengineering12(6), 658 (2025)

2025
[12]

In: Rodolà, E., Galasso, F., Masi, I

Panaccione, F.P., Sgaravatti, C., Pinoli, P.: Gemm-gan: A multimodal generative model conditioned on histopathology images and clinical descriptions for gene expression profile generation. In: Rodolà, E., Galasso, F., Masi, I. (eds.) Image Analysis and Processing - ICIAP 2025 Workshops. pp. 393–404. Springer Nature Switzerland, Cham (2026) 16 F.P. Panacci...

2025
[13]

In: International conference on machine learning

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

2021
[14]

Nature communications11(1), 3877 (2020)

Schmauch, B., Romagnoni, A., Pronier, E., Saillard, C., Maillé, P., Calderaro, J., Kamoun, A., Sefta, M., Toldo, S., Zaslavskiy, M., et al.: A deep learning model to predict rna-seq expression of tumours from whole slide images. Nature communications11(1), 3877 (2020)

2020
[15]

Bioinformatics38(3), 730–737 (2022)

Viñas, R., Andrés-Terré, H., Liò, P., Bryson, K.: Adversarial generation of gene expression data. Bioinformatics38(3), 730–737 (2022)

2022
[16]

Journal of Pathology Informatics12(1), 54 (2021)

Wang, N.C., Kaplan, J., Lee, J., Hodgin, J., Udager, A., Rao, A.: Stress testing pathology models with generated artifacts. Journal of Pathology Informatics12(1), 54 (2021). https://doi.org/https://doi.org/10.4103/jpi.jpi_6_21, https://www. sciencedirect.com/science/article/pii/S2153353922001766

work page doi:10.4103/jpi.jpi_6_21 2021
[17]

BioRxiv pp

Zheng, Y., Pizurica, M., Carrillo-Perez, F., Noor, H., Yao, W., Wohlfart, C., Marchal, K., Vladimirova, A., Gevaert, O.: Digital profiling of cancer transcriptomes from histology images with grouped vision attention. BioRxiv pp. 2023–09 (2024)

2023

[1] [1]

https://huggingface.co/ContactDoctor/Bio-Medical-Llama-3-8B (2024)

Bio-medical: A high-performance biomedical language model. https://huggingface.co/ContactDoctor/Bio-Medical-Llama-3-8B (2024)

2024

[2] [2]

Nature Reviews Bioengineering2(12), 991–1004 (2024)

van Breugel, B., Liu, T., Oglic, D., van der Schaar, M.: Synthetic data in biomedicine via generative artificial intelligence. Nature Reviews Bioengineering2(12), 991–1004 (2024)

2024

[3] [3]

Nature Medicine (2024)

Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F., Jaume, G., Chen, B., Zhang, A., Shao, D., Song, A.H., Shaban, M., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine (2024)

2024

[4] [4]

Advances in neural information processing systems30 (2017)

Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems30 (2017)

2017

[5] [5]

Advances in neural information processing systems32(2019)

Kynkäänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in neural information processing systems32(2019)

2019

[6] [6]

Bioinformatics39(Supplement_1), i111–i120 (2023)

Lacan, A., Sebag, M., Hanczar, B.: GAN-based data augmentation for transcrip- tomics: survey and comparative assessment. Bioinformatics39(Supplement_1), i111–i120 (2023)

2023

[7] [7]

Bioinformatics39(Supplement_1), i111–i120 (06 2023)

Lacan, A., Sebag, M., Hanczar, B.: Gan-based data augmentation for transcrip- tomics: survey and comparative assessment. Bioinformatics39(Supplement_1), i111–i120 (06 2023). https://doi.org/10.1093/bioinformatics/btad239, https://doi. org/10.1093/bioinformatics/btad239

work page doi:10.1093/bioinformatics/btad239 2023

[8] [8]

arXiv preprint arXiv:2504.03964 (2025)

Lee, S.A., Wu, A., Chiang, J.N.: Clinical modernbert: An efficient and long context encoder for biomedical text. arXiv preprint arXiv:2504.03964 (2025)

Pith/arXiv arXiv 2025

[9] [9]

Computers in Biology and Medicine170, 107997 (2024)

Liu, H., Zhang, Y., Luo, J.: Contrastive learning-based histopathological features infer molecular subtypes and clinical outcomes of breast cancer from unannotated whole slide images. Computers in Biology and Medicine170, 107997 (2024)

2024

[10] [10]

Auto- matica11(285-296), 23–27 (1975)

Otsu, N., et al.: A threshold selection method from gray-level histograms. Auto- matica11(285-296), 23–27 (1975)

1975

[11] [11]

Bioengineering12(6), 658 (2025)

Panaccione, F.P., Mongardi, S., Masseroli, M., Pinoli, P.: Biogan: Enhancing transcriptomic data generation with biological knowledge. Bioengineering12(6), 658 (2025)

2025

[12] [12]

In: Rodolà, E., Galasso, F., Masi, I

Panaccione, F.P., Sgaravatti, C., Pinoli, P.: Gemm-gan: A multimodal generative model conditioned on histopathology images and clinical descriptions for gene expression profile generation. In: Rodolà, E., Galasso, F., Masi, I. (eds.) Image Analysis and Processing - ICIAP 2025 Workshops. pp. 393–404. Springer Nature Switzerland, Cham (2026) 16 F.P. Panacci...

2025

[13] [13]

In: International conference on machine learning

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

2021

[14] [14]

Nature communications11(1), 3877 (2020)

Schmauch, B., Romagnoni, A., Pronier, E., Saillard, C., Maillé, P., Calderaro, J., Kamoun, A., Sefta, M., Toldo, S., Zaslavskiy, M., et al.: A deep learning model to predict rna-seq expression of tumours from whole slide images. Nature communications11(1), 3877 (2020)

2020

[15] [15]

Bioinformatics38(3), 730–737 (2022)

Viñas, R., Andrés-Terré, H., Liò, P., Bryson, K.: Adversarial generation of gene expression data. Bioinformatics38(3), 730–737 (2022)

2022

[16] [16]

Journal of Pathology Informatics12(1), 54 (2021)

Wang, N.C., Kaplan, J., Lee, J., Hodgin, J., Udager, A., Rao, A.: Stress testing pathology models with generated artifacts. Journal of Pathology Informatics12(1), 54 (2021). https://doi.org/https://doi.org/10.4103/jpi.jpi_6_21, https://www. sciencedirect.com/science/article/pii/S2153353922001766

work page doi:10.4103/jpi.jpi_6_21 2021

[17] [17]

BioRxiv pp

Zheng, Y., Pizurica, M., Carrillo-Perez, F., Noor, H., Yao, W., Wohlfart, C., Marchal, K., Vladimirova, A., Gevaert, O.: Digital profiling of cancer transcriptomes from histology images with grouped vision attention. BioRxiv pp. 2023–09 (2024)

2023