Biology subjects

Romesser, P.

Publications and source records attributed to Romesser, P..

4 recordsLinked to original sources

bioRxiv2026

Beyond Segmentation Accuracy: Evaluating Biological Fidelity in Spatial Pathology

Multiplex immunofluorescence (MxIF) resolves cellular identity, state, and spatial organization within tumor tissue, enabling higher-order representations of cancer ecosystems including immune infiltration, multicellular neighborhoods, and tissue niches. Yet these analyses depend on a deceptively simple first step: accurately identifying individual cells. Segmentation errors can alter marker quantification, cellular phenotypes, cell-cell relationships, and ultimately the spatial organization inferred from an image. Consequently, improvements in conventional segmentation benchmarks matter biologically only if they propagate to more faithful representations of the tumor microenvironment. We test this relationship directly using an expert-curated benchmark of 31,924 expert-reviewed cells on three slides under a leave-slide-out protocol. We compare a domain-specific self-supervised encoder, adapted to multiplex tissue imaging, with supervised, microscopy-specific, prompted generalist, and general-purpose vision models, and pass every method's predicted cells through an identical downstream pipeline. We evaluate fidelity across four levels: cell segmentation, phenotype assignment, spatial neighborhood structure, and tissue organization, comprising niches and tumor-immune interactions. The resulting hierarchy reveals that these levels do not improve in lockstep. Cellular phenotype assignment is comparatively robust to segmentation differences, whereas missed cells increasingly disrupt neighborhood structure and tumor-immune interactions. As a stringent spatial-interaction test, tumor-CD8+ contact recovery is poorly predicted by conventional segmentation quality (Panoptic Quality: {rho} = 0.18), while errors specifically at tumor-immune interfaces are strongly associated with contact fidelity ({rho} = -0.80). Methods with nearly identical segmentation accuracy can therefore produce substantially different biological representations. These results argue that cell segmentation for spatial pathology should be evaluated not only by boundary and instance accuracy, but also by preservation of the biological relationships that downstream analyses are intended to measure.

bioinformatics↗

bioRxiv2026

Discovering Latent Scientific Concepts through Discrete Representation Learning

Scientific imaging is undergoing a fundamental transformation. Advances in multiplex immunofluorescence, computational pathology, and electron microscopy now enable observation of biological and physical systems at unprecedented spatial resolution. Rather than simply classifying known objects, these data offer the opportunity to discover previously unknown organizational principles. However, scientific image datasets are typically highly curated, information-rich, and limited in size, making reliable representation learning particularly challenging. We show that standard continuous self-supervised adaptation can become unstable in this regime, allowing teacher representations to drift and degrading downstream transfer. We introduce Kodiak, a discrete representation learning architecture that discovers latent scientific concepts through Sinkhorn-balanced codebook assignments. Instead of learning unconstrained continuous embeddings, Kodiak constructs a reusable vocabulary of discrete concepts that provides stable supervision for both masked-patch prediction and cross-view representation learning within a single-stage framework. Across five scientific imaging domains, including multiplex immunofluorescence for pancreatic cancer pathology and scanning electron microscopy, Kodiak consistently improves representation quality under both continued foundation-model adaptation and training from scratch. Beyond improved classification and transfer,the learned concepts are interpretable, reusable, and produce spatially coherent concept maps that support applications such as cell phenotyping, segmentation-free patch phenotyping, spatial niche discovery, tissue organization, and materials characterization. We view Kodiak as the representation-learning engine of our foundation model for spatial biology. More broadly, our results suggest that discrete representation learning provides a principled foundation for discovering reusable latent scientific concepts that can accelerate scientific interpretation across imaging domains.

cancer biology↗

bioRxiv2026

A Paracrine Dietary Lipid Axis Constrains Antitumor Immunity in Liver Cancer

Overnutrition-related liver dysfunction and cancer are increasingly prevalent and highly resistant to immunotherapy. While metabolic dysregulation is a hallmark of hepatocellular carcinoma (HCC), how nutrient overload impairs antitumor immunity remains unclear. Here, we show that short-term Western diet (WD) exposure drives near-complete loss of CD8 T cell infiltration and antitumor function in HCC. We identify dietary linoleic acid (LA), the most abundant {omega}-6 fatty acid, as the dominant immunosuppressive driver. Cancer cell-restricted FADS2-mediated desaturation of LA to longer-chain {omega}-6 PUFAs drives their accumulation in the tumor interstitial fluid, suppressing infiltrating CD8 T cells via lipid peroxidation. FADS2 inhibition restores CD8 T cell function and sensitizes WD-driven HCC to PD-1-based immunotherapy. Further, the Parkinsons disease-associated deglycase DJ-1 protects LA-handling proteins from methylglyoxal-mediated glycation, sustaining tumoral immunosuppressive PUFA production. Across multiple independent human MASLD-HCC cohorts, LA metabolic activity correlates with CD8 T cell impairment, immune exclusion, and immunotherapy resistance. Overall, these studies identify a dietary lipid axis as a therapeutically actionable vulnerability in WD-associated HCC.

Cancer Biology↗

bioRxiv2025

Integrated histopathologic modeling of detailed tumor subtypes and actionable biomarkers

Accurate cancer subtyping with accompanying molecular characterization is critical for precision oncology. While machine learning approaches have been applied to both digital pathology and cancer genomics, previous work has been limited in sample size and has typically aggregated granular cancer subtypes into coarse groupings, likely obfuscating informative molecular and prognostic associations and phenotypic variation of more detailed tumor subtypes. Accordingly, we collated 378,123 hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) with matched targeted DNA clinical sequencing results and OncoTree detailed cancer subtypes from a real-world cohort of 71,142 patients. Using this scaled, granular dataset and a cancer subtype knowledge graph, we developed Mosaic: a family of calibrated machine learning models using H&E WSI embeddings to classify tumors and identify molecular phenotypes across 163 detailed subtypes. The cancer subtyping module (Aeon) achieved an area under the receiver operating characteristic curve (AUROC) of 0.992 overall, with 161/163 subtypes reaching an AUROC [≥] 0.90 and improved performance over a state-of-the-art genomics-based classifier. The genomic inference module (Paladin) achieved an AUROC [≥] 0.80 for 167 pairs of detailed subtypes and genomic targets. We further used the learned histopathologic representations to i) identify key associations of the histopathologic embeddings with clinical biomarkers; ii) identify unsupervised sub-clusters of tumors with genomic determinants of tumor phenotype; iii) specify granular diagnoses for cancers of unknown primary, evaluated by genomic associations and expected clinical outcome distributions; iv) annotate functional significance for variants of uncertain significance (VUS); and v) identify cases that mimic the phenotypic effect of known DNA variants on H&E in the absence of detectable DNA alterations. Taken together, this work advances our understanding of phenotypic variation of granular tumor subtypes, their relevance to enhanced diagnostics, and their potential utility in risk stratification with multimodal machine learning in cancer.

cancer biology↗