Publications

H-Optimus-1: A foundation model for computational histopathology

Pathology AI usually needs a different model for every task. H-Optimus-1 is one model with 1.1B parameters, trained on one of the largest histology datasets. It is hitting state-of-the-art across major external benchmarks.

A practical guideline for benchmarking unimodal and multimodal foundation models on spatial transcriptomics data in oncology

The benchmarks judging foundation models for spatial omics are too coarse to trust. This infographic poster lays out how to benchmark with transparent baselines, diverse datasets, biologically meaningful tasks, and new oncology-specific metrics. A practical standard for evaluating the next generation of models, presented at the American Association for Cancer Research Annual Meeting 2026.

Anna C. Wahl-Schaar, Dasha Valter, Lisa Sikkema, Stefan G. Stark, Ediem Al-Jibury, Tobias Heinen, Liam Gonzalez, Marin Scalbert, Tokuwa Kanno, Zelda Mariet +

A practical guideline for benchmarking unimodal and multimodal foundation models on spatial transcriptomics data in oncology

The benchmarks judging foundation models for spatial omics are too coarse to trust. This infographic poster lays out how to benchmark with transparent baselines, diverse datasets, biologically meaningful tasks, and new oncology-specific metrics. A practical standard for evaluating the next generation of models, presented at the American Association for Cancer Research Annual Meeting 2026.

Topics:

Bioptimus Research

Marin Scalbert, Charlie Saillard, Thomas Peeters, Liam Gonzalez, Dasha Valter, Felipe Llinares-López, Zelda Mariet, Rodolphe Jenatton

H-Optimus-1: A foundation model for computational histopathology

Pathology AI usually needs a different model for every task. H-Optimus-1 is one model with 1.1B parameters, trained on one of the largest histology datasets. It is hitting state-of-the-art across major external benchmarks.

Model:

H-Optimus

Topics:

Bioptimus Research

GitHub / bioRxiv

Saillard C, Jenatton R, Llinares-López F, Mariet Z, Cahané D, Durand E, Vert JP (Bioptimus)

H-optimus-0: A Foundation Model for Computational Pathology

H-optimus-0 is Bioptimus's inaugural open-source pathology foundation model — a 1.1B-parameter ViT-g/14 trained with self-supervised learning on over 500,000 whole-slide images from 4,000+ clinical centers worldwide. The model achieves state-of-the-art performance across tile- and slide-level benchmarks including cancer subtyping, biomarker prediction, and gene expression estimation. As the foundational release of Bioptimus's model family, H-optimus-0 has exceeded one million downloads and is widely adopted by pharmaceutical and academic researchers.

Read full publication

Model:

H-Optimus

Topics:

Bioptimus Research

Foundation Models for Computational Pathology

Bioengineering (MDPI)

Al-Asi H et al.

Pathology Foundation Models: Evolution, Current Landscape, Challenges and Opportunities from a Technical and Clinical Perspective

This peer-reviewed review surveys the technical and clinical evolution of pathology foundation models, from early task-specific systems to large Vision Transformer architectures trained on millions of slides. H-optimus-0 and H-optimus-1 are cited as leading examples of domain-specific pathology foundation models representing best-in-class scale and performance. The work places Bioptimus at the forefront of a rapidly maturing field with clear clinical translational potential.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Bioptimus Research

NeurIPS 2024 (Spotlight)

HEST-1k: A Dataset for Spatial Transcriptomics and Histology Image Analysis

HEST-1k assembles 1,229 spatial transcriptomic profiles each paired with a whole-slide image and metadata across 26 organs and 25 cancer types, providing a large benchmark for foundation model evaluation in spatial biology. H-optimus-0 achieves top performance on the HEST-Benchmark for predicting spatial gene expression directly from H&E images, ranking first among all evaluated foundation models. The work establishes HEST as the primary benchmark for evaluating pathology models on spatial transcriptomics — a domain central to Bioptimus's M-Optimus and STELA platforms.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Bioptimus Research

Computers in Biology and Medicine

Shvetsov N et al.

A Lightweight and Extensible Cell Segmentation and Classification Model for Whole Slide Images

This paper develops a lightweight, deployable cell segmentation and classification model for whole slide images using H-optimus-0 as a fixed encoder backbone integrated with a U-Net architecture. By cross-relabeling PanNuke and MoNuSAC annotations to create a unified seven-class dataset and distilling the H-optimus-0-based model, the authors achieve a 48× reduction in parameters while maintaining segmentation quality. The distilled model is integrated into QuPath, bringing Bioptimus-quality feature extraction to clinical digital pathology workflows.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Oncology Applications

Computers in Biology and Medicine

Balezo G et al.

MIPHEI-ViT: Multiplex Immunofluorescence Prediction from H&E Images using ViT Foundation Models

MIPHEI-ViT predicts multiplex immunofluorescence signals from routine H&E images using a U-Net architecture with H-optimus-0 as the encoder backbone, enabling cell-type characterization without requiring expensive mIF panels. Trained on the OrionCRC dataset and validated across five independent cohorts, the model achieves strong performance for key immune and stromal markers including T cells, B cells, and epithelial markers. The work demonstrates H-optimus-0's effectiveness as a general-purpose encoder for translating between complementary tissue analysis modalities.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Oncology Applications

ESMO Real World Data and Digital Oncology

Özden GM et al.

End-to-end deep learning for predicting breast cancer recurrence risk

This study presents an end-to-end deep learning framework for predicting breast cancer recurrence risk directly from whole-slide images, benchmarking multiple pathology foundation models in a clinical oncology setting. H-optimus-1 achieves top performance among all evaluated models, providing the strongest predictive signal for recurrence risk stratification. Published in ESMO Real World Data and Digital Oncology, the work delivers high-profile clinical validation of H-optimus-1's translational value in breast cancer prognosis.

Read full publication

Model:

H-Optimus

Topics:

Oncology Applications

Foundation Models for Computational Pathology

arXiv:2509.09923 (preprint)

Toledo Tan MJ, Kapetanaki M, Benos PV

Engineering Spatial and Molecular Features from Cellular Niches to Inform Predictions of Inflammatory Bowel Disease

This paper presents a computational framework that utilizes spatial transcriptomics to distinguish between Crohn's disease, ulcerative colitis, and healthy tissue. The authors use H-optimus-1 to predict spatial gene expression directly from routine H&E-stained slides, enabling more efficient identification of high-value tissue regions for further analysis. This integration provides an accurate and explainable tool for understanding the spatial organization of IBD pathology.

Read full publication

Model:

H-Optimus

Topics:

Spatial Biology & Transcriptomics

Oncology Applications

medRxiv (preprint)

Witali Aswolinskiy et al.

Uncertainty-Aware Prediction of Microsatellite Instability in Colorectal Cancer from H&E-Stained Whole Slide Images

This study develops an uncertainty-aware deep learning pipeline for predicting microsatellite instability (MSI-H) status directly from H&E-stained colorectal cancer whole slide images. H-optimus-0 is used for tissue patch encoding within the prediction framework. The approach advances clinically actionable biomarker prediction by incorporating calibrated uncertainty estimates alongside MSI predictions.

Read full publication

Model:

H-Optimus

Topics:

Biomarker Detection & Treatment Response

Oncology Applications

arXiv (preprint)

Kang M et al. (HKUST)

PathBench: A Comprehensive Comparison Benchmark for Pathology Foundation Models Towards Precision Oncology

PathBench is the most comprehensive pathology foundation model benchmark to date, evaluating 19 models on 15,888 whole-slide images from 8,549 patients across 10 hospitals and 64 diagnosis and prognosis tasks in common cancer types. H-optimus-1 achieves the best overall average ranking score (2.3), ranking first in lung and colorectal cancer tasks, and H-optimus-0 excels in molecular subtyping. This benchmark is the most rigorous public validation of Bioptimus models' superiority across the full clinical spectrum from diagnosis to prognosis.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Oncology Applications

arXiv:2508.04441 (preprint)

Jonas Ammeling et al.

Benchmarking Foundation Models for Mitotic Figure Classification

A comprehensive benchmark evaluating multiple pathology foundation models for mitotic figure classification in histopathology slides. H-optimus-0 and H-optimus-1 rank among the strongest models evaluated across multiple datasets. This work confirms the Bioptimus model family's capability for cell-level detection tasks critical for cancer grading and prognostication.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Oncology Applications

arXiv:2506.09095 (preprint)

Vivien van Veldhuizen et al.

Foundation Models in Medical Imaging: A Review and Outlook

This comprehensive review surveys the landscape of foundation models across medical imaging modalities, with substantial coverage of computational pathology. H-optimus-0 and H-optimus-mini are highlighted as leading examples of domain-specific pathology foundation models representing the state of the art. The work positions Bioptimus firmly within the broader medical AI research community.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Filiot A et al. (Owkin)

Distilling Foundation Models for Robust and Efficient Models in Digital Pathology

This paper demonstrates that H-optimus-0 can serve as an effective teacher model for knowledge distillation, yielding a compact student model (H0-mini) that achieves comparable downstream performance at significantly reduced computational cost. The distilled model achieves 3rd place on the HEST benchmark and outstanding robustness to staining and scanning variation. This work showcases the transferability and practical deployability of Bioptimus's foundation model architecture.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Nature Communications

Campanella G, Chen S, Singh M et al.

A Clinical Benchmark of Public Self-Supervised Pathology Foundation Models

This study establishes a benchmark of publicly available self-supervised pathology foundation models evaluated on six clinically relevant tasks spanning multiple organs and disease types, using slides from three medical centers. H-optimus-0 ranks among the top three models — alongside Prov-GigaPath and SP85M — achieving the highest average AUC (0.785) on biomarker prediction tasks, outperforming UNI and Phikon. The work provides direct evidence of H-optimus-0's competitive standing against all publicly available pathology AI models.

Read full publication

Model:

H-Optimus

Topics:

Foundation Models for Computational Pathology

Oncology Applications

npj Precision Oncology

Breen J, Allen K, Zucker K et al.

A Comprehensive Evaluation of Histopathology Foundation Models for Ovarian Cancer Subtype Classification

This rigorous study evaluates 14 histopathology foundation models for five-class ovarian carcinoma subtype classification, trained on 1,864 whole-slide images and validated across two independent external cohorts. H-optimus-0 achieves the highest balanced accuracy — 89% internal, 97% and 74% external — ranking as the best-performing model overall. The results provide compelling clinical-grade evidence for deploying Bioptimus's models in rare and complex cancer diagnostics.

Read full publication

Model:

H-Optimus

Topics:

Oncology Applications

Foundation Models for Computational Pathology

Pan S, Chen J, Secrier M

Teaching Pathology Foundation Models to Accurately Predict Gene Expression with Parameter Efficient Knowledge Transfer

This paper presents a parameter-efficient fine-tuning approach for adapting pathology foundation models — including H-optimus — to the task of predicting spatial gene expression directly from histology images. The method achieves competitive gene expression prediction performance with minimal additional training cost, making it highly practical for spatial biology applications. The work demonstrates the adaptability of Bioptimus models to the multi-omic data integration central to the M-Optimus and STELA roadmap.

Read full publication

Model:

H-Optimus

Topics:

Spatial Biology & Transcriptomics

Foundation Models for Computational Pathology

Computers in Biology and Medicine

Camillo Saueressig et al.

From Histology to Diagnosis: Leveraging Pathology Foundation Models for Glioma Classification

This paper applies pathology foundation models to glioma classification from H&E-stained whole slide images, evaluating their ability to distinguish tumor grades and molecular subtypes. H-optimus-0 achieves strong classification performance, and knowledge distillation from H-optimus-0 further improves results for this domain-specific brain tumor task. The work demonstrates Bioptimus models' effectiveness for rare and complex neurological tumor diagnostics.

Read full publication

Model:

H-Optimus

Topics:

Oncology Applications

Foundation Models for Computational Pathology