MORPHEUS

MORPHEUS (Masked Omics modeling for multimodal RePresentation learning across Histopathology and molEcUlar profileS) is a self-supervised pretraining strategy that learns a single shared representation of tumour tissue morphology and the molecular profiles measured from the same patient. It was developed at Institut Claudius Regaud with IRT Saint Exupéry and the Cancer Research Center of Toulouse.

Self-supervised pathology encoders such as UNI and Prov-GigaPath extract rich morphological features from whole-slide images (WSIs), but morphology alone misses molecular signal that resides in transcriptomics, DNA methylation, and copy-number data. Previous attempts to combine the two either treated omics as supervision for a slide encoder — aligning WSI embeddings to omics embeddings with a contrastive loss — or trained supervised late-fusion models with a separate encoder per modality. Neither yields a general-purpose multimodal encoder.

MORPHEUS instead treats each omics profile as a full input modality. Every modality is tokenized into a shared representation space, a large fraction of the omics tokens is masked, and one transformer encoder must reconstruct the missing content from the surviving omics tokens plus the WSI. The pretrained encoder can then be applied to histopathology alone or to any subset of the omics modalities, and — by choosing masks at inference — it can reconstruct any omics profile from any input combination that includes a slide.

#Key Features

• Masked omics modeling: 75% of omics tokens are masked during pretraining and reconstructed by a shared multimodal encoder with one lightweight decoder per modality, forcing the model to learn cross-modal structure rather than modality-specific shortcuts.
• Biologically structured tokenization: RNA features are grouped into the 50 MSigDB hallmark pathways, while methylation and copy-number features are grouped by chromosomal position into 50 clusters each, so tokens correspond to interpretable biological units.
• Compact slide representation: A Perceiver module compresses the thousands of patch embeddings in a slide into 16 learned prototype tokens, making joint attention over histology and omics tractable.
• Flexible modality subsets: One checkpoint serves WSI-only, WSI+RNA, and fully multimodal downstream tasks, matching clinical reality where omics availability varies per patient.
• Any-to-any omics reconstruction: Missing RNA, methylation, or copy-number profiles can be imputed from a slide alone or conditioned on whichever omics modalities are available.
• Modest compute: Pretraining runs for 100 epochs with a batch size of 128 on a single NVIDIA RTX 3090 Ti.

#Technical Details

Pretraining uses 4,718 TCGA cases spanning 32 cancer types for which WSI, RNA, methylation, and copy-number data are all available; every downstream evaluation is run on patients held out from that cohort. Slides are tiled at 20x magnification and embedded with UNI2-h, with 1,024 patch embeddings sampled per case during training. Two variants are reported: a single-block encoder and decoders (MORPHEUS) and a two-block variant (MORPHEUS-2L).

On five WSI-only subtyping tasks — glioma IDH status, ATRX mutation, 1p/19q codeletion, breast IDC versus ILC, and lung LUAD versus LUSC — MORPHEUS averages 92.9% AUC against 87.4% for the strongest baseline, a 6.3% relative gain over methods including ABMIL, TransMIL, and a contrastive omics-guided slide encoder trained on identical features. Gains are largest where the tasks are hardest: in the 10-shot setting it improves AUC by 11.1%, 11.0%, and 12.0% on ATRX, IDH, and 1p/19q over the best supervised baselines. For reconstruction, median Pearson correlation between predicted and true pan-cancer RNA features is 0.653 from histology alone and 0.739 when methylation and copy number are also supplied, and methylation predicted from histology in the glioma cohort reaches a median correlation of 0.459.

#Applications

The encoder supports molecular biomarker prediction directly from routine H&E slides, which matters most where the reference assay is slow or unavailable: WHO 2021 glioma classification depends on IDH, ATRX, and 1p/19q status, and MORPHEUS estimates these from morphology alone. It also serves survival prognosis under varying modality availability, and its reconstruction mode lets groups with slides but no sequencing budget impute transcriptomic, epigenomic, and copy-number profiles for exploratory analysis and cohort triage.

#Impact

MORPHEUS shows that masked modeling — the objective behind most language and vision pretraining — transfers cleanly to paired histology and multi-omics data, and that a single shared encoder can absorb modalities as heterogeneous as gigapixel images, pathway-level expression, and CpG methylation. Two constraints bound its current reach. Pretrained weights are not distributed: the model consumes patch embeddings from UNI, whose licence prevents redistribution, so the authors ship code and step-by-step reproduction instructions under a CC BY-NC 4.0 licence instead of a checkpoint. And pretraining draws on a single cohort of a few thousand cases, orders of magnitude smaller than the slide corpora behind pathology-only foundation models, which leaves headroom for scaling and makes the survival results noisier on small per-cancer subsets. The work is a preprint and has not been peer reviewed.