EXAONE Path 2.5

EXAONE Path 2.5 is a multimodal pathology foundation model from LG AI Research that enriches histopathology representations by aligning whole-slide images (WSIs) with multiple layers of molecular data — genomic, epigenetic, and transcriptomic. Rather than learning from imaging alone, it produces an integrated patient representation grounded in tumor biology by jointly modeling histology and multi-omics signals in a shared embedding space. Released as an arXiv preprint in December 2025, it continues the EXAONE Path lineage, following the patch-level EXAONEPath 1.0 and the end-to-end supervised EXAONE Path 2.0.

Computational pathology has converged on large self-supervised image models, but most such models see only pixels and miss the molecular context that drives cancer behavior. EXAONE Path 2.5 addresses this gap by treating histology and multi-omics as complementary views of the same patient and forcing them into alignment during training. The result is an embedding intended to reflect underlying tumor biology more comprehensively than image-only foundation models, while remaining usable from a slide alone at inference time.

The model is positioned for the growing set of clinical and translational tasks where morphology and molecular state interact — biomarker inference, subtyping, and outcome prediction — and is released with open weights on HuggingFace under a non-commercial license.

#Key Features

• Multi-omics alignment: Jointly models histology alongside genomic, epigenetic, and transcriptomic data, aligning all modalities in a unified embedding space so the slide representation carries molecular context.
• Multimodal SigLIP loss: Uses an all-pairwise contrastive objective across the heterogeneous modalities, rather than aligning only pairs, to learn consistent cross-modal representations.
• Fragment-aware rotary positional encoding (F-RoPE): A positional-encoding module that preserves spatial structure and tissue-fragment topology across a whole-slide image during slide-level aggregation.
• Domain-specialized internal encoders: Separate foundation models for WSI and RNA-seq supply biologically grounded embeddings that anchor the multimodal alignment.
• Slide-only inference: After alignment training, the model produces an integrated representation from histology alone, so the multi-omics knowledge is usable without molecular assays at test time.

#Technical Details

EXAONE Path 2.5 uses a two-stage inference pipeline: a patch encoder extracts embeddings from image patches (224×224 pixels), and a slide encoder aggregates patch-level features into a slide representation using spatial coordinates and tissue-contour information, with F-RoPE preserving spatial topology. During training, the multimodal SigLIP loss aligns these slide representations with embeddings from domain-specialized WSI and RNA-seq encoders across genomic, epigenetic, and transcriptomic modalities. The authors report that the approach matches or surpasses state-of-the-art pathology foundation models while using substantially fewer training samples and parameters than existing approaches. It was evaluated against six leading pathology foundation models on two complementary benchmarks: an internal multi-institutional real-world clinical dataset and Patho-Bench, a public suite of 80 tasks. (Exact parameter counts and per-task scores are not disclosed in the available materials.)

#Applications

EXAONE Path 2.5 targets computational pathology workflows where molecular context improves analysis of routine histology, including cancer subtyping, molecular biomarker prediction, and survival/outcome prediction from whole-slide images. Because the multi-omics knowledge is distilled into a slide-only representation, pathologists and translational researchers can apply it to standard H&E slides without requiring paired sequencing at inference, making it suitable for retrospective cohort analysis and as a feature extractor for downstream clinical prediction tasks.

#Impact

EXAONE Path 2.5 is part of a broader shift in computational pathology toward biologically grounded, multimodal foundation models that move beyond image-only self-supervision. By demonstrating competitive performance on the 80-task Patho-Bench suite with reportedly lower data and parameter requirements, it offers an efficiency-oriented alternative to larger image-only models. The open release of weights on HuggingFace lowers the barrier for academic adoption, though the non-commercial license and as-yet-undisclosed training data scale limit direct clinical deployment and full reproducibility.