Skip to content

Prov-GigaPath

Family: path-fm Modality: H&E whole-slide image tiles (256×256 at 20×, resized to 224 for the ViT) plus a slide-level encoder Released: 2024 (Nature) License: Code Apache-2.0; model weights gated, non-commercial research only (Microsoft Research / Providence terms) Code/checkpoint: github.com/prov-gigapath/prov-gigapath; checkpoint on HuggingFace (prov-gigapath/prov-gigapath) Also surfaced in: single-cell-spatial-omics, virtual-cells

Prov-GigaPath is a whole-slide pathology foundation model from Microsoft Research and Providence Health that pairs a tile-level encoder with a slide-level aggregator, so a single forward pass can turn an entire gigapixel slide into a fixed-length representation. It is pretrained on more than 1.3 billion tiles from over 170,000 real-world H&E slides drawn from 31 Providence-affiliated hospitals, making it one of the largest published pathology corpora to date. The Nature 2024 paper claims state-of-the-art results across cancer subtyping, mutation prediction, and pathology-genomics retrieval.

Architecture & training

The tile encoder is a ViT-style backbone trained with DINOv2-style self-supervision on roughly 1.3 billion 256×256 H&E patches sampled from 171,189 whole-slide images. The slide-level encoder uses a LongNet-based dilated-attention transformer that ingests the full sequence of tile tokens for a slide (tens of thousands per WSI) and is pretrained with a masked tile-modeling objective, producing a single slide embedding without needing weakly supervised MIL. The public checkpoint comprises both the tile and slide encoders; the authors report improved performance over UNI, CTransPath, and other pathology backbones across 26 downstream clinical and biomarker tasks.

Use in the AACR 2026 corpus

Posters mentioning Prov-GigaPath (n=2):

  • 1457 — Prediction of gene expression and molecular pathway activity from H&E whole slide images in non-small cell lung cancer …ted internal dataset of 67 commercial non-small cell lung cancer (NSCLC) patient samples with matched RNA-seq.Gigapath, a large-scale pathology foundation model pre-trained on over 170,000 WSIs, was used to extract patch and pat…
  • 1470 — Selecting representative histologic sections for cost-efficient 3D spatial transcriptomics and tumor microenvironment reconstruction …mune aggregates, and mucosa—were segmented using a graph neural network. For each section, we extracted Prov-GigaPath embeddings and tissue class area proportions. Pairwise similarities, incorporating histology features and z-d…

Sessions mentioning Prov-GigaPath (n=1):

What's missing / where evidence is weak

  • Only the tile encoder is being used. Both poster mentions (1457 NSCLC gene expression; 1470 3D spatial-transcriptomics tile selection) extract patch / tile embeddings — neither exercises the LongNet slide-level encoder that is Prov-GigaPath's actual architectural contribution.
  • No head-to-head benchmark in the corpus. None of the posters compare Prov-GigaPath against UNI / UNI2-h or CONCH on a shared task; the choice of backbone reads as convenience, not the result of evaluation.
  • The session mention is glancing. The April-20 session reference is a name-drop ("using gigapath to … model pathology slide") with no method detail, so it adds no evidence beyond the posters.
  • No discussion of the gated weights or non-commercial terms in any corpus mention, despite this being a real adoption barrier for industry-adjacent work.

Takeaway

Prov-GigaPath shows up as a working tile-feature backbone in two specialized methods posters — gene-expression prediction from H&E in NSCLC, and tile-similarity-based section selection for 3D spatial transcriptomics — plus one passing session reference. The AACR 2026 corpus thus uses Prov-GigaPath the same way it uses UNI: as a frozen patch encoder feeding a downstream model. The slide-level LongNet encoder, which is the paper's main novelty, never appears in the corpus, suggesting that even where Prov-GigaPath is preferred over UNI, it is not yet preferred because of its slide-level capability.

FM comparison & 2026 status

Where it sits in the FM landscape. Prov-GigaPath is the two-stage (tile + slide) encoder in the pathology FM family — a ViT-style DINOv2 tile encoder paired with a LongNet dilated-attention slide encoder, which is its architectural differentiator from the encoder-only UNI / UNI2-h / Virchow line. At ~1.1B parameters total (the LongNet slide encoder adds substantial weight on top of the ViT tile backbone) it sits between UNI2-h (681M) and Virchow2G (1.85B). The 1.3B-tile / 171K-WSI Providence corpus is one of the three largest pathology pretraining datasets, alongside Paige's 3.1M-WSI Virchow2 corpus and the Mahmood Lab's 350K-WSI UNI2 corpus.

Direct peers.

Model Architecture Pretraining corpus Public weights Headline benchmark
Prov-GigaPath DINOv2 ViT tile + LongNet slide encoder 1.3B tiles, 171K WSIs (Providence) Apache-2.0 code, gated HF weights, non-commercial 26-task SOTA over UNI/CTransPath (Nature 2024); top tumor-type classification (2026 benchmark)
Virchow DINOv2 ViT-H, 632M params 1.5M WSIs (Paige real-world) CC-BY-NC, HF gated Pan-cancer detection, rare cancers (Nature Medicine 2024)
Virchow2 / Virchow2G DINOv2 ViT-H (632M) / ViT-G (1.85B) 3.1B / 1.9B tiles, 3.1M WSIs CC-BY-NC, HF gated Highest overall CPTAC 7-task average (2026 clinical benchmark)
Hibou-B / Hibou-L DINOv2 ViT-B (85.7M) / ViT-L 1M+ WSIs (HistAI) Apache-2.0 open Patch classification + batch-effect correction
PathChat / PathChat+ Vision-language LLM (UNI/CONCH-based encoder + LLM) 1M+ visual-language instructions Demo only, non-commercial Diagnostic Q&A across pathology specialties (Nature 2024 + arXiv 2025)

What changed in 2025-2026. Prov-GigaPath itself has not had a v2 release — the model card on prov-gigapath/prov-gigapath retained the original Nature 2024 weights through 2025-2026, gated behind a HuggingFace contact-information form. Two things shifted in 2025-2026: (1) the clinical benchmark frontier moved — Campanella et al. (Nat Commun 2025) and follow-up CPTAC-based studies put Virchow2 first, Prov-GigaPath and UNI tied for second, reordering the 2024 Nature paper's "SOTA" claim; (2) Virchow2G at 1.85B parameters demonstrated that further scaling beats Prov-GigaPath on tile-level tasks, putting pressure on the LongNet slide-level architecture to justify its compute cost. The non-commercial / Microsoft-Providence terms remained an adoption barrier for industry-adjacent groups, more restrictive than Hibou's Apache-2.0 and comparable to UNI2-h's CC-BY-NC-ND.

AACR-relevant use cases. The corpus shows three live use modes: (1) H&E-to-RNA prediction in NSCLC — poster 1457 uses Gigapath embeddings as the tile-feature backbone for predicting matched RNA-seq gene expression and pathway activity from H&E in commercial NSCLC samples, the most "biomarker-discovery" use case in the corpus; (2) 3D-ST section selection — poster 1470 uses Prov-GigaPath embeddings + tissue-class area proportions to compute pairwise WSI similarity for cost-efficient 3D spatial transcriptomics, a methodologically novel application that exercises tile-level features at scale; (3) Pathology TME modeling — the April 20 AM Decoding Cancer Ecosystem session name-checks "gigapath" in the context of modeling pathology slides for tumor microenvironment heterogeneity, though without methodological detail. Conspicuously absent: no AACR 2026 poster exercises the LongNet slide-level encoder, which is Prov-GigaPath's actual architectural novelty over UNI — every corpus use is a tile-feature extraction that UNI2-h could equally serve. The "virtual cell" framing (the topic the dossier is also surfaced in) does not yet connect Prov-GigaPath to single-cell or spatial transcriptomic FMs in any 2026 abstract.