Threads

Most computational pathology foundation models are patch encoders: they embed a 224- or 512-pixel tile, and anything larger has to be assembled by a separately trained aggregator. That is a poor fit for clinical practice, where the unit of diagnosis is a gigapixel whole-slide image and the cohorts that matter most — treatment response, prognostication in a rare subtype — often contain fewer than a hundred patients.

Threads, released in January 2025 by the Mahmood Lab at Brigham and Women's Hospital, Massachusetts General Hospital, and Harvard Medical School, takes a different route. It is a slide-level encoder that maps a whole-slide image of any size directly to a single 1024-dimensional embedding, so a downstream task becomes a logistic regression rather than a training run. Its distinguishing idea is the supervision signal: rather than aligning slides to pathology reports or to tissue-site labels, Threads is pretrained by contrastive alignment against next-generation sequencing profiles taken from an adjacent section of the same tissue.

The pretraining cohort, MBTG-47K, pairs 47,171 H&E sections across 39 organs with matched transcriptomic or genomic profiles — the largest paired histomolecular corpus assembled for foundation model development. Threads sits alongside the lab's other slide-level model, TITAN, and its patch-level predecessors UNI and CONCH, and is benchmarked against Prov-GigaPath, CHIEF, and PRISM.

#Key Features

• Molecular pretraining supervision: Cross-modal contrastive learning aligns each slide embedding with the embedding of bulk RNA sequencing (MGH, TCGA, GTEx) or a targeted SNV/indel/CNV panel (BWH) drawn from adjacent tissue.
• Universal slide embeddings: A single forward pass produces one vector per slide regardless of slide dimensions, so downstream work is linear probing rather than training a multiple-instance-learning aggregator from scratch.
• Molecular prompting: Averaging encoded molecular profiles per class yields prototypes a query slide can be assigned to by nearest L2 distance — classification without any task-specific training. IDH prompts built on TCGA-GBMLGG and applied to EBRAINS reach 0.960 AUC, against 0.961 for a trained linear probe.
• Label efficiency on rare events: Because only a linear head is fit, tasks with small cohorts remain tractable; Threads also serves as an initialization for full fine-tuning when labels are plentiful.

#Technical Details

Threads pairs a frozen CONCHv1.5 patch encoder — a ViT-L (307M vision parameters) initialized from UNI and fine-tuned on 1.17 million image-caption pairs — with a two-head gated attention-based multiple instance learning slide encoder of 11.3M parameters that projects 768-dimensional patch features to a 1024-dimensional slide embedding. Expression profiles are encoded by a modified scGPT adapted to bulk RNA (12 transformer blocks, 8 heads, 512-dimensional gene embeddings, initialized from a pancancer checkpoint trained on 5.7 million cells); mutation panels go through a 4-layer MLP with 10.1M parameters. Pretraining ran on 4x80GB A100s for up to 101 epochs, sampling 512 patches per slide, with early stopping on RankMe.

Evaluation covers 54 tasks from 23 cohorts across 17 sources: 8 subtyping and grading, 21 mutation prediction, 12 IHC status, and 13 treatment response and survival tasks. Under linear probing, Threads improves absolute performance by 6.3% over PRISM, 9.9% over Prov-GigaPath, and 6.7% over CHIEF, winning 44/54, 54/54, and 49/54 tasks respectively. It reaches 98.3% AUC on breast cancer subtyping and 91.9% quadratic Cohen's kappa on colorectal grading — within 2.3 points of a dedicated ABMIL model trained on the same data. Ablations show performance peaks at two attention heads and declines by 1.0% at six, and that a transformer slide encoder underperforms the ABMIL design by 6.3%.

#Applications

Threads targets the biomarker and prognostication questions that sit downstream of routine H&E: predicting mutation and IHC status without ordering the assay, subtyping and grading, estimating treatment response, and stratifying survival. Because a new task requires only a logistic regression over precomputed embeddings, it suits translational groups working with retrospective trial cohorts of a few hundred slides — the setting where training a bespoke model is not viable. Molecular prompting extends this to cross-modal slide retrieval and to classification in cohorts where no labeled training slides exist at all.

#Impact

Threads reframes what a pathology foundation model should be supervised on. Report text and tissue-site labels describe how a slide was categorized; sequencing describes what the tissue is made of, and the paper's data-scaling curves show that this signal is more efficient per training sample than the far larger report-supervised and self-supervised corpora used by PRISM, Prov-GigaPath, and CHIEF. The paper remains a preprint awaiting peer review, and the model is not yet distributed: the lab's TRIDENT toolkit lists Threads among its supported slide encoders with weights pending, and no model card or public checkpoint has been posted. Reported gains are also confined to retrospective, in-silico benchmarks — no prospective clinical evaluation is reported.