You can deploy this model and start earning money today!

IshaChhikara/
cx-uni-abmil-7cohort

Name: cx-uni-abmil-7cohort
Availability: InStock

chromosomal-instability

copy-number-signatures

attention-mil

image-classification

mit

Views

No views yet

cx-uni-abmil-7cohort

A Gated Attention-MIL head that predicts the 17 Drews 2022 chromosomal-instability signatures (CX1–CX17) from whole-slide H&E pathology images, using frozen MahmoodLab/UNI tile embeddings as input.

This is a 5-fold ensemble trained on 289 TCGA patients across 7 cohorts.

Headline results (5-fold cancer-type-stratified CV)

Pan-cohort Spearman ρ, mean ± std across folds, on 13 TCGA-active signatures:

Signature	ρ ± std	95% bootstrap CI	q (BH)	Tag
CX3	+0.384 ± 0.096	[+0.255, +0.459]	<0.001	★ H1
CX1	+0.366 ± 0.052	[+0.246, +0.456]	<0.001	★ H1
CX12	+0.359 ± 0.048	[+0.248, +0.452]	<0.001	★ H1
CX6	+0.271 ± 0.046	[+0.146, +0.372]	<0.001
CX10	+0.236 ± 0.112	[+0.098, +0.322]	<0.001
CX17	+0.205 ± 0.142	[+0.040, +0.251]	<0.05
CX2	+0.193 ± 0.125	[+0.032, +0.268]	<0.05
CX5	+0.189 ± 0.067	[+0.028, +0.247]	<0.05
CX4	+0.171 ± 0.135	[+0.034, +0.262]	<0.05
CX16	+0.149 ± 0.174	[-0.003, +0.227]	n.s.
CX9	+0.125 ± 0.193	[-0.068, +0.147]	n.s.
CX11	+0.112 ± 0.076	[-0.064, +0.165]	n.s.
CX8	+0.087 ± 0.092	[-0.006, +0.216]	n.s.
CX7, CX13–15	n/a	inactive in TCGA	—

Mean active ρ = +0.219 across 13 active signatures
3 H1 hits (ρ ≥ 0.30)
5 signatures at q < 0.001 (CX1, CX3, CX6, CX10, CX12)
9 signatures at q < 0.05
Patient-level bootstrap with n_boot = 10,000

Architecture

WSI -> tile (224 px @ 0.5 mpp, Macenko-normalized)
   -> UNI (frozen) -> (N tiles, 1024)
   -> GatedABMIL (this model) -> (17,) CX activity prediction

GatedABMIL is the gated-attention multiple-instance-learning head from Ilse et al. (ICML 2018), with hidden = 512, dropout = 0.3, multi-task regression onto log1p-then-z-scored CX activities. Trained with AdamW (lr = 1e-4) and Multi-Task Huber loss. The same 5 fold checkpoints are used as an inference ensemble.

Quick start

python

import torch
from huggingface_hub import snapshot_download
from modeling_cx_abmil import CXEnsemble

# Download model weights + code
local = snapshot_download(repo_id="IshaChhikara/cx-uni-abmil-7cohort")

# Load 5-fold ensemble (CPU is fine; the head is tiny)
model = CXEnsemble.from_pretrained(local, device="cpu")

# Get UNI embeddings for your WSI tiles (see cx-histology repo for tiling)
tile_embeddings = torch.load("my_slide_uni_embeddings.pt")  # (N, 1024)
tile_embeddings = tile_embeddings.unsqueeze(0)              # (1, N, 1024)

# Predict 17 CX activities
cx = model.predict(tile_embeddings)
print(cx.shape)  # torch.Size([1, 17])

Intended use

Research only. Not for clinical use.
Predicting CIN signature activities from histology images of solid tumours in TCGA-style cohorts.
Feature analysis, attention rollout, comparative studies.

Limitations and important caveats

1. Pan-cohort ρ partly reflects cancer-type classification

Within-cohort Spearman ρ (recomputed per cancer type) is much smaller than the pan-cohort number:

Cohort	n	Within-cohort mean ρ	Within-cohort max ρ
CHOL	32	+0.122	+0.337
KICH	9	+0.082	+0.707 (small n!)
READ	80	+0.003	+0.189
MESO	58	+0.001	+0.202
DLBC	31	-0.063	+0.169
UVM	24	-0.061	+0.196
UCS	55	+0.037	+0.302

Pan-cancer cohorts differ in mean CIN burden by ~10×; some of the pan-cohort ρ comes from the model classifying cancer type and inferring its mean CIN. The honest within-biology signal is weaker.

2. Scaling caveat

A larger 15-cohort version (IshaChhikara/cx-uni-abmil-15cohort, n = 1,454) shows the small-cohort effect sizes do not preserve at scale — mean active ρ drops to +0.13 with the same architecture. The 7-cohort numbers are statistically robust but partly an artefact of cohort selection. Use the 15-cohort model if you need a more conservative estimate.

3. Calibration

Slope of regression(predicted_log1p, true_log1p) is far below 1 (0.05–0.16 across signatures). Predictions are rank-correct but heavily shrunken toward the mean. For "rank patients by predicted CX1" this is fine; for "estimate patient X's CX1 magnitude" use post-hoc isotonic calibration.

4. Inactive signatures

CX7, CX13, CX14, CX15 are essentially zero in TCGA copy-number profiles (≥ 93% of patients zero). Predictions for these signatures are emitted but should not be interpreted; the training pipeline auto-masked them via per-fold variance threshold. The predict_dict() method hides them by default.

5. UNI dependency and license

The encoder (MahmoodLab/UNI) is gated on Hugging Face. You must request access from the authors and agree to their license terms separately. We do not redistribute UNI weights.

Training data

289 TCGA patients with both diagnostic FFPE H&E slides (open-access via GDC) and CX1–CX17 activity labels computed from ABSOLUTE copy-number calls via the CINSignatureQuantification Bioconductor package (Drews 2022 algorithm).
7 TCGA cohorts: DLBC, CHOL, UVM, READ, MESO, UCS, KICH.
5-fold cancer-type-stratified CV, seed = 20260419.

Methods overview

Macenko stain normalization of 224 × 224 H&E tiles at 0.5 µm/px, tissue-fraction-min = 0.5, Laplacian-var-min = 100.
Frozen UNI ViT-L/16 forward → (N tiles × 1024) embeddings per slide.
GatedABMIL head: bag-level pooled prediction of (17,) CX activities.
Loss: Multi-Task Huber on log1p-then-z-scored labels, with per-fold variance-threshold mask to exclude TCGA-inactive signatures.
AdamW (lr = 1e-4, wd = 1e-5), early stopping on val loss.

Full training code, ablations, and bootstrap analysis at github.com/ishaschhikara316/cx-histology.

Citation

This work is in preparation. If you use this model, please cite:

bibtex

@misc{cxuni-abmil-7cohort-2026,
  title  = {Pan-cancer prediction of CIN signatures from H\&E with a frozen UNI foundation model},
  author = {Chhikara, Isha},
  year   = {2026},
  note   = {In preparation. Code: github.com/ishaschhikara316/cx-histology},
}

Please also cite the upstream sources you depend on:

Drews et al., A pan-cancer compendium of chromosomal instability, Nature 606, 976–983 (2022).
Chen et al., Towards a general-purpose foundation model for computational pathology, Nature Medicine (2024) — UNI.
Ilse et al., Attention-based deep multiple instance learning, ICML (2018).
Macenko et al., A method for normalizing histology slides for quantitative analysis, ISBI (2009).

Files

modeling_cx_abmil.py — model class + ensemble wrapper
inference_example.py — minimal usage example
config.json — architecture + training metadata
fold0.pt … fold4.pt — 5 fold checkpoints (state_dict + per-fold label mu/sd)
LICENSE — MIT