PathLUPI

Precision oncology increasingly turns on molecular information — driver mutations, expression-based subtypes, prognostic signatures — that is expensive and slow to obtain by sequencing every patient. Predicting those readouts directly from the haematoxylin-and-eosin slides that hospitals already produce is an attractive shortcut, but models trained only on whole-slide images (WSIs) have to discover the link between morphology and genotype from weak, slide-level labels, and they generalize poorly across centres.

PathLUPI attacks this with learning using privileged information (LUPI). Bulk transcriptomic profiles are available for the training cohorts but not at deployment, so instead of treating them as an input, the method treats them as supervision: during training a transcriptomic branch shapes the histology encoder so that its embeddings are anchored to the molecular state of the tissue. At inference the transcriptomic branch is discarded and only the WSI is required. The result is a set of genome-anchored embeddings that carry molecular context a purely visual model would not learn.

The work comes from Cheng Jin, Hao Chen, and collaborators at the Hong Kong University of Science and Technology together with clinical and computational groups in Guangzhou, Macau, Hangzhou, Hong Kong, and Boston, and was posted as a preprint in June 2025. It builds on the patch-level pathology foundation model CONCH rather than replacing it, and is best understood as a training strategy that upgrades an existing histology encoder.

#Key Features

• Privileged transcriptomic supervision: Paired bulk RNA-seq guides representation learning during training and is dropped at test time, so deployment needs nothing beyond a digitized slide.
• Pathway-structured molecular branch: Fifty pathway-specific MLPs, organized around MSigDB Hallmark gene sets, encode the transcriptome so that supervision is aligned to interpretable biological programs rather than to raw gene vectors.
• Region-aware slide encoding: A re-embedding transformer compresses patch features into 50 latent spatial regions before gated attention pooling, keeping slide-level inference tractable on gigapixel images.
• Broad task coverage from one embedding: The same representation supports mutation calling, molecular subtyping, and survival modelling across many cancer types with only lightweight task heads.
• Morphology-genotype interpretability: Attention over the learned embeddings surfaces cellular morphological signatures associated with specific genotypes and pathways, giving pathologists a visual handle on the prediction.

#Technical Details

PathLUPI extracts 512-dimensional patch embeddings with CONCH, re-embeds them with a region-aware transformer, and fuses the histology and transcriptomic branches through shared cross-attention before gated attention pooling. Internal training and validation used 6,427 TCGA cases spanning 13 cancer types (BLCA, BRCA, COAD/READ, ESCA, HNSC, KIRC, GBM/LGG, LIHC, LUAD, LUSC, STAD, SKCM, and UCEC) with matched bulk RNA-seq. Evaluation covered 49 molecular oncology tasks over 11,257 cases from 20 cohorts, including CPTAC, the EBRAINS Digital Tumor Atlas, and two independent breast cancer centres contributing 2,045 and 1,527 WSIs. Across 25 biomarker prediction tasks the mean AUC improved 4.48% over the strongest multiple-instance-learning baseline (4.83% internal, 2.68% external), with AUC ≥ 0.80 on 14 biomarker and subtyping tasks — for example 0.992 for IDH1 mutation in glioma and 0.840 for BRAF in colorectal cancer. Survival modelling reached a mean C-index of 0.693 internally and 0.640 externally, exceeding 0.70 in five major cancer types.

#Applications

The practical target is molecular pre-screening in settings where sequencing is a bottleneck: flagging which patients are likely to carry an actionable alteration, proposing a molecular subtype from the diagnostic slide, or producing a prognostic score during routine sign-out. Because inference needs only a scanned slide, the workflow fits laboratories that lack in-house genomics, and the released per-task checkpoints let groups evaluate the model on their own cohorts without retraining the backbone.

#Impact

PathLUPI's contribution is methodological as much as empirical: it shows that transcriptomic data is more useful as a training-time constraint on the histology encoder than as an inference-time input, an idea that transfers to any setting where a richer modality exists retrospectively but not prospectively. The evaluation is unusually broad for this literature — 49 tasks, 20 cohorts, multiple external centres — which makes the reported gains harder to attribute to cohort-specific overfitting. Limitations are real: the model inherits its visual capacity, and its gated access terms, from CONCH; the released checkpoints are per-task rather than a single general encoder; the weights carry a non-commercial CC BY-NC-ND 4.0 licence while the code is MIT; and the work remains a preprint under review.