cs.MAAug 25, 2026

Towards Trustworthy Biological Alignment in TabPFN-Probed Pathology Foundation Models

Authors: Ushashi Bhattacharjee, Alloy Das, Saria Hannan, Tirtho Roy, Koushik Howlader, Soumik Sarkar

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Histology and transcriptomics provide complementary views of tissue biology, capturing spatial morphology and molecular activity, respectively. Pathology foundation models (PFMs) learn rich morphological representations from H&E images, yet strong downstream performance alone does not establish whether these representations encode biologically meaningful and robust molecular information. We present a training-free framework for auditing biological alignment in frozen PFMs using spatially paired histology and transcriptomics from HEST-1k, evaluated on 240 samples spanning three organs. Multiple frozen PFMs are used to extract H&E representations, while gene expression is aggregated into biologically interpretable pathway-level programs. We use TabPFN as a pretrained probe to quantify the extent to which these molecular programs can be decoded from frozen image representations without task-specific gradient updates. Beyond predictive performance, our audit examines whether pathway decodability generalizes across tissue sections, patient groups, and tissue types; whether representations exhibit section-level or other shortcut dependencies; and whether predictions remain stable under small image perturbations and context resampling. This multi-tissue evaluation distinguishes molecular programs that are consistently encoded from those that are tissue-specific, unstable, or shortcut-sensitive. Our framework therefore provides a systematic approach for assessing not only what biological information pathology foundation models encode, but also how reliably that information persists under clinically relevant sources of variation.

Figures & tables

Figure 1: Training-free morpho–molecular audit. Spatially paired H&E and transcriptomics from HEST-1k are encoded by a frozen pathology foundation model (UNI, Phikon, Phikon-v2, Lunit, or CTransPath); a frozen TabPFN probe then decodes gene expression and pathway activity in-context. No parameters are updated, so decodability reflects the representation and the underlying biology rather than a trained prediction head.
Figure 2: The four audit operators. Applied to the same frozen pipeline: (1) negative controls (matched vs. label-shuffled vs. random-representation), (2) distribution shift (IID vs. organ-OOD), (3) benign image perturbations, and (4) the section-identity shortcut. A representation is trustworthy only if decodability is high and survives all four.
Figure 3: Decodability across encoders. Per-organ decodability (mean PCC, K=512K{=}512 ; 95%95\% CI over seeds) across five frozen pathology encoders (one modern large-scale PFM, UNI, and four established pathology encoders) on Breast, Skin, and Brain, for (a) pathways and (b) individual genes. Only the encoder changes, so the in-context probe ranks representations directly (colourblind-safe Okabe–Ito palette).
Figure 4: Label efficiency and in-context scaling. (a) UNI+TabPFN attains a pathway PCC comparable to the reported HEST-1k trained-probe reference using ≈\approx 23 ×\times fewer labeled context examples, while the other frozen encoders never reach that level at any context size. (b–d) In-context scaling per organ: pathway decodability versus context size KK for all five encoders ( 95%95\% CI band over seeds); only UNI rises with context.
Figure 5: Robustness under shift. Per organ, the IID-to-organ-OOD drop in pathway decodability for all five frozen encoders (dashed line = label-shuffle floor). High IID scores do not guarantee transportable molecular alignment: UNI retains roughly half its signal on Breast and Skin but collapses on Brain, while the substitutes have no signal to lose.
Figure 6: Trustworthiness controls. Per organ, matched decodability ( K=512K{=}512 , solid) vs. the within-section control ( K=256K{=}256 , the largest context feasible within one section, hatched) and the label-shuffle floor (dashed). UNI’s residual remains above the floor, although the unequal context sizes limit a direct effect-size comparison; the substitutes’ controls meet or exceed their matched scores.
Figure 7: Spatial prediction of pathway activity. For one representative section per organ (Breast, Skin, Brain), rows show the most-decodable Hallmark pathway. Columns show the ground-truth spatial map beside reconstructions from the training-free TabPFN probe and the kNN, Ridge, and MLP baselines, all sharing the same frozen UNI features and K=512K{=}512 context; per-panel PCC is annotated. TabPFN reproduces the spatial pattern most faithfully, and decodability is strongest in structured tissue (Skin) and weakest in Brain—mirroring the quantitative trends in the main text.

Explore similar work

CardsList

  1. Spatial Transcriptomics-Guided Alignment Enhances Molecular Profiling in Pathology Foundation Model

    May 29, 2026Fengtao Zhou, Yingxue Xu, Zhengyu Zhang +20Spatial TranscriptomicsCross-Modal Alignment

    Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong SAR, China · Department of Pathology, Nanfang Hospital, Southern Medical University, Guangzhou, China · Department of Pathology, School of Basic Medical Sciences, Southern Medical University, Guangzhou, China +16

  2. CellPath-Bench: A Multidimensional Benchmark for Whole-Slide Cellular Representations in Pathology Foundation Models

    Aug 21, 2026Bokai Zhao, Yiyang Zhang, Hanqing Chao +6Computational PathologyPathology Foundation Models

    School of Artificial Intelligence, University of Chinese Academy of Sciences. · Brainnetome Center, Institute of Automation, Chinese Academy of Sciences. · DAMO Academy, Alibaba Group. +1

  3. Data-Efficient Multimodal Alignment for Histopathology-based Molecular Prediction

    Jun 29, 2026Dominik Winter, Dominik Vonficht, Loïc Le Bescond +6Computational PathologyCross-Modal Alignment

    AstraZeneca Computational Pathology and Biomarkers, Munich, Germany · AstraZeneca, Early Oncology and Translation Medicine, Cambridge, UK