TabPFN探针审计病理基础模型的生物学对齐
原标题：Towards Trustworthy Biological Alignment in TabPFN-Probed Pathology Foundation Models
AI 摘要
该研究提出一个无需训练的框架，用于审计病理基础模型（PFMs）的生物学对齐性。作者利用HEST-1k中空间配对的H&E组织学图像与转录组数据，以TabPFN作为预训练探针，在240个样本（涵盖乳腺、皮肤、脑3种器官）上评估UNI、Virchow、Prov-GigaPath、Phikon-v2等冻结PFMs的表示。结果显示分子可解码性高度依赖组织类型，通路活性比单个基因表达更稳定可恢复，且强可解码性并不总意味着可信对齐——分布偏移下性能下降，切片身份等捷径信息仍被强编码。
正文节选
Towards Trustworthy Biological Alignment in TabPFN-Probed Pathology Foundation Models Abstract Histology and transcriptomic data provide complementary views of tissue biology through spatial morphology and molecular activity. However, pathology foundation models (PFMs) encode rich tissue morphology, but strong downstream performance does not necessarily indicate that their representations capture robust biological information. We present a training-free framework for auditing biological alignmen