FLEXMIL：用于临床与转化研究的灵活多模态多示例学习框架

Background: Multiple-instance learning (MIL) provides a powerful framework for training deep neural networks in settings lacking detailed annotations, particularly in digital pathology, where slide-level labels are routinely available, but region- or tile-level annotations are scarce. However, most existing MIL toolkits suffer from limited maintenance, functionality, and restrictive licenses, hindering adoption in clinical-translational research. Methods: We present FLEXMIL, a flexible, end-to-end MIL framework designed to unify pathology, omics, and clinical data analysis. FLEXMIL was developed in Python (V3.10) to enable survival analysis, classification, and regression with robust cross-validation, automatic data splitting, and standardized reporting to ensure reproducibility. A co-attention fusion module was implemented to integrate multimodal data including clinical information, histopathology images and omics data, while also supporting single-modality experimentations. FLEXMIL generates slide-level predictions, patient-level summaries, and attention-based heatmaps to facilitate biomarker discovery, visual interpretation and hypothesis generation. Results: We evaluated FLEXMIL across diverse translational use cases, including biomarker discovery, tumor target expression prediction, survival analysis and demonstrated its versatility and robust performance. Importantly, a significant advancement was observed when FOLR1-related transcriptomic signatures were integrated with features extracted from H&E-stained images, leading to enhanced predictive performance for FOLR1 expression levels. Specifically, this multimodal approach boosted not only the AUC from 0.72 to 0.83 in binary classification, but the correlation coefficient for continuous values predicted from 0.5 to 0.78 in TCGA-LUAD (N=460). Same approach further improved the AUC of the predicted FOLR1 protein expression from 0.78 to 0.83 in a commercial cohort (N=69) and enhanced overall survival prediction (C-index increased from 0.59 to 0.62) in TCGA-LUAD (N=334), outperforming models based on image data alone. In the biomarker identification task, we predicted tumor-infiltrating-lymphocytes (TILs) in 85 TCGA-TNBC samples with %TIL annotated by two pathologists. At a 10% TIL threshold, FLEXMIL achieved an AUC of 0.89, highlighting the strong predictive value of histopathology features. In addition, FLEXMIL can generate attention heatmaps for image-only and multimodal (co-attention) models, with support for multi-head views and transcriptomic signature-guided explanations to enable transparent slide-level interpretability. Conclusions: FLEXMIL provides a flexible, scalable and interpretable platform that bridges computational modeling and clinical insight, advancing the development of integrative biomarkers for precision oncology.