Eric Durand

Accurate prediction of patient outcomes remains a major challenge in oncology. While recent machine learning (ML) approaches often rely on bulk omics lacking spatial resolution or histology-based multiple instance learning (MIL), spatial transcriptomics (SpT) provides a unique opportunity to capture both molecular content and tissue architecture. However, no generalizable ML framework has yet been established to exploit SpT for patient-level outcomes. We present SpaMIL, a flexible and interpretable MIL framework designed for SpT, with a distillation strategy that enables deployment for hematoxylin and eosin (H&E) slides alone. We evaluate the framework by predicting survival from glioblastoma (GBM) patients, a clinically compelling setting given its aggressiveness with a median survival of only 15 months and the lack of prognostic clinical variables. We analyzed 76 GBM cases from the MOSAIC dataset: 43 with matched SpT, H&E, single-nucleus RNA-seq (scRNA-seq), bulk RNA-seq, and clinical variables, and 33 with H&E for external validation. We developed two main architectures: abMIL, tailored to SpT’s spatial molecular structure, and MabMIL, which distills SpT-derived representations into H&E. Model interpretability was achieved through a Shapley-based framework linking prognostic predictions to cell-type compositions via SpT deconvolution. In benchmarking across the five GBM MOSAIC modalities, SpT-based abMIL achieved unprecedented prognostic accuracy (median C-index: 0.72, standard deviation: 0.04), outperforming all other modalities, including established clinical predictors. PCA and deconvolution-based SpT representations surpassed recent foundation models, suggesting the need for further research on SpT foundation models. Our interpretability analysis highlighted malignant and non-malignant cell subpopulations associated with favorable or poor prognosis, consistent with recent reports. Finally, MabMIL maintained strong performance while enabling H&E-only deployment, with improved condorance index over H&E-only baselines in both internal (0.59 vs. 0.57) and external (0.62 vs. 0.55) cohorts.

Background: Glioblastoma (GBM), the most aggressive primary brain tumor, has a median survival of approximately 15 months. Twenty percent of patients survive beyond three years, but known clinical factors like age, performance status, resection extent, and MGMT promoter methylation status do not fully explain the observed outcomes. Objective: Our objective was to identify novel histology derived biomarkers associated with end-of-spectrum overall survival (OS) to provide novel biological insight with a translational potential. Methods: We analyzed a total of 748 GBM patients from 3 different cohorts, uniquely enriched in long survivors (n=98 with overall survival (OS) > 5y including n=196 with OS≥3y), with clinical data and H&E slides obtained from the primary tumor at baseline. We propose an interpretable machine learning (ML) methodology for the discovery of histological biomarkers. Our method learned to segment each H&E slide into three distinct regions associated with long-term survival, short-term survival, and non-informative tissue. We characterized these regions by integrating unsupervised learning, nuclei segmentation, blood vessels detection, pathologist annotations, and multimodal data including spatial transcriptomics from n=31 patients of the GBM MOSAIC dataset to discover fully interpretable biomarkers. Results: Our OS prediction model using histology and clinical data as input achieved an area under the curve (AUC) of 0.85 for the classification of patients between OSPLIN2 expression and lipid accumulation- is consistently enriched in histological areas predictive of poor prognosis. Conclusion: Our interpretable ML methodology identified a novel prognostic impact of biological processes and cell types according to distinct tumor regions of GBM. These results pave the way for spatially-informed biomarkers to improve risk stratification and for personalized spatially-targeted therapeutic strategies. Key highlights: Our ML model identified histological biomarkers predicting prognosis independently from known clinical factors The region of lowly infiltrated white matter enriched in malignant cells including a mesenchymal-like phenotype is predictive of poor prognosis Angiogenesis is increased in areas predictive of long survival in main non-necrotic tumor regions. The subtype of macrophages expressing PLIN2 and associated with increased lipid metabolism was associated with poor prognosis in all GBM regions. Highlights:

Recent advancements in probe-based, full-transcriptome, high-resolution technologies for Formalin-Fixed Paraffin-Embedded (FFPE) tissues, such as Visium CytAssist, Chromium Flex (10X Genomics), and GeoMx DSP (Nanostring), have opened new opportunities for studying decades-old archival samples in biobanks, facilitating the generation of data from extensive cohorts. However, the experimental protocols can be labor-intensive and costly; therefore, it is thus essential for researchers to carefully evaluate the strengths and limitations of each technology in relation to their specific research objectives. Here, we report the results of a comparative analysis of the three methods mentioned above on FFPE archival tumor samples from four non-small cell lung cancer, four breast cancer and six diffuse large B-cell lymphoma. We highlight some relative advantages and disadvantages of each method in the context of operational challenges, bioinformatic analysis and biological discovery. Our results show that: 1) all three methods yielded good-quality, highly reproducible transcriptomic data from serial sections of the same FFPE block; 2) GeoMx data contained mixtures of cell types, even when pre-selecting areas with cell type-specific markers; 3) high-throughput spot-level (Visium) or cell-level (Chromium) data enabled the identification of tumor heterogeneity within and between patients, which could be used to identify targeted therapies. Our data support the use of Visium and Chromium for high-throughput and discovery-driven projects, while the GeoMx platform could be suited for addressing specialized questions on targeted regions. All data generated from this study, including GeoMx, Visium, Chromium, H&E, and expert annotations are publicly available.