DREAMER-S: Deep leaRning-Enabled Attention-based Multiple-instance approaches with Explainable Representations for Spatial Biology

We introduce an attention-based multiple-instance learning (MIL) framework, DREAMER-S, that learns, from slide-level labels alone, to pinpoint regions in 3D imaging hypercube datasets whose signatures are most predictive for a downstream target classification, acting as a filter over massive, multi-channel image sets. We demonstrate DREAMER-S in its application to tissue images captured by Quantum Cascade Laser IR imaging, where the model assigns attention weights to each spectrum, producing spatial “importance maps” that localize class-relevant regions without requiring manual annotation, and automatically surfaces the most informative spatial instances for downstream analysis. As the DREAMER-S MIL attention layer yields ROI maps and an instance-level filter, the technique is broadly applicable across spatial biology tasks to narrow regions-of-interest in high-content imaging datasets. We demonstrate the approach on a chemotherapy-response task in colorectal cancer patient-derived xenograft (PDX) models, separating a chemo–sensitive class (CRC0344) from a less responsive class (CRC0076) with a best validation F1-score of ~0.95 using a compact, residual-free linear network tuned by grid search. The attention mechanism both accelerates computation and functions as a data-reduction stage, enabling targeted interpretation with SHAP that highlights biochemically meaningful spectral features (nucleic acids, lipids, amide features) while excluding sample artifacts such as paraffin wax. In validating DREAMER-S from an explainability perspective, we connect model saliency to cellular physiology, finding: (1) unsupervised embeddings of the high-attention spectra stratify samples by treatment (FOLFOX, ABT-199, combination, vehicle), and (2) selected spectral markers correlate with pro-apoptotic proteins (Bim, Puma) measured independently in the same PDX system, overall supporting a mechanistic link between spectral signals and apoptosis pathways. Overall DREAMER-S supports efficient, explainable and interpretable analysis of high-content spatial-biology imaging sets.