Macrophage signature-based prediction of cancer treatment response using MIL-attention
Predicting immunotherapy response from single-cell data remains difficult due to patient-level labels, extreme class imbalance, and highly heterogeneous macrophage states. We present a Multiple Instance Learning (MIL) framework that treats each patient as a bag of macrophage embeddings derived from a single-cell RNA foundation model. The architecture incorporates an attention-based pooling…
Detecting immunotherapy success from single-cell data is challenging because patient labels are scarce, class imbalance is extreme, and macrophage states vary widely. The researchers introduce a Multiple Instance Learning (MIL) method that groups each patient into a "bag" made up of macrophage embeddings generated from a single-cell RNA foundation model.
To keep the model size manageable, the design uses an attention-based pooling system with less complexity, dropout regularization, and explicit attention penalties to boost stability when working with small samples.
To handle uneven clinical datasets, the MIL outputs are fine-tuned using a blend of focal loss and supervised contrastive objectives. This combo sharpens class distinctions and enhances representation clustering all at once. When tested on three cancer datasets, this strategy outperforms both pseudobulk aggregation and ordinary MIL approaches.
By employing attention-weighted attribution and examining transcriptional regulation, the study identifies unique macrophage programs involving interferon and antigen presentation pathways. Responders show interferon networks, while non-responders exhibit hypoxia-linked regulatory modules. This finding highlights the potential of MIL to pinpoint predictive and mechanistically interpretable immune states.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
