Cross-validated machine learning and deep learning recover a TGF-β–SMAD–RUNX1 regulatory axis underlying glioblastoma myeloid commitment

This Short Communication provides an independent computational corroboration of a TGF-β–SMAD–RUNX1 regulatory axis previously implicated in preceding study of glioblastoma myeloid commitment. Using a fully annotated single-cell myeloid atlas from eleven newly diagnosed human glioblastomas, three independently structured machine-learning architectures: elastic-net logistic regression, gradient-boosted trees, and a multilayer perceptron-were evaluated using patient-grouped cross-validation to distinguish TREM2-positive macrophages from homeostatic microglia. Classification performance remained high after removal of the marker genes originally used to define the two populations. Gene-level feature attribution was unstable under multicollinearity, motivating analysis in a curated transcription-factor activity space. Within the terminally committed population, RUNX1 emerged as a top-decile discriminant of commitment depth and remained stable across ten of eleven leave-one-patient-out exclusions. A structurally independent multilayer perceptron independently corroborated RUNX1 using permutation-based feature importance. Together, these results provide an orthogonal machine-learning and deep-learning test of the regulatory signal identified in the preceding study and support computational cross-validation as a strategy for testing transcription-factor nominations in single-cell analyses.