Deep learning integration of single-cell and spatial transcriptomics reveals a neutrophil-associated mesenchymal niche and a candidate translocon dependency in lung adenocarcinoma

In lung adenocarcinoma, epithelial–mesenchymal transition drives treatment resistance and metastasis but has been dissected mainly through SPP1 ⁺ macrophages, leaving neutrophils under-resolved; we asked whether integrating consensus non-negative matrix factorization, variational-autoencoder label transfer, reference-based spatial deconvolution and transformer-based in silico perturbation could resolve this niche and nominate tumor-intrinsic dependencies. We built a machine learning framework integrating seven single-cell RNA-sequencing cohorts, two spatial transcriptomic cohorts, bulk RNA sequencing with survival, CRISPR essentiality and three immunotherapy cohorts. Unsupervised consensus matrix factorization defined malignant programs; deep generative models (scVI/scANVI) resolved neutrophil states; probabilistic deep learning (cell2location) with random forest multi-view modeling mapped spatial niches; and a transformer-based single-cell foundation model (Geneformer) performed in silico gene deletion perturbation across three coupled state transitions, feeding a four-endpoint orthogonal validation cascade. Consensus factorization resolved four malignant meta-programs. The mesenchymal/interferon program MP3 carried prognostic information only in the context of the other three programs (hazard ratio 1.43 per 1 SD, 95% CI 1.12–1.82, p = 0.0045), and the composite score z (MP3) − z (MP4) was prognostic on its own (hazard ratio 1.40, p = 3.4 × 10 −6 ), reciprocally balanced by a protective alveolar-like MP4 (0.72, p = 0.010). We resolved an OSM-primed neutrophil axis forming a partitioned dual circuit with SPP1 + macrophages; spatially, OSM flux was directed toward macrophages and fibroblasts rather than toward malignant cells, so any effect on the malignant compartment is likely indirect. Spatial deconvolution localized the circuit to a reproducible two-compartment niche across both cohorts. Foundation model perturbation followed by orthogonal validation nominated SEC61G with convergent support across four orthogonal endpoints (overall survival hazard ratio 1.64, p = 1.3 × 10 −6 ; higher expression in immunotherapy non-responders in two of three cohorts, significant in one). The cascade re-derived SEC61G, a gene with independent published support in lung adenocarcinoma, de novo—we therefore treat it as a positive control rather than a new nomination, and the property we report as new is that its prognostic signal is independent of 7p11.2 copy number. The foundation model perturbation plus multi-endpoint validation framework offers a generalizable strategy—foundation model perturbation followed by multi-endpoint orthogonal validation—for nominating candidate targets in plasticity-driven solid tumors.