Machine Learning based Classification of Ocular Pathogens using Host-transcriptomic Signatures: a proof-of-concept

Diagnosis of infectious keratitis, uveitis, and endophthalmitis is constrained by the scarcity of pathogen-derived material in small-volume ocular specimens. The host transcriptome is an abundant alternative signal: infection elicits pathogen-class-specific transcriptional responses that may permit pathogen identification without direct detection. We tested this premise as a proof of concept using publicly available in-vitro infection RNA seq. We assembled 330 human RNA-seq samples (177 infected, 153 mock) spanning six pathogens (Herpes Simplex Virus-1, Human Cytomegalovirus Virus, Aspergillus Fumigatus, Candida Albicans, Pseudomonas Aeruginosa, Mycobacterium Tuberculosis) across 36 BioProjects and 24 host cell types.

Per-study differential expression yielded a robust-gene feature matrix. Pathogen classification was evaluated with Random Forest under leave-one-study-out (grouped) cross-validation and contrasted against random cross-validation to quantify study-level information leakage. Under random (non-grouped) cross-validation, balanced accuracy reached 0.85-0.92 across gene panel sizes (8-700 + genes). Under honest, leakage-corrected leave-one-group-out validation, balanced accuracy was lower than random cross-validation and remained stable across panel sizes and gene-selection strategies, peaking at 0.70 (60-100 gene panel) and averaging 0.519-0.60 across the full range of gene sets tested (8 to 9,942 genes; chance level = 0.14 for 7 classes). Unsupervised UMAP visualization of the robust gene set showed distinct clustering by pathogen and by pathogen class (viral, fungal, bacterial), with cross-kingdom pathogen pairs showing more distinct separation than within-kingdom comparisons.

Beyond established diagnostics such as PCR and metagenomics, host transcriptomic signatures analyzed via machine learning offer a promising alternative approach to pathogen classification, achieving balanced accuracy substantially above chance under rigorous, study-independent validation. These findings support the feasibility of host-transcriptome-based diagnostic classification, with dataset provenance as the primary current limitation to clinical translation.