A pathway-guided kernel MLP architecture for interpretable cross-disease transfer from rheumatoid arthritis to hematological malignancies

Gene expression profiling has improved understanding of disease-specific molecular processes; however, the extent to which transcriptomic patterns from chronic inflammatory conditions can be reused across hematological malignancies remains unclear. Here, we examine the transferability of molecular knowledge from Rheumatoid Arthritis (RA) to hematological malignancies using a pathway-guided kernelized multilayer perceptron (MLP) architecture incorporating biological prior knowledge from the KEGG pathway database. The model was trained repeatedly on RA data across 80 independent runs to assess pathway selection stability. Notably, leukemia-related pathway representations showed high selection stability, with Acute Myeloid Leukemia- and Chronic Myeloid Leukemia-related pathways retained in 73 and 79 of 80 replications, respectively. A frequency-based binomial filtering strategy yielded 116 robust pathways from an initial pool of 320, improving transfer performance to the initial Acute Myeloid Leukemia (AML) cohort from 0.69 to 0.82 AUROC. Independent validation in a second AML cohort further demonstrated reproducible transferability, achieving a mean AUROC of 0.75 across 80 repetitions. Comparative analyses across additional hematological malignancies, including Chronic Myeloid Leukemia (CML), Myelodysplastic Syndromes (MDS), and Acute Lymphoblastic Leukemia (ALL), revealed differential transferability of the RA-derived pathway representations across disease contexts. Together with the frequent selection of leukemia-related pathways during RA source-domain training, these results indicate that RA-derived pathway representations show differential transferability across multiple hematological malignancies, rather than being limited to AML. In contrast, transfer to Prostate Cancer remained near random, indicating substantially weaker generalization to a biologically distant solid-tumor context. Adjustment for estimated cellular composition reduced RA-to-AML transfer performance, but the residualized transfer remained significantly above random expectation, suggesting that cellular composition contributes to the observed transferability but does not fully account for it. Independent analysis using the PID pathway database yielded consistent performance trends, supporting the robustness of the identified transferable pathway-level patterns across alternative pathway collections. Collectively, these findings suggest that pathway-constrained, stability-aware transfer learning can identify reproducible transcriptomic patterns exhibiting heterogeneous transferability from RA across hematological malignancies. These computational associations should be considered hypothesis-generating and warrant further biological and experimental validation to determine their underlying mechanisms.