MIRCID: Inferred Hub-miRNAs Drive Cross-Task Improvements in Drug Mechanistic Modeling

MIRCID, a framework for comparing gene expression with inferred transcription factor activity and miRNA expression, has been developed to enhance cross-task improvements in drug mechanism-of-action (MoA) modeling. This innovative approach addresses the challenge of limited matched miRNA measurements in perturbational transcriptomes.

The core of MIRCID, HubmiRNet, infers 414 pan-cancer hub miRNAs from 977 L1000 landmark genes, demonstrating an impressive Pearson correlation coefficient of 87.72%. Additionally, the 1,298-output variant of HubmiRNet surpassed SiCmiR on the full-miRNA task, achieving a remarkable 71.21% performance compared to 67.30% previously reported.

When comparing the impact of miRNA augmentation versus TF activity, the study found that miRNA augmentation provided more consistent gains across evaluations. The generic embedding controls also exhibited model-dependent utility, while complementarity analyses revealed a distinct, partially linearly recoverable representation that preserved the original gene-derived structure.

Furthermore, illustrative rescue cases showcased how improved classification was linked to biologically plausible miRNA patterns in samples with weak transcriptional signatures. These findings strongly support the use of inferred HubmiRs as a biologically informed re-coding of transcriptomic data for perturbational drug modeling. However, the recovery of measured perturbational miRNA responses still requires validation through further studies.