SignalPath-Finder: Identifying TOR complex downstream target genes via pseudo-peak ranking and autoencoder-based optimization

Document Type

Article

Publication Date

6-2026

Department

College of Computing; College of Forest Resources and Environmental Science

Abstract

Identifying downstream genes regulated by signaling pathways remains challenging because signaling complexes generate nonlinear, context-dependent transcriptional responses that are difficult to detect in heterogeneous transcriptomic datasets. Although large volumes of RNA-seq data are publicly available, most datasets are not generated through pathway-specific perturbations, limiting the ability of conventional statistical or network-based approaches to identify downstream effectors. Here, we developed SignalPath-Finder, an AI-driven computational framework for extracting pathway-level regulatory signals from heterogeneous RNA-seq cohorts. The framework first reorders transcriptomic samples using signaling complex components as reference anchors and then applies a pseudo-peak transformation to convert heterogeneous expression profiles into aligned bell-shaped patterns. Genes exhibiting similar distributional and structural characteristics are identified through statistical similarity metrics and grouped using unsupervised clustering. To capture nonlinear regulatory structure beyond correlation-based approaches, SignalPath-Finder incorporates a cluster-wise autoencoder-based representation learning module that learns low-dimensional latent manifolds underlying gene expression profiles. Genes are prioritized according to their contribution to the learned latent representation, enabling identification of candidate downstream genes through a two-stage autoencoder-based selection framework. Using the TOR complex in Populus trichocarpa and a compendium of 628 RNA-seq samples, SignalPath-Finder identified coherent downstream modules and enriched both known and previously unrecognized downstream genes with literature-supported functions, showing improved recovery of known downstream genes (KDGs) compared with conventional correlation and gene regulatory network methods. Overall, SignalPath-Finder provides an AI framework that integrates distributional pattern discovery with autoencoder-based representation learning to identify downstream regulatory genes from heterogeneous transcriptomic datasets and facilitates generation of biologically testable pathway-level hypotheses.

Publication Title

Artificial Intelligence in the Life Sciences

Link to Full Text

Share

COinS