RNAStabFormer: Region-Aware Multi-Task Hybrid Learning for RNA Stability Prediction from Pulse-Chase Transcriptomics

Abstract

RNA stability is a major post-transcriptional regulator of gene expression, yet sequence-based prediction from pulsechase transcriptomics remains difficult because labels depend on time window, quantification region, and replicate quality. We present RNAStabFormer, a controlled RNA stability framework centered on a Region-aware Multi-task Hybrid Transformer (RAMHT). RAMHT encodes 5'UTR, CDS, and 3'UTR nucleotide context, adds a CDS codon stream, upgrades engineered sequence features into a tabular interaction branch, and uses gated multi-task regression to predict four ENCODE BrU-seq/BruChase-seq stability proxies, with exon total 6 h/0 h as the primary task. Across 26 outer splits including 23 chromosome holdouts, a heterogeneous three-member RAMHT ensemble achieves 0.773 mean Pearson correlation on the primary task, statistically matching an engineered-feature XGBoost baseline (0.773; mean paired delta +0.000004; bootstrap 95% CI [- 0.003845, +0.004077]; Wilcoxon p = 0.8613). The ensemble improves over the strongest single RAMHT member (0.768 to 0.773; 23/26 split wins), while a strict nested XGBoost+RAMHT blend reaches 0.775. Same-split checks also exceed frozen full-length mRNA-LM embeddings (0.760) and public LAMAR-DR transfer (0.180). Gate, ablation, and recoding analyses show that engineered sequence grammar remains dominant, while nucleotide and codon branches provide complementary CDS-local signal. RNAStabFormer narrows the gap between neural RNA sequence modeling and strong tabular baselines while retaining an extensible architecture for interpretation and biological integration.