DisenTE: Sparse Pattern-Context Modeling for Interpretable Translation-Efficiency Matrix Completion

In the realm of data-rich science, object-by-context matrices often present partially observed data, where dominant object effects can mask smaller but significant context-dependent variations. This challenge was examined in a translation-efficiency atlas of 9,494 5 UTRs spanning 78 cellular and tissue contexts. The researchers introduced DisenTE, a neural model capable of handling these complexities.

DisenTE employs a sequence-conditioned architecture, integrating distinct sequence and context branches linked by a sparse low-rank pattern-context channel. Each module merges a sequence-derived activation with context-specific deployment weights, yielding a dictionary where both sequence and context components can be analyzed independently.

When subjected to five-fold within-panel entry masking, DisenTE demonstrated an impressive UTR-centered residual Spearman correlation of 0.641 +/- 0.005, a marked improvement over the 0.304 +/- 0.003 achieved by the best reference model. The learned dictionary preserved 11 out of 20 candidate modules, including CTM 6, which exhibited the strongest overlap with an external TOP set and a cap-proximal pyrimidine pattern.

CTMs 5 and 7 also shared similarities with the set but featured purine-containing consensuses, suggesting their potential as TOP-set-associated factors.

The findings bolster CTM 6's status as a top sequence anchor and position CTMs 5 and 7 as key elements of the TOP set. Importantly, DisenTE outperformed the evaluated references, completing the dataset with greater accuracy while concurrently generating module-level summaries of its context-dependent variations.