Regional Signals Matter: Rethinking Global Pooling for mRNA Foundation Models

Region-aware pooling of mRNA embeddings (5′ UTR, coding sequence, 3′ UTR) improves linear-probe performance versus global mean pooling, boosting Pearson correlation by 11–25% on translation and stability tasks.

• Why it matters: Global mean pooling can dilute signals from short regulatory regions, making benchmark scores reflect readout limitations rather than how well mRNA foundation models preserve region-specific information.
• What they did: The authors propose an architecture-agnostic region-aware readout that averages embeddings within each mRNA region and concatenates them for linear probing, then evaluate 20 frozen models across mRNABench and a new region-aware benchmark.
• The result: Across 380 model–dataset–target evaluations, region-aware pooling increases Pearson correlation by 11–25% on translation and stability tasks without statistically significant regression declines, and classification AUPRC changes range from −3% to +11%.

AI summary of the abstract · may contain errors · not medical advice · read the paper.