AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-19
Evolutionary Informed Translation from Protein to Nucleic Acid Sequence
The nucleic acid sequence chosen for a protein affects its function and differs by species, yet existing translation models treat sequences in isolation and tend to favor frequently used codons. The authors introduce PINATA, which uses sequences from related species in multiple sequence alignments to guide predictions. Across 61,934 gene trees and 882 species, it generated sequences that more closely mirror natural ones than existing language models, improving exact native codon recovery by 17.95 percentage points.