Lifespan prediction in protostomes from trinucleotide composition

Background Promoter CpG composition has been proposed to predict vertebrate lifespan, with higher CpG density hypothesized to buffer age-related DNA methylation drift. Recent comparative analyses in protostomes suggest that lifespan associations extend to other short motifs across multiple genomic regions. We tested whether these sequence features can predict lifespan in protostomes. Results Using lifespan and genomic data for 466 species spanning eight protostome phyla, we compared di-, tri-, and tetranucleotide lifespan predictors across different sets of genomic regions. Models were fitted using least absolute shrinkage and selection operator (LASSO) regression with 327 species for model development and evaluated in a fixed-holdout set of 139 species. Conserved genes were identified with Benchmarking Universal Single-Copy Orthologs (BUSCO). The best-performing model combined relative frequencies (Frequency) and observed/expected ratios (O/E) of 64 trinucleotides across six BUSCO-associated regions: exons, introns, and two regions each around the start and stop codons. Fixed-holdout predictions achieved a predictive R² of 0.594 for ln(lifespan), a mean absolute error of 5.24 years, and a median fold error of 1.91. The model outperformed models based on phylogenetic relatedness or genome size, as well as an individual-gene upstream CpG O/E model adapted from previous promoter-based approaches. Frequency and O/E contributed complementary regional information, with Frequency coefficient magnitudes most prominent around translation initiation and O/E coefficient magnitudes in upstream and downstream flanking regions. In leave-one-taxon-out validation, models were trained excluding any species from the target taxon. At higher taxonomic ranks (phylum to order), the models captured relative lifespan differences within excluded groups, although absolute lifespan estimates were inaccurate for phyla and classes. At lower ranks (family and genus), the models captured little within-group variation but produced lower prediction errors, suggesting that their higher accuracy largely arose from correctly estimating the average lifespan of each family or genus, rather than predicting variation among its species. Conclusions Regional trinucleotide composition supports lifespan prediction across multiple invertebrate phyla and provides information beyond that captured by the phylogenetic predictors. Its ability to recover lifespan associations in taxa absent from training supports the potential applicability to genomic lifespan estimation in any animal group.