Machine-Learning Models Trained on Genomic Summary Statistics Assess Extinction Risk and Recovery Potential of...

Forecasting extinction risk from genomic data is challenging because of interactions among environment, demography, and genomic erosion. Here, we use known outcomes from forward simulations as a proof-of-concept study to test whether machine-learning (ML) models can discriminate relative extinction risk and recovery potential from genomic summary statistics, and identify the most informative predictors. Using individual-based, forward-in-time simulations, we generated genomic and fitness data for 11,200 populations across 112 demographic parameter combinations (ancestral population size: 100-10,000, bottleneck size: 4-50, fecundity: 8-64). We evaluated populations at three time points relative to bottleneck onset: 100 years before, 10 years before and at onset. We then assessed their subsequent fate up to 200 years after the end of the bottleneck, classifying populations as extinct or recovered. From 10 sampled individuals per population, we derived 738 predictors spanning diversity, heterozygosity, genetic/realised/masked load, and ROH-based inbreeding (FROH). Unsupervised analyses primarily captured historical demography and did not separate outcomes. In contrast, supervised models achieved high performance at bottleneck onset, correctly predicting population fate (i.e., future extinction or recovery) from genomics alone with ≈86%-87% accuracy. However, extinction was harder to predict correctly than recovery because simulated extinctions were rarer. Bottleneck size and fecundity provided strong baseline predictability, with genomics adding modest but complementary information. The upper-quantile ROH/FROH summaries were consistently more informative than mean diversity or mean load. A complementary single-genome analysis showed that reference-genome-style summaries retain useful signal when demographic information is sparse. Together, these results highlight the promise of ML-informed extinction-risk assessment to translate genomic indicators into more actionable, early-warning signals for biodiversity conservation.