Detection of deleterious variants by protein language models is improved by nucleotide conservation at three evolutionary timescales in rice

Predicting the effect of genetic variants can boost the efficiency of crop breeding. This calls for developing and applying novel approaches to predict mutation effects on plant fitness based on evolutionary data. Here, we examine the potential of protein language models (pLMs) to predict highly conserved sites in the rice genome, using nucleotide conservation as a proxy for variant effects on fitness. We leveraged ESM-1v, a pLM used to generate sequence representations, which delivered input features for three supervised models (MLP, CNN, XGBoost) to predict evolutionarily conserved sites across species, at three evolutionary timescales: angiosperms, monocots, and temperate grasses (C3 grasses). Results show that we can predict nucleotide conservation in the rice genome at all three timescales. Our supervised models substantially outperformed the random expectation (P4 from 0.61 to 0.67 vs. 0.40 to 0.43), with comparable performance across model architectures. Importantly, predicted conservation at smaller evolutionary timescales showed stronger associations with rare variants in rice. Furthermore, we showed that combining our models with ESM-1v can lead to stronger functional enrichment for fitness effects than ESM-1v alone. Our study demonstrates that supervised machine learning models trained on evolutionary data can accurately prioritize deleterious variants in silico , potentially reducing the reliance on resource-intensive annotations and providing clade-specific information about variant effects.