Tree Genetics & Genomes Biological sciences

CLAP-HMM v2: a cross-species adaptive transfer learning framework integrating domain adversarial training and few-shot meta-learning

Wang, Liuyan, Xu, Yingfan, Guan, Xuemei
Springer Nature · 2026-08-21 · DOI 10.1007/s11295-026-01755-8

Abstract

The huge genome of coniferous trees and the extremely scarce validation genes lead to a long-term inefficiency in the exploration of stress-resistant genes. This study proposes the cross-species adaptive transfer learning framework CLAP-HMM v2, which systematically integrates domain adversarial (DANN), dynamic MMD regularization, and MAML few-shot meta-learning, successfully solving the negative transfer problem caused by the over 320 million years of evolutionary distance between poplar and pine trees. Across four representative pine species, including Pinus taeda , Pinus tabuliformis , Pinus albicaulis , and Pinus radiata , the model achieved an F1-score of 0.912 (improving by 14.1% compared to traditional transfer and by 31.0% compared to from-scratch training), AUC-PR reached 0.947, and the F1 drop rate in the 1:50 extreme imbalance scenario was only 7.3%, with the transfer efficiency index (TEI) increasing by 93%. The whole-genome scan predicted 1238 high-confidence candidate stress-resistant genes ( P > 0.90), among which 512 were highly homologous to known R genes, and the enriched pathways significantly pointed to drought response, MAPK cascade, plant-pathogen interaction, and conifer-specific terpene synthesis. Attention and HMM visualization precisely located stress elements such as W-box and ABRE. The candidate gene priority list produced by this framework can directly guide CRISPR/Cas9 knockout and multi-gene stacking, providing an efficient computational framework for prioritizing candidate genes and supporting future experimental validation in conifer stress-resistant molecular breeding.

Read at publisher →