New AI large model sharpens data-driven soybean breeding

HEFEI -- Chinese researchers have unveiled an upgraded version of "Fengshu" 2.0, a vertical large model tailored for soybean science, at a national symposium on soybean research, according to Anhui Agricultural University in Hefei, capital of East China's Anhui province.

The model is expected to facilitate data-assisted breeding of soybeans, which is a major grain, oil and feed crop in China.

"Fengshu" 2.0 integrates soybean germplasm resources, genomic and coding sequences, proteins, transcripts, gene expression data, disease samples, breeding trial records, as well as both domestic and international literature, to form a structured knowledge system specifically for soybean research.

The original "Fengshu" model was jointly developed in February by the university and the Institute of Crop Sciences under the Chinese Academy of Agricultural Sciences, as a generative AI breeding platform covering the full life cycle of soybeans.

The two institutions have since completed an iterative upgrade, with the new version featuring a multimodel collaborative analysis mechanism.

According to the research team, "Fengshu" 2.0 embeds interfaces for multiple general-purpose large models, each capable of independently answering domain-specific questions.

A dedicated summarization module then synthesizes their responses by integrating domain knowledge, proprietary data, and structured comparison rules. It evaluates consensus, discrepancies, and evidence completeness across models to generate clearer and more reliable conclusions, thereby enhancing the stability of professional knowledge services in soybean research.

Currently, "Fengshu" 2.0 can assist in disease diagnosis, parental line screening for breeding, virtual cross design, phenotypic data analysis, molecular marker-assisted breeding and functional analysis of candidate genes, all centered on key agronomic traits of soybeans.

"Compared with traditional experience- and experiment-based breeding, 'Fengshu' 2.0 represents a shift toward data-assisted breeding and precision decision-making," said Wang Xiaobo, a professor of the university. "It improves data processing efficiency, reduces repetitive work and optimizes breeding strategies."

Wang added that the model will be further integrated with climate, soil and cropping patterns of major soybean-producing regions across China to support elite seed breeding, large-scale cultivation, scientific research and agricultural extension services.

The team said it will continue to expand field trials, accumulate professional data and refine algorithms, while strengthening validation of model outputs against real-world breeding experiments.

Future efforts will focus on enhancing the platform's capabilities in germplasm exploration, trait analysis, variety selection and production management, with the goal of facilitating the development of new soybean varieties featuring high yield, superior quality and stress tolerance.