Crop breeding is all about improving the traits of plants – from higher yields, nutritional value and better taste, to resistance to pests and diseases. To meet ever-increasing expectations from growers and society, crop breeders need to accelerate trait discovery.

In 2024, InstaDeep, a pioneering artificial intelligence (AI) company, approached Syngenta with a hypothesis: that their state-of-the-art large language model, AgroNT, could help Syngenta Seeds significantly accelerate trait discovery to enhance the global food system.

Lynne Mumm, Syngenta Technology Licensing Manager, explains, “The idea was intriguing. Syngenta has world-class scientists with expertise in bioinformatics and AI, and they are extremely busy focusing on the myriad of today’s agricultural challenges. InstaDeep’s team was looking at how we can use AI to solve tomorrow’s challenges.”

The collaboration between InstaDeep and Syngenta has resulted in an expanding      integration of Large Language Models (LLMs), such as AgroNT, into Syngenta’s human-led research pipeline.

Bringing the power of Large Language Models to trait discovery

The initial six-month feasibility study brought together Syngenta’s world-leading proprietary trait research capabilities and InstaDeep’s foundational LLM, Agronomic Nucleotide Transformer (AgroNT).

In a previous research project, AgroNT was trained on trillions of nucleotides from 48 agriculturally relevant crop species to interpret the complex language of the genetic code.

“AgroNT is based on the same transformer architecture that powers ChatGPT,” says Gurnek Singh, Head of Business Development, AI for Life Science at InstaDeep. “Where ChatGPT was trained on about 300 billion words to model language, we pre-trained the AgroNT on approximately 10.5 million genomic sequences comprising trillions of base pairs. The data set of 48 species included genomes of row crops, fruits, legumes, vegetables and species important for industry and research, so that it could learn the language of plant DNA.”

In this way, a next-generation AI technology has been able to learn from nature, improving our understanding of how genes are regulated and potentially enabling a new level of trait control and crop performance.

The initial phase of the collaboration focused on AI-mediated trait design for both corn and soybeans.

Joe Clarke, Syngenta Fellow in Genomics continues, “These cutting-edge AI tools are enabling us to understand the language of DNA regulation in ways that we have never been able to do before. They are enabling the adoption of AI-assisted design approaches that optimize sequence of trait genes to improve performance, reduce risk and accelerate our ability to deliver to farmers.” 

AI-assisted trait design

AgroNT has been developed to support plant trait research as Syngenta’s scientists look for solutions that address modern challenges in agriculture.

In adopting the model, Syngenta is leveraging the potential of AI to enable its scientists to achieve more ground-breaking science, faster.

“AI has become a powerful tool in Decision Sciences,” explains Joe, “allowing our scientists to extract insights from vast amounts of DNA sequence and genomic data to improve key scientific decisions on how to improve traits in plants.”

A scientific collaboration moving to the next stage 

“When looking for a collaboration partner, we wanted to work with a leader in the field,” says Gurnek. “Syngenta is built to find the most important innovations and leverage them to improve what they do. The collaboration has been a tremendous experience – not just because of the exciting results but because of how well the InstaDeep and Syngenta teams have worked together and learned from one another.

Our collaboration was so enjoyable that time on the project seemed to pass very quickly.”

Joe agrees: “When we met InstaDeep, they had a mature AI Platform and were truly focused on agriculture. I was impressed that they knew the value of their algorithms, our data and the partnership. That made it easy to put together a collaboration with them.” 

This initial phase of using the LLM has already revealed several advantages.

“Trait discovery has historically been hypothesis driven with decisions made based on empirical evidence following rigorous testing,” says Joe. “These AI tools allow prediction – they offer foresight on how genes should perform and how to improve that performance. This will accelerate our ability to deliver traits to farmers”.

The collaboration is now moving into a second phase to explore more ways to bring highly desirable traits to farmers.