Similar to how AI is used for early disease detection in healthcare, AI is now increasingly being applied to identify crop diseases, to forecast crop yields and optimise resource management.
Companies like Enko and Moa Technology are using AI to discover new mechanisms of action and active ingredients for crop protection products.
AI techniques for analysing genomic data in healthcare are being applied to crop breeding. Benson Hill’s CropOS platform, for instance, uses AI to generate soybean breeding predictions with up to 85% accuracy. And just as AI enables personalised medicine, it’s allowing for precision agriculture. John Deere’s See & Spray technology, for example, uses AI to apply herbicides only where needed, reducing usage by up to 66%.
NVIDIA – which provides hardware and software solutions for various AI applications – became involved in both the pharmaceutical and agricultural industries around a decade ago. It has collaborated with equipment manufacturers like John Deere, leading to the development of AI-driven autonomous farming systems.
The company has also been working with Corteva Agriscience, using AI to create new seeds with high yield in different weather and terrain, and that are disease resistant.
That’s thanks to AI-based approaches like ESMFold and DiffDock NVIDIA is claiming significant improvement in the speed and accuracy of protein structure predictions, or protein folding modelling.
ESMFold is a state-of-the-art AI method developed by Meta AI that can quickly generate accurate protein structure predictions, while DiffDock is used in drug discovery for predicting how small molecules interact with proteins. Used in combination, the methods can accelerate the design and evaluation of new molecules to explore designs of new fungicides, herbicides, insecticides, and other seed treatments.
In a recent benchmarking exercise, Corteva used NVIDIA NIMs (or Inference Microservices – pre-packaged, ready-to-use software containers that make it easier and faster to deploy AI models, especially for generative AI applications) for ESMFold and DiffDock to compare their current open-source based deployment to NVIDIA NIMs and found that the speed up was significantly faster than the open-source-based deployment.
“Our platform was developed for the pharmaceutical industry,” explains Johnny Israeli, NVIDIA’s head of drug discovery. “However, you can use it in agriculture as well.” Today’s AI algorithms, he says, are more powerful and capable than a decade ago. While it’s the case that AI models are only as good as the data they are trained on, today we have much larger databases of billions of molecules that we didn’t have a few years ago, says Israeli.
“The data, the algorithms and the computing platforms have all improved rapidly to be able to take on the challenges and problems that need solving in ag.” This serves to shorten the R&D cycle in an increasingly cost-effective way. Cost, however, remains the main bottleneck, he suggests. “I think the biggest challenge that we see for this industry is ROI.” Will we see, he asks, enterprises “prioritise the lower hanging fruit business problem that will demonstrate ROI and drive further adoption?”.
Faster molecule identification
Cost is probably the sector’s biggest choke point, agrees Daniel Ferrante, a partner and AI leader in R&D and data strategy at Deloitte. Ferrante develops AI solutions to accelerate drug discovery and improve the research processes. Concerning the challenge that AI models are only as good as the data they are trained on, he responds: “Everybody hears that small data is a problem for AI. That’s not quite true. Noisy data is more of a problem. Good, curated data is perfect, and we can use and make great progress with that.”
He reiterated that newer models in AI are now being used a broader way across cross multiple different fields. “When DeepMind started AlphaFold for protein language model, the focus was on life science and drug discovery,” he explains. “We can use that in the plant space just the same because the data and the mentality is the same. It’s just a different type of application.”
He says AI models are significantly accelerating the discovery and development of new crop protection products. AI, he explains, can rapidly search and analyse vast molecular libraries, scientific publications, and patent records to identify promising candidate molecules for new pesticides. This process, which traditionally took years, can now be completed in a fraction of the time.
On top of the time savings, rapidly advancing AI processes can also potentially make “better” products, he believes. AI feedback loops operate through a series of stages that resemble the rotational motion of a “crank”, he notes – a cyclical process of continuous improvement and adjustment.
“Once this feedback loop and the whole crank is put together, we are trying to teach the whole process that has been has embedded in years of practice into the data set so that all these models will learn this.”
Seed treatments
Ferrante is particularly excited about AI being deployed so that new crop developments can avoid the delays of long regulatory processes. “When we began, we were really excited about potentially helping with new GMOs to increase nutrition or make crops more water resistant,” he explains. GM crop development in the US, however, has a 13-year average time-to-market and $136 million cost. “The goal is to try and inject these properties in the coating in the treatment of the seed, because that’s a different process. And that’s something that can have all sorts of different downstream benefits for things like soil health.”
AI’s energy dilemma
It’s thanks to this process of continuous improvement and adjustment that the AI sector hopes to overcome another of its big challenges: its high energy use. Microsoft’s greenhouse gas emissions, for example, have increased around 30% compared to 2020, largely because of its prioritisation of AI development. Considering the ag industry wants to leverage AI largely to reduce environmental impact, “it is a concern, and people need to take into account”, admits Ferrante. But once you get the system “cranking”, the energy use is expected to flatten out, he says.
Zihan Wang, who heads up high-tech strategy and operations at NVIDIA, agrees. “If every agrochemical company can do virtual screening on our platform as opposed to the traditional way of computing, that means less experiments, less waste, and less energy consumed in the experiment. That is a kind of additional of value we’re bringing here.”
Johnny Israeli and Daniel Ferrante are scheduled to speak at the AI in Agriculture Forum at the World Agri-Tech Innovation Summit in San Francisco in March.
