Weeds are a major challenge in agriculture: competing for critical resources and killing crops. While herbicides have been pivotal in weed control in the last fifty years, their overuse has led to herbicide resistance in 273[1] species globally, with annual losses to herbicide resistant blackgrass in the UK estimated to be $500million[2] , and annual weed losses of $3.3billion in Australia and $11billion in the USA[3].
As the world’s population grows and demand for food increases, so too does the need for new herbicide modes of action that can overcome these resistant “superweeds” and help farmers protect their crops.
At Moa, new mode of action herbicide candidates are discovered using high-throughput screening of chemical libraries in GALAXY followed by extensive profiling in glasshouse. Profiling of chemistries in the glasshouse is not only labour-intensive but can also be subjective, relying on the expertise and the judgment of the person analysing the growth and damage to plants.
Through the Knowledge Transfer Partnership (KTP) program sponsored by UKRI and in collaboration with the Department of Computer Science in Loughborough University, we have been harnessing deep learning to automate chemical library phytotoxicity testing process. Specifically, I have developed an advanced AI-driven solution to analyse glasshouse imagery, improving the accuracy, efficiency and scalability of herbicide activity assessment.
To achieve this, our collaboration started by collecting image data from our glasshouse testing, capturing images of control and treated plant samples across multiple weed species. Each image was then manually annotated to specify the plant species and identify visible leaf matter, creating an adequate dataset for deep learning model training.
We then used these annotations to build object localisation models, enabling automated identification of plant species and precise detection of plant matter in both control and treated samples. By analysing these paired images contrastively, our approach quantifies relative growth differences, allowing a statistical estimation of each compound’s efficacy with improved speed and stability. This process was iteratively refined to optimise model performance by expanding the dataset, including more species, and analysing relationships between model predictions and manual human evaluations.
This new AI tool scored an impressive 95% accuracy when evaluated against real glasshouse data, demonstrating robustness across different levels of herbicide activity and species.
Our AI tool is now being integrated into Moa’s glasshouse pipelines, to give scientists in a standardised, scalable, reproducible process, reducing manual labour and improving the overall quality of assessment.
What next?
AI cannot find and develop new, safe and effective herbicides all on its own, out of nowhere. It needs data: the higher quality the better. Moa has a huge bank of data, from the 750,000+ compounds our platforms have already screened. So now we are starting to use this new tool to help clean up and interrogate our data: re-analysing archived experiments, identifying potential misclassifications and recovering previously overlooked compounds. This will help us – and AI – to make better decisions.
This project was made possible by a Knowledge Transfer Partnership between Moa Technology and the Department of Computer Science at Loughborough University. The author would like to thank Professor Baihua Li, Dr Haibin Cai, and Professor Qingqing Meng at Loughborough and Dr Laura Davies and Dr Nasir Rajabi at Moa Technology for their collaboration on the project.
The results of the collaboration will be presented at VISAPP 2025 in Porto, Portugal and have been submitted for publication in Machine Learning with Applications journal.
[1] Heap, I. (2024). The International Herbicide-Resistant Weed Database. Retrieved from https://www.weedscience.org.
[2] Varah, A. et al. (2020). The costs of human-induced evolution in agriculture. Nat. Sustain., 3(1), 63–71.
[3] Chauhan, B. S. (2020). Grand challenges in weed management. Frontiers in Agronomy, 1, 3.