Today weed resistance affects more than 250 weed species infesting nearly 100 crops. As a result, weed management costs have doubled for many farmers, and we’re overusing a dwindling number of herbicides at ever higher rates. The root of the problem is the difficulty in finding new molecules which can selectively control weed plants via novel mode of actions (MoAs). We therefore need to discover new improved herbicides with novel MoAs using innovative tools and techniques so we can keep our crops safe.
A good herbicide molecule should have the following characteristics:
- High Selectivity: It should control specific weeds while being safe for the crops.
- Safe: It should have minimal to no toxicity to humans, animals, and non-target plants.
- Environmental Safety: It should break down quickly in the environment to benign substances.
- Resistance Management: It should help manage or reduce the risk of weed resistance.
- Economic Viability: It should be cost-effective for farmers to use.
Like drug discovery, finding a good herbicide molecule with a novel MoAs, is often compared to finding a needle in a haystack and is an enormous complex challenge. In this analogy, the haystack consists of order of 1060 to 10100 synthetically feasible molecules [1]. For context there are 1022 to 1024 stars in the observable universe. This is a gigantic chemical space in which to search and find a novel chemical compound that satisfies the plethora of desirable characteristics listed above. In addition, currently only a fraction of this chemical space has been synthesized, let alone tested on weed plants in an efficient data informative way. State of the art Artificial Intelligence (AI) algorithms offers a possible solution, by designing and assessing huge numbers of molecules virtually, thereby reducing the chemical search space into something more manageable for the search of the needle.
Artificial neural network (ANN) has been successfully applied in the process of searching for novel molecules in drug discovery. In fact, a type of ANN known as Recurrent Neural Network (RNN) have been used successfully for tasks such as natural language processing, music and painting generation. Generative Pre-trained Transformer (GPT) which is the adaptation of RNN for natural language processing is employed to understand and generate human-like text based on the input it receives. One such example is the well known ChatGPT application.
Similarly, there are multiple examples where this kind of adaptation has been applied to enable successful de novo design of novel drug molecules. However, it should be noted success is dependent on the availability of high-quality data.
Like ChatGPT which has been trained on huge and diverse dataset of text such as books, articles, websites etc, a model for herbicide can be trained to generate novel herbicide using public and inhouse chemical data. Ideally this data would be the chemical structures of molecules which has been tested on known species of weed plants in the lab and glass house and shown to control/inhibit their growth. Given the complexity involved in such task, Transfer learning (TL) strategy is used commonly. Transfer learning is a technique in machine learning (ML) in which knowledge gained from one task is utilized to enhance performance on a related task. For example, we can first train the model on large herbicide-like or drug-like molecule structure to generate novel and valid chemical molecules to get an initial model. This model (also known as Prior) can then further fine-tune using a small but focused set of chemical compounds which has shown to inhibit weed plants. Effectively, this gives us a model which has learned to generate not only valid but novel molecules with potential herbicidal activity. The strategy is depicted in Figure 2 as published here [2] where Prior (the initial model) is trained on ChEMBL data (a manually curated chemical database of bioactive molecules) [3] followed by two stages of fine tuning, first with molecules shown to control weeds in our glass house experiments and then a finally on a focused ‘project’ specific dataset. Since the training of the model is done in stages, this is also known as ‘staged learning’.
At Moa we have implemented a staged learning strategy to train Generative AI models using our unique high quality data sets. These models are routinely used to design novel herbicide molecules belonging to specific chemical classes using data associated with their respective projects. By integrating generative AI alongside traditional design methods not only are we able to expand our search space but also generate new molecules with desirable herbicidal properties within significantly reduced timeframes. Preliminary results show that the model can design compounds which inhibits growth of weed plants in our early lab tests with a hit rate of around 24%.
Given the complexity and low success rate of discovering new herbicides, along with the rising costs of managing resistant weed species, we must employ every possible technique to maximize our chances of success. By adopting Generative AI techniques, we can increase the chance of finding new and better herbicide molecules. Ultimately reducing the burden on the farming community and ensuring future food safety.
References:
- Schneider, G., Fechner, U.: Computer-based de novo design of drug-like molecules. Nat Rev Drug Discov 4(8), 649{663 (2005). doi:10.1038/nrd1799
- Loeffler, H.H., He, J., Tibo, A. et al. Reinvent 4: Modern AI–driven generative molecule design. J Cheminform 16, 20 (2024). https://doi.org/10.1186/s13321-024-00812-5
- Anna Gaulton, Louisa J. Bellis, A. Patricia Bento et al. ChEMBL: a large-scale bioactivity database for drug discovery, Nucleic Acids Research, Volume 40, Issue D1, 1 January 2012, Pages D1100–D1107, https://doi.org/10.1093/nar/gkr777