The idea to use computers as supporting tools in drug discovery and design is not new. Already in 1962 chemists were developing so called quantitative structure–activity relationship (QSAR) techniques, which could be used for drug design. Those techniques were using early machine learning algorithms to identify suitable molecules for the desired application scenario.

Today more potent AI methods can be used such as Generative Adversarial Networks (GANs). Such generative AI models identify suitable solutions for the given chemical problem by letting two neural networks produce solutions and compare the generated results. One of the networks, the generative network, is constantly trying to improve the quality of its solutions, while the other discriminative network is trying to identify, which solutions are artificially created and which are from a training data set. In the process of training the AI both networks are constantly evolving and improving. An example of such a model is druGAN.

A further machine learning method, which can be applied in generative chemistry is Reinforcement Learning (RL). Here, the AI is trained to make better decisions in decision making processes by getting a reward for good decisions. In practice the AI agent is interacting with the environment and gets a feedback about the interaction. Depending on the positive or negative nature of the feedback, the AI modifies its behaviour and becomes better in reaching the defined goal. Using this method generative AI can be trained for example to become better in discovering very specific chemical compounds and structures. An example is ReLeaSE (Reinforcement Learning for Structural Evolution).

Those AI model are no longer only in the theoretical testing stage, but are also practically used in the industry. One example is the collaboration between the chemical company Evonik and the leader in AI, IBM Watson. Here the focus of the use of AI lies exactly on the development process for new chemicals and should help with the optimization and acceleration of the process to find desired solutions quicker and at lower costs.

This improvement is created by substituting the old process of finding chemicals with the desired properties by trial and error with a predictive AI, which can identify chemical compounds with ideal properties much faster. In the specific case of Evonik and IBM AI-based image processing is used to generate the needed novel materials asked for by customers, which helps the company to stay competitive in the chemical industry. The image processing AI architecture is used here to detect the patterns in a material science application, just like it would find similarities between photographs.

While the development and discovery of chemical compounds with AI is now already an established and very vibrant field, from the chemical companies’ perspective many uncertainties around patents for generative AI solutions still remain, which need to be answered to secure the return on investment for the discovery of new chemical substances with AI. Answers on the most pressing questions around AI patenting in the chemical industry will be answered by Laura Fè, head of the Munich office of Murgitroyd, at the CEIPI Executive IP management days in Strasbourg. Here you can already watch an interview with her on the topic: