Symposium AI in Chemistry

In de novo drug design new chemical compounds are generated by deep learning algorithms that work quite well. However, these algorithms result in many candidates, which means we need to prioritize which molecules we want to test first.

Interdisciplinary work from Leiden University showed that we can use AI to assess whether molecules can be synthesized quickly and to predict if they actually bind to their intended target, thus increasing the efficiency of de novo drug design.

AI and ML technologies are demonstrating to be powerful in many application fields, and the number of successful projects in science is increasing.

In this talk we will describe how CAS leverages different techniques and combinations thereof to optimize internal workflows and enhance its products, discussing advantages of mixed approaches to obtain highest quality.

Nouryon conducts smarter experimentation with its formulations by using software to store the data generated in experiments. Within the software, the data is processed using AI/machine learning and this provides data-driven insights. Customers requesting new formulations, provide Nouryon with the appropriate parameters.

The AI engine proposes the most optimal formulation. This enables Nouryon to respond very quickly to customers request with the best possible solution with just one or a few experiments. Before AI, many experiments were needed to arrive at the best possible formulation, creating longer time to solution and wasteful experiments.

The quick rise of Artificial Intelligence has spread to all facets of our society. Therefore, also drug discovery is changing where the influence and catalytic effect of AI cannot be denied. While the application of computational chemistry and computer aided drug discovery long predates the current AI era [1], recently advanced new tools have become available.

Central to drug discovery in the public domain are large database which provide (ideally literature obtained) bioactivity data for a large group of (protein) targets and chemical structures [2,3]. Machine learning and Large Language Models support and accelerate the drug discovery process. Often these methods exist at the intersect of the fields of chemistry, biology, and informatics.

In this talk I will give an overview of research going on at the computational drug discovery group in Leiden. Central in our research is the usage of machine learning and the combination of chemical and biological information. I will highlight some examples we have published previously and finish with an outlook of cool new possibilities just around the corner [4–9].

Note that quite a few of our tools are on our GitHub for download: https://urldefense.com/v3/__https://github.com/CDDLeiden__;!!H1puPpzZ!WIPEx_SO2c7-94nkJuRF9uZzyrktcnLu8ezRAu4Fxec11kV85dagaSnwDO43UV4umjkv_i5VsgDFGnM2KmHFWbaIuAzO8bWnfB6TtSA$

Andrius Bernatavicius, PhD Candidate LACDR/LIACS, Leiden University

Recent advances in deep learning are rapidly expanding the frontiers of virtual screening and de novo drug design. Here, we present PCMol - a GPT-like generative

model that uses the internal protein representations of AlphaFold for conditioning the model to generate novel compounds for thousands of different targets. We illustrate the effectiveness of this approach by investigating the properties of generated molecules and comparing it to other existing methods.

The code is available at: https://github.com/CDDLeiden/PCMol

Preprint: https://chemrxiv.org/engage/chemrxiv/article-details/65d47632e9ebbb4db9c63988 

When building Generative AI solutions for R&D in the chemical domain, we quickly find that LLMs only have very limited expertise. Retrieval Augmented Generation (RAG) has recently emerged as the core method to adapt Large Language Models (LLMs) for Question Answering and Chat grounded in private data sets. We have developed a synthetic data approach for finetuning the neural search in RAG to specific domains and customer data sets, with relevance improvements of over 200% in some use cases. We will discuss work on adapting Gen AI pipelines based on open data from ChemRxiv, as well as more recent work to build RAG agents for high-recall question answering.

AI depends on machine learning, and learning requires access to data and knowledge to learn from. Learning is done with both positive and negative input, because we need to understand the boundaries of what we are interested in. Second, when we find contradictions, we need to be able to learn where that is coming from, and therefore, the provenance of our knowledge and data is essential too.

This talk will focus on the use of semantic web technologies like Resource Description Framework and ontologies in chemistry, how we can (and should) improve the current dissemination approaches, and how this can be used in AI in chemistry. This story will be told from a cheminformatics perspective, vuilding on more than two decades of open science experience.