A practical guide to implementing AI
In this ebook, you’ll discover the key considerations which every leader needs to take in order to successfully implement AI in their drug discovery pipelines.
Imagine you’re trying to find the correct key to unlock a treasure chest, but there are billions of keys to choose from. You don’t know which one will work, and all you can do is test them endlessly one by one. This is what drug discovery has traditionally been like for scientists – e.g. with high-throughput screening – where you test everything you can and see what sticks. It’s a painstaking and expensive process that could take years often ends in disappointment.
The emergence of Artificial Intelligence (AI) arrives with revolutionary power in drug discovery – it doesn’t just promise to speed up the process but also transform it. People are realising the potential of AI to reduce time, money, and resources. In chemistry alone, from generating new molecules tailored to specific needs (generative chemistry) to mapping out the most efficient ways to create them (synthesis planning) and predicting their performance realistically (compound property prediction), AI is accelerating every step of the drug discovery pipeline.
In this blog, we’ll dive into these three impactful areas and discover the role of AI in drug discovery chemistry.
Generative chemistry – AI-powered molecule design
Exploring and targeting the correct active compounds with suitable pharmacokinetic properties in vast chemical space is challenging. Early generative chemistry often provided mixed-quality suggestions. Chemists were forced to spend hours filtering through suggestions of unstable, inappropriate or synthetically complex molecules to find their useful series and chemistries.
Numerous generative chemistry methods are available, from classical models, where medicinal chemistry transformations are iteratively applied, to AI and machine learning approaches such as auto-encoders. The latest generative chemistry methods tend to use transformer models, the foundation of large language models (LLMs), such as ChatGPT. They are quicker, cheaper to train and often very powerful compared to other methods – they can work with larger datasets. These models enable scientists to generate a broader variety of chemical structures with greater confidence in their synthetic accessibility.
The best methods in generative chemistry are those which harness the user’s expert knowledge, in a concept which we call Augmented Chemistry®. By combining AI’s ability to explore diverse strategies without bias with the intuition and experience of expert scientists, we can unlock new possibilities in drug discovery and development. To learn more about augmenting inspiration with generative chemistry, read Matt’s IBI article.
For a practical walkthrough of what generative chemistry is, where it fits in the discovery workflow, and the best practices to avoid common pitfalls, download our free ebook.
Synthesis planning – AI as the master molecular chef
Designing a molecule is one thing, but figuring out how to actually make it is another challenge. This is where AI steps in with synthesis planning – creating a recipe for how to build a molecule.
Traditionally, scientists relied heavily on their expertise and long literature searches to figure out the best way to synthesise a compound by trial and error. Now, AI emerges as the hero in this area, not only collating literature much more quickly but also predicting synthetic routes in minutes. Some AI methods even integrate direct links to vendor libraries – with all the information about stock, price, and delivery times – allowing synthesis to begin as soon as possible.
This is where our collaboration with Chemical.AI comes in. If you license both StarDrop and ChemAIRS®, you can connect to ChemAIRS directly from StarDrop and run Chemical.AI’s synthetic accessibility scoring and retrosynthetic route analysis without leaving your workflow. Run a quick accessibility score across a dataset to see at a glance how easy each compound is to make, helping you prioritise which ones to take forward. Then, once you’ve identified your leads, a retrosynthetic search lays out the possible recipes for each – every route labelled with the number of steps, a difficulty score, and an estimated cost per gram to synthesise – so you can weigh your options and choose the most practical path to the bench. And it goes beyond planning a single synthesis: analysing these routes can spark new ideas for your SAR, surfacing shared intermediates and related analogues worth exploring.
Nonetheless, the intricacies of identifying a feasible synthetic route can be tricky for current AIs to predict: Can AI propose novel synthetic routes beyond the reactions it was trained on? Are the suggested routes scalable in large-scale manufacturing? This field is evolving rapidly, with exciting research underway, and we anticipate significant advancements in the near future.
Compound property prediction – AI enables smarter and better decisions
AI can help you predict your compound’s properties – that’s great. But how does it compare to traditional models?
Whilst traditional QSAR methods rely on molecular descriptors, advanced deep-learning imputation approaches like Cerella(TM) can leverage existing experimental data to support predictions (see our recent blog on QSAR vs imputation). It can also tell you about the confidence in each prediction; which additional data you should obtain to support compound prioritisation; and which of your current experimental measured data points may require re-evaluation due to potential errors.
By providing information on uncertainty, such methods can guide you not only to the compounds with the best predicted properties but also to those which are most likely to succeed in your project. To see this in action, watch our webinar, Does your model know its limits? Leveraging uncertainties to find better compounds, where real-world case studies show how uncertainty-aware modelling helps you make better decisions and find compounds with a greater chance of success.
Beyond basic predictions, AI can identify and harness complex relationships in large, sparse and noisy data sets – common in many drug development organisations – to inform predictions that generate new insights and support your understanding of mechanisms of action or adverse outcome pathways. AI methods can identify the early-stage frequently measured endpoints that are most correlated with late-stage endpoints. By suggesting cheaper early-stage measurements that inform predictions of expensive late-stage outcomes, AI can guide your decisions on compound progression. This can enable you to minimise expensive late-stage experiments.
AI in chemistry – An ally, not a replacement
AI isn’t here to take over our jobs – it’s here to make them better. It is about embracing and accepting the role of AI in your drug discovery journey. If you feed your AI relevant data, it should recognise patterns, make suggestions and predictions, and help you guide compound, assay or experiment prioritisation more effectively. To unlock the full value AI can deliver, we should be open to adapting our workflows and rethinking how we integrate technology into our research. Read the blog.
Ready to transform to AI-guided drug discovery?
Simply complete the form and our team will be happy to chat about how Cerella could improve your discovery workflows.
About the author
Nathan Brown, PhD
Director of Science, Optibrium
Nathan Brown leads Optibrium’s Research and Application Science teams, developing new methods while ensuring customer feedback directly shapes R&D priorities. He was previously Director of Digital Chemistry at Healx, where he led the development of AI-powered computational methods for rare disease drug discovery, and before that led the Cheminformatics team at BenevolentAI and founded the In Silico Medicinal Chemistry team at The Institute of Cancer Research, delivering significant impact on drugs in active clinical trials. A globally recognised thought leader in cheminformatics and computational chemistry, Nathan is a Fellow of the Royal Society of Chemistry and sits on the Editorial Advisory Boards of the Journal of Chemical Information and Modeling and ChemMedChem. He invented the first multi-objective de novo molecular design system, published in 2004, and has authored over 50 peer-reviewed papers and four books. He won the 2017 Corwin Hansch Award for his outstanding contributions to the field.
More AI in Drug Discovery resources
How does generative chemistry work, and how can it help me?
The role of generative chemistry in drug discovery A key difficulty in finding new drugs is the sheer size of…
How much do AI drug discovery platforms cost?
Which AI platform do you need? The first thing you’ll need to do is decide what you need to achieve…
