Is there a single AI model that can accurately predict whether a molecule would be a good drug against a disease? We explored this question in our recent publication, "Current Methods for Drug Property Prediction in the Real World”. Our insights have, since then, driven the development of DeepMirror's state-of-the-art drug discovery platform.
The motivation
Drug discovery is a lengthy and costly process, where the pre-clinical development alone costs an average of ~£35-55M per drug and takes >8yrs (Wellcome Trust, 2023). To make things worse in our fight against disease, only 1 in 10,000 drug candidates are FDA approved after clinical trials (Boyd et al., 2021). AI has burst into the space promising to revolutionise the speed and cost at which drugs are developed (DeepMirror, Jan 2023, Jayatunga et al., 2022, Nature biopharma dealmakers News, May 2021).
A wealth of diverse AI approaches have been published and tested on predicting properties of small molecule drugs. Generally, these publications are centred on evaluating just one or at most a handful of AI methods.
As a medicinal or computational chemist trying to solve a specific property prediction problem, how can you choose which AI approach to take when there are multiple available methods and evaluation of performance is spread across a growing list of publications, which are hard or very time consuming to reproduce? Is there a specific method for molecular properties that can “fit-them-all”, or should you keep track of what specific methods works best for every type of property prediction?
These were the questions we set out to answer in our recent publication (Green et al., 2023). Our goal was to compare many different AI methods in a detailed and unbiased way, to find out the most effective ones for specific drug properties, aiming to make drug discovery more efficient and less expensive.
Molecular properties take center stage
In drug development, predicting the properties of drugs (small molecules) before testing them in the laboratory is crucial to reduce the time and resources required to bring safe and effective new drugs to patients. Two main types of property predictions are crucial: properties that describe how “drug-like” a molecule is, such as it’s absorption, it’s distribution in the body, how it gets removed from the body and how toxic it is; and properties that describe how good a drug is at binding it’s target and exerting an effect against a disease (affinity, potency).
In total, we tested 184 AI approaches against 44 public datasets.
The AI leader board
Overall, our research highlighted the need for different AI approaches for different datasets. For example, we found that (somewhat unsurprisingly!) traditional methods perform better for low dataset sizes and datasets with affinity measurements, whereas modern AI methods (such as deep learning) perform better at higher dataset sizes and some drug-like property datasets.
We found that selecting the right featuriser was also dataset dependent. Featurisers are methods that turn molecular structures into a numerical format that computers can understand. Expert features (properties derived by cheminformaticians) worked best for affinity property datasets; yet molecular descriptors (chemical properties of a molecule) and Natural Language Processing (features derived from letter sequences such as molecules SMILES) worked best for drug-like property datasets.
All in all, we did not find a single model “to rule them all” in small molecule drug property prediction. Our study emphasizes the need for a comprehensive methodology when building AI for predicting drug properties, that compares many methods and selects the best for each case. By benchmarking a wide range of AI methods across various datasets, we have developed a platform that can intelligently adapt to the specific needs of each dataset. This adaptability is critical in a field as diverse and complex as drug discovery.
“Our work highlights that practitioners do not yet have a straightforward, black-box procedure to rely on, and sets the precedent for creating practitioner-relevant benchmarks.” (Green et al., 2023).
Making it accessible to everyone
How can one access AI for drug discovery? Companies could hire teams that develop bespoke models for their prediction needs. But this is costly and one often does not know whether these models are useful until after many months of development.
This is why we decided to make it simple for companies to get started with AI for drug property predictions by building an easy-to-use cloud app for drug design. On the app, users can quickly use thousands of models and test whether the best automatically selected AI method works for their problem within days instead of months.
Want to try it out yourself? Request Early Access here.
References
Boyd, N.K., Teng, C., Frei, C.R., 2021. Brief Overview of Approaches and Challenges in New Antibiotic Development: A Focus On Drug Repurposing. Front Cell Infect Microbiol 11, 684515. https://doi.org/10.3389/fcimb.2021.684515
Green, J., Cabrera Diaz, C., Maximilian H Jakobs, Dimitracopoulos, A., Mark van der Wilk, Greenhalgh, R.D., 2023. Current Methods for Drug Property Prediction in the Real World. https://arxiv.org/abs/2309.17161
Jayatunga, M.K.P., Xie, W., Ruder, L., Schulze, U., Meier, C., 2022. AI in small-molecule drug discovery: a coming wave? Nat Rev Drug Discov 21, 175–176. https://doi.org/10.1038/D41573-022-00025-1
The Path to AI-driven Drug Discovery - Part 2: How is AI disrupting Drug Discovery? DeepMirror January 2023 https://www.deepmirror.ai/post/the-path-to-ai-driven-drug-discovery-part-2-how-is-ai-disrupting-drug-discovery
Nature biopharma dealmakers News, May 2021: https://www.nature.com/articles/d43747-021-00045-7
Unlocking the Potential of AI in Drug Discovery, Welcome Trust, 2023: biopharma: https://wellcome.org/reports/unlocking-potential-ai-drug-discovery
コメント