1. Repurposing existing drug candidates
The use of AI in drug discovery so far is perhaps best demonstrated in drug repurposing, where AI can rapidly identify alternative indications for existing molecules. Over 250 companies are currently working on repurposing drugs through AI, with COVID-19 having provided a unique opportunity to apply this quick and flexible approach to drug discovery.
Baricitinib, a Janus kinase inhibitor for rheumatoid arthritis, was identified as a potential treatment for COVID-19 by BenevolentAI using their knowledge graph platform. It received emergency use authorization from the U.S. FDA in 2020 for treatment of COVID-19 in hospitalised patients followed by full approval in 2022, based on the results of four randomised clinical trials.
Most compounds identified by AI repurposing approaches are still under evaluation in clinical trials. The outlook is positive: data availability and quality are expected to improve as data becomes more diverse and accessible. That should fuel efforts in this space.
2. Drug target identification
AI techniques can rapidly build molecular disease models and much more efficiently identify druggable targets and biomarkers than traditional methods. An enormous volume of biomedical data is available, although the integration of multiple unstructured datasets is challenging. AI can be employed to extract and analyse findings from unstructured datasets, such as journal articles and omics databases as well as imaging and real-world patient data. Knowledge graphs are used to identify novel connections between entities, although the capabilities of these approaches are limited by the quality of standardisation and labelling of underlying datasets.
Initial programs using AI in drug target identification have moved through discovery and preclinical development, and at least 20 drugs with novel disease-target associations identified by AI are progressing through phase 1 and 2 studies. As companies expand datasets and feed findings back into AI algorithms, increasing numbers of drugs with novel disease-target associations — or entirely novel targets — are expected to emerge.
3. Small molecule drug design
Using available chemical structure data, AI can simulate complex chemical properties or enable the design of drug structures significantly faster and more accurately than traditional methods. Within this use case, companies can use AI to screen existing chemical libraries or to generate novel chemical designs. The availability and usability of underlying datasets remain key challenges in this use case, with training sets being comparatively small compared with the full chemical space of billions of compounds.
In addition, data availability varies across different target classes, with kinases and G protein-coupled receptors being the most well characterised, which limits generalisable models and the novelty of resulting drug candidates.
Individual AI-driven tools are already an integral part of the drug design process for small molecules, with larger predictive solutions undergoing iterative development. Small molecules designed using AI are significantly more common than antibodies designed using AI at this point. Clinical programs from companies such as Exscientia and Insilico Medicine are part of the first wave of AI-designed small molecule drugs undergoing phase 2 trials, the results of which are likely to begin to illustrate the maturity and future potential of this use case.
4. Antibody drug design
Antibody design is a growing use case for AI through both optimisation of existing structures and de novo candidate design. To date, few AI-designed antibodies have reached the clinic, and the more complex nature of these molecules poses distinct challenges compared with small molecule drug design — such as with the computational capabilities required to run larger models. AI models for antibody design are also limited by the availability of datasets for antibody sequences and antibody-antigen pairs. In addition, with a large proportion of training data being derived from the same libraries used for traditional antibody design approaches, many of the traditional challenges, such as balancing specificity and affinity, persist.
The ecosystem of researchers and companies focused on AI for antibody drug design is growing, with a flurry of announcements from large pharma companies over the past year disclosing innovative internal capabilities or collaborations with start-ups or big tech.
Most recently, more than US$1 billion was secured by Xaira Therapeutics, which plans to initially focus on de novo antibodies, having employed researchers with experience designing leading diffusion models for protein and antibody design alongside genomics and proteomics groups. Continued collaborations between AI platforms and pharma companies, as well as an increase in standardised and open source data, are expected to grow the maturity of this use case.
Outlook: Maturity of AI across use cases
AI maturity in life sciences in a diverse landscape and varies across use cases (see FIgure 4). While applications like repurposing existing drug candidates and target identification have made significant strides, others like antibody drug design are still in the relatively early stages.
