Pharma’s use of Artificial Intelligence/Machine Learning (AI/ML) is growing quickly. The industry’s challenge now is to figure out how and when to incorporate these tools to bring value to different parts of the business cycle, from target identification to drug discovery and optimization, clinical trials, manufacturing, and even commercialization.
Artificial intelligence is a broad technology focused on enabling a machine to perform a particular task the way a human would, but much faster. Machine learning is a subset of AI that allows a machine to automatically improve its performance from past experience.
Some of the fastest-moving advances to date are coming from applications of AI/ML to research and development. Increasingly sophisticated AI algorithms have enabled the simultaneous exploration of hundreds of genomic sequences to discover unique drug targets, screen thousands of small molecules in silico and determine their pharmacological profiles, predict drug effects and potential toxicities, and create visual outputs on how drugs, targets and pathways interact. The result has increasingly been a deeper, better understanding of biological processes and outcomes, creating opportunities for streamlining the drug discovery process.
In November, for example, the team behind Google’s DeepMind AI platform described in Nature the use of its latest program, AlphaFold, to predict how proteins fold – a problem that researchers have struggled with for nearly 50 years. The researchers trained AlphaFold for several weeks on a public database of 170,000 proteins and their shapes; once trained, the platform was able to predict protein shapes in a few days and with similar accuracy to lab-based methods which have sometimes required decades of work. This new advance is expected to help decipher the function of thousands of unsolved proteins in the human genome, and make sense of disease-causing protein variants, leading to a wealth of potential new biomarkers and targets.
AI/ML is also greatly speeding drug screening and optimization. For example, in 2020 Exscientia became the first company to advance an AI-designed molecule to clinical trials – DSP-1181, a potential serotonin regulator with application in obsessive-compulsive disorders. By allowing an AI algorithm to generate and screen possible compounds against their chosen target, Exscientia advanced their candidate molecule to the clinic in 12 months, rather than the typical five years required from discovery to clinical trials. The company is now preparing clinical studies for a second AI-designed drug whose discovery and preclinical development has taken less than 14 months.
Pharmaceutical developers are also increasingly forming partnerships with AI companies. For example, Hyundai Pharmaceuticals recently teamed with Pharminogen to use the latter company’s AI platform for target discovery and the design of new drug candidates. The two organizations believe that by working together they can advance a new drug candidate to clinical trials within six months.
Clinical trials are further being improved through applications of AI. Large amounts of real-world data are generated daily from electronic medical records, scientific publications, health insurance claims, health apps, wearables, biometric devices, and even call center transcripts and social media, enabled by new natural language processing (NLP) algorithms that permit the rapid analysis of diverse text-based information. Applications of AI to this growing mountain of data will help improve disease characterization and the stratification of patient populations for clinical trials, as well as inform better study design and optimize patient enrollment and retention.
Even pharmaceutical commercial operations are beginning to benefit from AI, including manufacturing, supply chain, and sales and marketing activities. GlaxoSmithKline has opened an AI center in London. Among the many applications that the company is exploring is the use of AI models to help understand and predict consumer preferences pertaining to drug taste, texture, size, color, etc. GSK is also developing a model that can forecast how an upcoming cold and flu season might unfold in different regions, predicting when peaks and troughs of infection might occur. Such information could help the company advise consumers, inform advertising spend, and alert retailers on when to stock up.
Challenges remain to the widespread adoption of AI in pharma. The quality as well as quantity of data needed to effectively train an algorithm is important, with training sets requiring at least two to three years of data to be effective. In addition, the talent pool of potential employees with the right data science skills is currently limited. Issues regarding data privacy also remain to be addressed. Moreover, it is difficult as yet to provide the business value/return on investment for many of the potential uses of AI outside of speeding early-stage R&D. But it is clear that pharma’s use of AI is here to stay.