AI is highly applicable and compatible with drug discovery methods. Artificial intelligence (AI) and machine learning (ML) are two techniques that are quickly being integrated into various research practices, including drug discovery and development (1). AI refers to machines or systems, typically software, that mimics human thinking to perform tasks while ML is a facet of AI that allows machines/systems to collect data and learn from tasks. As the foundational knowledge of diseases grows in size and complexity, so do the data sets, particularly genetic data which is one of the major sources of insights in recent years. Thus, AI approaches help analyze large amounts of data, recognize patterns, and most importantly, identify targets and compounds (1). Compounds can also be generated from scratch by AI. It can help limit the number of compounds synthesized for drug development and help reduce the design cycle by virtually tinkering with structural modifications prior to committing to them by chemical means. In fact, AI is estimated to reduce costs by 70% in the drug discovery phase through these efforts (2).
Case study #1: Repurposing old drugs is made easier with AI. The COVID-19 pandemic led to a renaissance of repurposing drugs and one company made notable contributions through AI efforts. In early 2020, BenevolentAI used an algorithm to scour over 50 million journal articles to find drugs that can prevent severe illness due to the SARS-CoV-2 virus. Within four days, they had a list of candidates which included baricitinib, a drug usually used to treat rheumatoid arthritis (3). Baricitinib was validated in subsequent studies and now it is a treatment approved by the FDA for certain patients and is backed by the Word Health Organization as a treatment option for severe cases of COVID-19. This two-year approval timeline is highly desirable and demonstrates the transformative nature of AI in providing new therapies, particularly against immediate public health threats.
AI’s incorporation into the later phases of R&D can also be beneficial. AI’s predictive capabilities can provide insights into how a drug may interact with a target even before cell and animal studies. These insights can help make informed decisions on how these studies will be conducted by providing estimates for dosing and pharmacokinetic properties. AI can aid in the management and analysis of data attained through clinical trials and post-market research (2).
AI has limitations. The accuracy of AI predictions will largely depend on the data that is available. When there is little data on a particular disease, demographic or factor, the results will be biased towards what that data set is enriched in. For example, most publicly available genomic data comes from individuals of European descent. In the realm of diagnostics, this has already led to poor predictions in terms of diagnosis for minority groups. In terms of drug development, genetic mutations may be unique to different demographics and less-represented demographics will not be able to benefit from AI approaches. Likewise, there are less studies on rare diseases and these data sets may also be limited. The availability of data and AI software also may be limited as they might be considered intellectual property (2).
AI cannot replace human oversight. In addition to intellectual property rights, data, particularly genomic data can be considered a private asset. Scientists must be wary of violating patient privacy while using AI. Furthermore, efforts to translate specific knowledge into code have been challenging. For example, AI can design a drug, but it cannot provide the steps and resources required to synthesize it (2). Chemists will still play a vital role in making new therapies part of reality. Finally, intentions and the goals of AI research lie with the humans involved as AI algorithms can be exploited for maleficence.
Case study# 2: AI can be applied to identify and produce harmful compounds. Currently, AI projects in R&D are centered around developing new therapies faster and more cost effectively. However, in the wrong hands, approaches can be reversed to instead create dangerous compounds. This theory was tested by a team at Collaborations Pharmaceuticals using a machine learning algorithm originally designed to assess toxicities. Within 6 hours, they produced over 40,000 compounds that can do harm (4). The list also included known biochemical warfare agents such as venomous agent (VX), a banned nerve agent. More alarmingly, the inverted algorithm put forth deadly compounds that bore no similarity to known agents but had predicted toxicity that far surpassed known agents as well.
Ultimately AI has already begun revolutionizing the drug development process and will continue to do so but like any emerging technology it is not without its risks and limitations. Researchers must take great care in the data sets they feed their AI, as well as their own intents to meet all public health needs, including for the treatment of rare conditions, despite how lucrative conventional and well-established drug targets may be (and bioweapons for that matter). Keeping privacy and equity in mind can enable new insights and pave the way for new treatments in the near future.
________________________________________________________________________________________________________________________
- Bender, A. & Cortés-Ciriano, I. Artificial intelligence in drug discovery: what is realistic, what are illusions? Part 1: Ways to make an impact, and why we are not there yet. Drug Discovery Today 26, 511–524 (2021).
- Insider Intelligence. Big pharma is using AI and machine learning in drug discovery and development to save lives. https://www.insiderintelligence.com/insights/ai-machine-learning-in-drug-discovery-development/ (2022).
- Richardson, P. et al. Baricitinib as potential treatment for 2019-nCoV acute respiratory disease. The Lancet 395, e30–e31 (2020).
- Cardoso -Founder, J. Financial Times. The dark side of using AI to design drugs. https://joaquimcardoso.blog/the-dark-side-of-using-ai-to-design-drugs/ (2022).