Artificial intelligence (AI) has the potential to radically change the way drugs are discovered, engineered, and tested. While the mathematics that enable AI and machine learning (ML) have been around for quite some time, in some cases over a thousand years, our current computational ability to collect, store, process, and analyze multidimensional complex data has enabled a burst of AI applications. This AI explosion has accelerated drug discovery and is poised to translate into impactful therapeutic advances within urology and the field of medicine in general.

When AI is discussed in medicine, what is most commonly being referred to is some iteration of ML. ML can be broadly divided into supervised and unsupervised ML. Whereas supervised learning incorporates a known outcome (such as tumor size or response to therapy) into its prediction, unsupervised learning seeks to find patterns in data where there isn’t a clear outcome that is attempting to be predicted. Additionally, “reinforced ML” utilizes a strategy for sequential decision-making that is learned from known data. These concepts are reviewed in depth elsewhere,^1–3 but these basic details will allow the understanding of some examples where AI is being used in the drug discovery process to accelerate the critical tasks of drug target identification, drug engineering, and optimization of clinical trials.

First, AI has the potential to accelerate drug target identification through precise modeling of protein structure. Our understanding of protein structure was recently transformed through public release of the AlphaFold database by a collaboration between EMBL’s European Bioinformatics Institute and Google’s DeepMind group. Using ML, this database predicts a protein’s 3D structure from its amino acid sequence with very high accuracy, allowing the scientific community access to over 200 million protein structure predictions, most of which were previously unknown. This information will certainly accelerate the process of drug molecule design and identification of lead compounds and druggable protein binding pockets, a task which previously required time-consuming protein crystallization.

This ML-generated information on protein structure and target identification can be used to design and evaluate potential drug targets either de novo or through repurposing of known chemical entities. The number of all possible small molecule drugs in the universe is extremely vast (one estimate puts the number at 10⁶⁰ potential molecules).⁴ Given this amount of complexity, even a very high throughput synthesis and screening program can only explore a miniscule portion of all potential molecular iterations. However, ML can be used as a solution at multiple points in this pipeline to triage and invent the most likely drug candidates, saving both time and resources. Pharmaceutical companies and academic institutions are currently utilizing AI to develop predictive algorithms of reactions needed to synthesize complex molecules. This in silico development using ML allows drug makers to reverse engineer molecular structures based on desired mechanism of action from other molecules that may have similar physical parameters. ML algorithms can also use known portions of molecular structures to form predictions about the properties of a compound. This process not only assists in predicting potential toxicities and mechanisms of action but also can be used to determine the feasibly of actually making a molecule using a process called retrosynthesis. Finally, combinations of molecules can be queried to identify potential synergistic mechanisms. Within genitourinary malignancies, the ability to computationally predict which previously approved therapies may act synergistically to more effectively treat cancers can identify novel combinations that wouldn’t have been discovered otherwise.

Beyond the use of AI in small molecule design, ML is also being used to design even more complex therapeutic molecules that have not been previously discovered. With increasing antibiotic resistance associated with pathogens causing UTIs, development and identification of new, effective antibiotics could significantly improve treatment of recurrent UTIs. For example, researchers at MIT (Massachusetts Institute of Technology) utilized AI to identify a new antibiotic for treatment of drug-resistant E coli.⁵ They developed an ML algorithm to screen more than 100 million compounds in a few days, a task that was previously prohibitive by time and cost.⁵ They identified a compound that was predicted to have (1) powerful antibiotic activity, (2) a different chemical structure than existing antibiotics (to avoid resistance), and (3) low toxicity to human cells. This compound, halicin, also successfully treated Clostridium difficile, Acinetobacter baumannii, and Mycobacterium tuberculosis, among others.⁵

Finally, ML can be applied to improve the success rates of oncology clinical trials, which often have high failure rates. ML can be used at various levels of clinical trials to help them run more effectively by designing more precise cohort compositions, enhancing patient recruitment, and improving patient monitoring.⁶ Additionally, trials evaluating targeted therapies are becoming more common and often require molecular characterization of tumors to match patients with appropriate targeted therapies. ML approaches can be applied to identify molecular alterations in novel ways. For example, one study trained a deep learning network to detect FGFR3 mutations from digitized hematoxylin and eosin slides of bladder cancer tumors to predict which patients had these mutations. Given the success of erdafitinib, an FGFR3 inhibitor, in bladder cancer, this approach could allow for targeted therapy without the expense and time of full tumor sequencing.⁷

As AI permeates the drug discovery process, we must be cognizant to identify and mitigate potential ethical considerations that arise. One concern is the potential for introducing bias in terms of current health inequalities. Care must be taken to curate and consistently review training datasets used for AI so that health disparities are not exacerbated by these new technologies. Additionally, ML modeling and predictions often rely on structured, large-scale data which require computational support and, in the case of patient data, accurate cataloging and confidentiality protections. This will require development of new regulations for oversight and consistent monitoring of data in an open and transparent manner.

Despite challenges that arise with the rapid integration of AI in medicine, we are optimistic that the promise of more efficient and rapid drug development will lead to valuable advances for patients and serve as a catalyst to transform the way we deliver medicine.