Artificial Intelligence in Drug Discovery: A PRISMA-Based Systematic Review of Methods, Challenges, Taxonomy, and Evaluation Practices

Bringing a new medicine from a scientist's idea to a pharmacy shelf is a monumental undertaking. It typically takes a decade or more and costs billions of dollars, yet the vast majority of potential drugs fail before they ever reach a patient. Most of these failures happen because the drug does not work well enough or causes harmful side effects that only appear late in the testing process. To speed up this slow and expensive journey, researchers have turned to artificial intelligence. In this context, artificial intelligence refers to computer systems that can learn from data to find patterns and make predictions without being explicitly programmed with every single rule. One of the most powerful tools in this field is machine learning, where computers analyze massive amounts of information to improve their own performance over time. Another key concept is the drug discovery pipeline, which is the step-by-step process of finding a disease target, screening millions of chemical compounds, and refining the best candidates into safe medicines. For years, this process relied heavily on human intuition and slow, manual testing, but the sheer number of possible chemical combinations is so vast that human effort alone cannot explore them all.

A comprehensive new review of scientific literature has mapped out exactly how artificial intelligence is currently being used to tackle these challenges. The researchers behind this study, led by Mahmud Ahmed and colleagues, did not conduct a new experiment in a lab. Instead, they acted as master archivists, gathering and analyzing 618 separate scientific papers published between 2019 and 2026. They searched through three major databases containing millions of records, filtering out duplicates and irrelevant studies to focus only on high-quality research that applied artificial intelligence to drug discovery. Their goal was to cut through the noise of a rapidly growing field, where thousands of new papers appear every year, to understand which methods actually work, where the field is heading, and what problems remain unsolved. They wanted to create a clear picture of the landscape, organizing the chaotic mix of techniques into a structured framework that scientists and industry leaders can use to move forward.

The review revealed that the field has exploded in recent years. In 2019, there were only 11 relevant studies, but by 2026, that number had surged to 159, showing a fourteen-fold increase. The most common approach found in these studies is machine learning, which appeared in nearly every single paper reviewed. Within that broad category, deep learning—a type of artificial intelligence modeled after the human brain's neural networks—was used in about 78 percent of the studies. While older methods like quantitative structure-activity relationship modeling, which predicts how a molecule will behave based on its shape, are still widely used, newer technologies are rising fast. Graph neural networks, which treat molecules like maps of connected points to understand their structure, and transformer models, which are the same type of technology behind advanced language tools, have become dominant forces, particularly in papers published in the last two years. These newer tools are especially good at designing entirely new molecules from scratch rather than just testing existing ones.

The researchers organized these findings into a clear system, showing which artificial intelligence tools are best suited for each stage of the drug discovery process. The most frequent application is virtual screening, where computers rapidly test millions of chemical compounds to see which ones might bind to a disease target. This accounted for more than half of the studies. The second most common area is de novo design, where artificial intelligence invents new chemical structures that have never existed before. While these tools are powerful, the review found that the field is still struggling with how to prove they work reliably. A major issue is that many studies use only one type of measurement to judge success, often focusing on a single score that can be misleading when the data is unbalanced. For instance, if a dataset contains mostly inactive compounds, a model could appear highly accurate simply by guessing that everything is inactive, yet it would fail to find the few active drugs that matter. The review found that very few studies reported the specific metrics needed to catch these errors, making it difficult to compare different methods fairly.

Another significant finding is that while the computer models are becoming more sophisticated, the real-world application of these models is lagging behind. The biggest challenges identified were not technical failures of the algorithms themselves, but rather the difficulty of integrating these digital predictions into actual wet-lab experiments and clinical trials. Only a small fraction of the studies included in the review tested their models on completely new, independent data to see if they would work in a different setting. Furthermore, many of these powerful artificial intelligence systems are "black boxes," meaning that even the scientists who built them cannot fully explain why the computer made a specific prediction. This lack of transparency is a major hurdle for regulators, who require clear explanations before approving new medicines. The review also highlighted that the data used to train these models is often incomplete or biased, which can lead to models that work well in a computer simulation but fail when applied to real biological systems.

The geographic distribution of this research is also telling, with the majority of studies coming from China, the United States, and India. The review noted that India's contribution is larger than previously recognized, suggesting a growing hub of innovation in that region. Despite the rapid growth and the impressive volume of research, the authors conclude that the field is at a critical juncture. The technology has advanced to the point where it can generate novel drug candidates and predict their properties with high accuracy, but the infrastructure to validate, interpret, and safely deploy these candidates is not yet ready. The review suggests that the next major step is not just building bigger models, but creating systems that are transparent, reproducible, and capable of working across different types of data, such as combining genetic information with chemical structures. Until these gaps are filled, the promise of artificial intelligence to drastically shorten the time and cost of bringing new life-saving drugs to market will remain only partially realized.