Closing the data loop in AI-driven drug discovery

The pharmaceutical industry is facing mounting pressure to accelerate drug discovery while reducing costs and improving success rates. With traditional methods proving slow and expensive, artificial intelligence (AI) is emerging as a key tool in transforming the process. According to recent reports, the cost of developing new drugs has nearly doubled every nine years since the 1950s, a trend known as Eroom’s Law. Today, the average cost of bringing a new drug to market ranges from $1 billion to $2.5 billion, and the entire process typically spans 10 to 15 years, with failure rates exceeding 90%. These figures underscore the urgent need for innovation in drug discovery. Paul Belcher, director of protein research strategy at Cytiva, notes that the primary financial burden in drug development lies in the clinical phase. As such, efforts to improve success rates and shorten timelines during this stage are crucial. AI is being viewed as a promising solution, offering the potential to streamline processes and enhance the quality of drug candidates before they even enter clinical trials. Belcher emphasizes that AI's ability to predict interactions between molecules and disease targets could significantly reduce the risk of costly failures later in development. One of the earliest and most promising applications of AI in drug discovery is in hit identification. This process involves screening vast libraries of molecular entities against specific disease targets, such as proteins, to identify molecules that can bind effectively. Traditionally, this was done through physical screening, which is both time-consuming and resource-intensive. However, AI is enabling a shift toward predictive design, allowing researchers to generate drug candidates from scratch and simulate their interactions with disease targets before investing in extensive research and development. Belcher describes this transition as moving from empirical screening to a more efficient, data-driven approach. By leveraging AI, drug companies are no longer constrained by the limitations of physical screening. Instead, they can explore a broader range of molecular possibilities and eliminate subpar candidates early in the process, thereby conserving time and resources. This advancement has led to an increased demand for high-throughput laboratory systems capable of validating and characterizing the many AI-generated compounds now entering the pipeline. Despite these gains, challenges remain. Current AI models struggle to accurately predict the kinetic properties or developability of new compounds. As a result, every AI-designed candidate must still undergo rigorous laboratory validation. Traditional screening workflows, designed for identifying hits at scale, are ill-equipped to handle the complexity of modern AI-generated candidates. These workflows often rely on binary or threshold-based techniques, yielding low-fidelity data that fails to capture the nuanced characteristics of advanced drug candidates. To address these limitations, the industry is calling for more comprehensive and high-quality data. Many existing AI models are trained on publicly available datasets, which are often incomplete or biased. Belcher points out that these datasets frequently highlight only positive outcomes, omitting the numerous failed experiments that are essential for training robust models. This publication bias limits the accuracy and reliability of AI predictions, creating a bottleneck in the development of effective drug candidates. As the field continues to evolve, the integration of AI into laboratory systems is becoming increasingly critical. Researchers and industry leaders alike recognize that the future of drug discovery depends on overcoming these data and methodological hurdles. With ongoing advancements in AI technology and a growing emphasis on data integrity, the pharmaceutical industry is poised to make significant strides in its quest for more efficient and effective drug development.