Artificial intelligence (AI) holds promise to speed drug discovery by helping scientists develop and test new drug candidates. Traditionally, the development of one drug can cost over $2 billion and take a decade—but only about 10% of new drugs emerge successfully from clinical trials, able to recoup their development costs. Thus, using AI to develop drugs in less time and for lower costs is a valuable move for companies and patients alike. Here’s a look at how AI is transforming drug discovery, and challenges scientists are wrestling with along the way.
AI in drug discovery today
The recent development of generative AI has turbo-charged its effectiveness as a drug discovery tool. Language-based generative tools (such as those powering ChatGPT or Google Gemini) use language learning models (LLMs) to analyze data, allowing application to the biochemical “languages” of the sciences. This includes sequences for DNA, RNA, and proteins, as well as chemical and molecular structures—all of which can be expressed in language formats. In short, scientific disciplines such as biology and chemistry can be described in words, which allows generative AI tools to work on their components.
Drug discovery products Search Now Search our directory to find discovery-related products for your research needs.
One of the benefits of using LLMs and generative AI in drug discovery is that biologists and chemists can query AI using common language, rather than specialized computer skills. Darragh McArt, Founder and CEO of Sonrai Analytics, believes that the most profound effect of AI is in early discovery areas, such as target identification, biomarker discovery, and patient stratification. “AI has had the greatest impact where data volume and complexity exceed human capacity to interpret efficiently, where large-scale omics, imaging, and clinical datasets can be analyzed in parallel,” McArt says. “AI is also increasingly influential in hypothesis generation, candidate prioritization, and translational decision-making, helping teams move from correlation to actionable insight more quickly.”
Sonrai Analytics uses AI to integrate multimodal data. This is a task too enormous for one research group or company using traditional analytical approaches, because typically the modalities are in different data formats. “AI enables integration across previously siloed datasets, [allowing] researchers to identify patterns that span molecular, phenotypic, and clinical domains,” says McArt. “This makes AI particularly valuable in precision medicine contexts, where understanding heterogeneity is critical.”
AI is accelerating analysis of genomic and proteomic data, identifying new disease targets, leads, and drug candidates with therapeutic potential. AI speeds up the latter by predicting properties of drug compounds, or—in other applications—designing drugs de novo according to desired properties. AI allows researchers to optimize potential drugs “in silico” because it can use multiple parameters —such as chemical stability, pharmacokinetics, and absorption—to predict structures, molecular interactions, and effects.
As one of the first companies to leverage generative AI in drug discovery, Insilico Medicine used AI to find a novel target in the lung disease idiopathic pulmonary fibromatosis, as well as an effective therapeutic1,2 in markedly less time compared to conventional methods. They describe their proprietary “Pharma.AI” platform as working toward the idea of pharmaceutical superintelligence, in which a fully autonomous platform is able to find and create an improving or curative therapeutic, as well as identify a biomarker in patients.3 They’ve also developed a new class of small molecule inhibitors showing promise in overcoming resistance to immune checkpoint blockades in treatment of solid cancers, a long-standing problem in cancer immunotherapy.4,5 Another company that uses generative AI for precision medicine and small molecule therapeutics discovery, Relay Therapeutics, developed a new PI3Kɑ inhibitor drug for a specific type of breast cancer that reduced tumors in 81% of patients.6
AI is also speeding up clinical trials, a drug development process that is vital yet slow and clunky; it can take up to 18 months simply to enroll people in a mid-stage clinical trial. Companies are now using AI to streamline and speed up this process. For example, AI can identify the patients most appropriate for clinical trials to help enroll them faster and more easily. The company Tempus uses AI to analyze de-identified patient records, with the aim of recruiting patients that are best-qualified for particular clinical trials. Insilico Medicine has shown that their AI tool InClinico predicted clinical trials outcomes with 79% accuracy.7 BioPhy developed a platform with AI that not only simulates clinical trials for drug developers, but it also uses patient medical records to simulate outcomes from different patient populations. In short, optimization with AI can make the entire drug discovery process faster, reducing years of work into weeks, while also testing more drug candidates with greater accuracy.
Where AI is falling short
So what’s not to love? Despite its tremendous advantages in drug discovery, AI can fall short. “The main challenges with AI in drug discovery are structural,” says McArt. “Poor data quality, fragmented datasets, and lack of standardization can severely limit the reliability of AI outputs.” He recommends transparency and prioritizing strong data foundations using well-annotated datasets and reproducible data pipelines. “AI should be embedded within existing scientific workflows rather than treated as a standalone solution,” says McArt. “When AI is built on robust infrastructure and designed for real-world use, its limitations become manageable and its value far more sustainable.”
Lest anyone view AI as the powerful supercomputer HAL from “2001: A Space Odyssey”, it is sobering and occasionally amusing to learn that today’s AI in drug discovery sometimes suggests nonsense chemical compounds. It can also suggest a molecule that a human medicinal chemist might identify as chemically unstable or impossible to synthesize, or would have obviously undesirable effects. In fact, only a few drugs in clinical trials today were actually discovered with AI—but it is early days yet, in a very long process.
Moving forward
Any failings in AI’s capabilities reflect how people have trained the algorithms in the first place. AI improves when trained with high-quality data. Indeed, an iterative training and experimenting process is showing promise. In this process, AI is used to generate hypotheses, identify patterns, and give scientists ideas for new directions. This is followed by real-life lab experiments conducted by humans (or in some cases, AI-controlled robots), and the results are fed back into the AI model so that it can “learn” the new information. This method can help to reduce future optimization, and also keep AI models grounded in reality.
Genentech (part of the Roche Group) uses such a “lab-in-a-loop” method to target neoantigens for cancer vaccine discovery. Roche is collaborating with AWS and NVIDIA to expand its AI potential. Recursion Pharmaceuticals is also collaborating with NVIDIA to further AI-assisted drug discovery. Biotech companies are now offering more sophisticated “foundation models” of generative AI specifically trained on biological data (e.g., AWS HealthOmics and NVIDIA BioNeMo). For new drug design, using a foundation model means less time doing general training of the AI algorithms, and more time zeroing in on biological relevance. Other pharma companies such as Sanofi, BMS, and Eli Lilly are also partnering with AI companies.
But more data isn’t enough to make AI better at finding new drugs. One still needs to know how best to fit the data together in the context of biological mechanisms operating within complex disease systems—and that’s still a human’s job. “Human judgment will remain essential wherever context, interpretation, and ethical responsibility matter,” says McArt. “This includes framing the right scientific questions, assessing biological plausibility, and deciding how to act on analytical results.” Though a powerful tool, AI cannot replace the human understanding that places results into a broad biochemical and clinical context. “Decisions such as prioritizing therapeutic hypotheses, balancing risk in clinical development, and interpreting ambiguous or conflicting evidence require a deep understanding of biology, disease, and patient impact,” notes McArt. In drug discovery as in many fields, it will be interesting to observe how scientists’ roles change over time as we train AI in increasingly sophisticated tasks.
References
1. Ren, F., Aliper, A., Chen, J. et al. A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models. Nat Biotechnol 43, 63–75 (2025). https://doi.org/10.1038/s41587-024-02143-0
2. Xu, Z., Ren, F., Wang, P. et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nat Med 31, 2602–2610 (2025). https://doi.org/10.1038/s41591-025-03743-2
3. Insilico Medicine unveils winter edition of Pharma.AI accelerating the path to pharmaceutical superintelligence. Press release. Accessed January 5. 2026.
4. JMC Publication: Insilico’s AI platforms enable discovery of potent, selective, oral DGKα inhibitor to overcome checkpoint resistance. Press release. Accessed January 5, 2026.
5. Hongfu Lu, et al. Design, Synthesis and Biological Evaluation of Novel, Potent, Selective and Orally Available DGKα Inhibitors for the Treatment of Tumors. Journal of Medicinal Chemistry 2025 68 (23), 25011-25025. DOI: 10.1021/acs.jmedchem.5c01943
6. Relay Therapeutics announces updated data for RLY-2608+ Fulvestrant further demonstrating clinically meaningful progression free survival at ASCO 2025. Press release. Accessed January 5, 2026.
7. Aliper, A., et al. Prediction of Clinical Trials Outcomes Based on Target Choice and Clinical Trial Design with Multi-Modal Artificial Intelligence. Clin Pharmacol Ther, 114: 972-980. https://doi.org/10.1002/cpt.3008 (2023).