Developing a new medicine is difficult. Scientists must understand a disease, identify a biological target, find a compound or biologic that modulates it, and show that it is safe and effective. Most experimental drugs fail before reaching patients.
AI systems can examine large collections of chemical and biological data, identify patterns, suggest molecules, and help researchers choose validating experiments. The technology can drive discovery faster with more focus, but it cannot eliminate the complexity of human biology or replace laboratory and clinical evidence – in short, “it still happens at the bench”.
Trend: AI Is Moving from Prediction to Creation
Earlier drug-discovery software was often used to screen existing collections of compounds. Newer generative AI systems can propose molecules that have never been made.
These models learn from known chemical structures and biological information. Researchers can ask them to design candidates with several desired qualities, such as binding to a disease-related protein while remaining practical to manufacture. AI can also help study proteins, identify possible targets, and connect information from genes, cells, and scientific literature (European Journal of Medicinal Chemistry).
This does not mean that AI can press a button and produce a medicine. It means scientists can explore a much larger set of possibilities before deciding what to make and test.
Trend: Discovery Is Becoming a Continuous Learning Cycle
One of the most important changes is the connection between AI and laboratory automation.
Newer systems connect computer modeling, chemistry, and biological testing in a repeating “design-make-test-learn” cycle. AI proposes candidates, scientists or robots test them, and the results guide the next predictions. Even a failed compound can reveal why a prediction was wrong and improve later designs.
Researchers increasingly recommend combining AI with automated synthesis, biological testing, physical simulations, and clear estimates of uncertainty. Models should also be judged on new experiments, not only on their ability to reproduce known results (Expert Opinion on Drug Discovery).
Innovation: Models Are Learning to Use More Kinds of Information
Another major innovation is the development of multimodal AI. These systems combine several types of information rather than relying on chemical structures alone.
A multimodal model might analyze a protein’s shape, gene activity in diseased cells, chemical features, and earlier experiments. Scientists are also adding rules from chemistry and physics and teaching systems to signal uncertainty. These advances may reduce unrealistic suggestions and give models a broader view of each problem.
Challenge: Computer Success Does Not Guarantee Laboratory Success
An AI-generated molecule can appear promising on a screen and still fail when scientists try to make or test it. It may be chemically unstable, toxic, difficult to manufacture, or unable to reach the intended tissue. It may work in a simplified laboratory test but fail inside a living organism.
Human biology is extraordinarily complex. Proteins change shape, cells communicate, and genetics, metabolism, and the immune system can all influence a drug. Model results are therefore useful hypotheses, not established facts. AI should help scientists select better experiments, not create the illusion that experiments are unnecessary.
Challenge: AI Depends on the Quality of Its Data
An AI system learns from the information it receives. In drug research, that information may be incomplete, inconsistent, biased toward successful experiments, or concentrated around well-studied diseases and molecules.
Negative findings are especially important. Researchers publish successes more often than failed experiments, so a model may learn mainly from positive or familiar examples and struggle with new biology.
Reviews of generative AI in drug discovery identify data scarcity, limited interpretability, uncertain performance on unfamiliar problems, high computing demands, and the lack of standard evaluation methods as continuing barriers (European Journal of Medicinal Chemistry).
Better AI will require well-designed experiments, careful data management, common testing standards, and more honest reporting of failure.
Challenge: Clinical Proof Is Still Developing
Several drugs discovered or designed with help from AI have entered clinical trials. One widely discussed example is In Silico Medicine’s INS018_055, which advanced into Phase II testing for idiopathic pulmonary fibrosis, a serious lung disease.
That progress is meaningful, but it does not prove that AI-designed drugs are more likely to succeed. Clinical outcomes depend on the target, molecule, trial design, patient selection, and scientific team. Published assessments caution that reported AI programs are a small and selected group, while unsuccessful programs receive less attention (Expert Opinion on Drug Discovery).
It is useful to separate three claims. AI may complete a research task faster. It may help scientists make a better decision. It may eventually increase the chance that a drug succeeds in patients. Evidence for speed is growing, but evidence for improved clinical success remains limited.
Conclusion: The Next Milestone Is Better Evidence
Regulators are developing expectations for AI used in drug development. A key idea is that a model should be evaluated according to its “context of use,” meaning its specific purpose and the consequences of an incorrect result.
An AI tool that helps scientists select early laboratory experiments does not carry the same risk as one used to support a decision about a drug’s safety or effectiveness. The U.S. Food and Drug Administration has proposed a risk-based framework for assessing the credibility of AI models used to support regulatory decisions (FDA guidance).
The FDA and European Medicines Agency have also published principles covering human oversight, documentation, data management, performance testing, and continued monitoring (FDA and EMA guiding principles).
The lesson is simple: the more important the decision, the stronger the evidence and oversight should be.
Conclusion: The next milestone is better evidence
AI is likely to become a normal part of drug research. It can help scientists compare more possibilities, identify useful patterns, avoid some unproductive experiments, and learn more from each result.
The most important progress may not come from a machine independently inventing a cure. It may come from thousands of smaller improvements in how researchers select targets, design molecules, plan experiments, and stop weak programs earlier.
AI will not eliminate biological uncertainty, clinical trials, or expert judgment. Its value will be determined by how well it connects predictions with real-world evidence.
The future of AI in drug discovery will not be measured by the number of molecules a model can generate. It will be measured by whether those ideas lead to safer, more effective medicines for patients. The success of AI platforms in new drug discovery and development will hinge upon the integration of in silico algorithms with actual data-driven biological validation. We must “let the data speak”.
