The core challenge in drug discovery has never been a shortage of scientific ideas. It has been the ability to connect and interrogate the right evidence, at the right point in the research process, before committing years and billions of dollars to the wrong candidate. That is where AI is already making a difference.
58% of researchers now use AI tools for work, up from 37% in early 2024. Most use it for time-consuming tasks: finding and summarising research, comparing sources, extracting patterns from complex datasets and making sense of large bodies of information. For drug discovery, those use cases are useful but limited.
The bigger opportunity is embedding AI into scientific workflows, where AI can be used to search, synthesise, prioritise, document evidence and increasingly coordinate multi-step processes under human oversight.
Turning fragmented data into evidence
Drug discovery starts by connecting evidence from scientific literature, compound data, reaction dynamics, biological pathways and clinical findings. These sources sit in separate systems, making it difficult for researchers to consolidate knowledge quickly or consistently.
AI can help bring these sources together. In target identification, it can interrogate disease biology, connect evidence across datasets and suggest prioritised targets for validation. In hit identification and lead optimisation, it can help narrow thousands of compounds into a workable shortlist.
In design-make-test-analyse cycles, it can help teams refine hypotheses earlier, recommend which routes should be tested next and explain why a particular experimental path appears most promising.
The impact should be measured in practical terms: Are we increasing speed, success rates, or efficiency? Good metrics include shorter time-to-insight, faster iteration, more viable hypotheses tested per project, fewer late-stage terminations and better reuse of datasets, models and research tools.
Improving the economics of discovery
Drug discovery and development is long, costly and high risk, often taking 10 to 15 years and costing more than a billion dollars. If AI can help identify non-viable candidates sooner, run more cycles in less time and connect biology and chemistry data more effectively, it could change how resources are allocated across the pipeline.
More importantly, AI can help reduce the time it takes to get effective treatments to patients.
This matters most in areas such as rare diseases, where small patient populations and limited commercial incentives make traditional development models difficult to sustain. Many patients with rare conditions wait years, sometimes indefinitely, for treatments that never reach the clinic. AI-enabled drug repurposing offers one practical route forward.
One example of this is RARE Hope, a non-profit focused on rare neurological diseases, which used knowledge graphs and biological relationship data to analyse nearly 2,000 existing drugs, scoring and prioritising candidates based on disease-target interactions.
The approach significantly narrowed the field for expert review, demonstrating what becomes possible when AI is applied to well-structured data with a clear scientific question.
From assistive tool to active partner
Not every AI system plays the same role. Predictive machine learning can help model compound properties, biological activity or experimental outcomes. Generative AI can support literature synthesis, knowledge discovery and early hypothesis exploration.
Agentic AI systems go a step further by coordinating multi-step workflows across datasets, tools and decision points.
The most advanced systems are beginning to function as orchestration layers, interpreting a research question, selecting the right datasets or tools, retrieving evidence, applying domain context and producing an output that scientists can interrogate.
In some workflows, this is already moving researchers from asking ‘what evidence exists?’ to ‘which route should we test next, and why?’. That shift matters because it changes the role of AI from a passive assistant to an active participant in the discovery process.
Agentic AI also creates the possibility of more continuous, 24/7 scientific workflows. Systems can be designed to monitor new evidence, trigger the next step in a defined process, prepare candidate options for review, and keep research activity moving between human decision points.
That does not make AI a substitute for scientific judgement. Its strongest role is augmenting researchers to take on the work of searching, synthesising, prioritising and workflow coordination, allowing humans to focus on experimental design, validation and original thinking.
Confidence depends on strong foundations
Despite the potential of AI, adoption remains uneven. Only 22% of researchers currently describe AI as trustworthy, and nearly half feel undertrained in how to use it. Bridging that gap requires organisations to build the conditions in which AI can be used reliably.
Trust starts with transparency. One of the most significant barriers to AI adoption in research is the black box problem: Researchers can see an output but not the reasoning behind it. This becomes even more important as AI systems move from answering individual prompts to coordinating multi-step workflows.
Human experts must remain responsible for interpretation, validation and final decisions, but they also need enough visibility to understand how an AI system reached a recommendation.
The data foundation matters just as much. Most scientific data was created for human consumption, not machine reasoning. Making it fit for AI requires that researchers can establish where data originated, how it was processed and whether it is current.
It also requires consistency: the same concept described in different ways across different systems, or named differently, will degrade the quality of any AI output built on top of it.
FAIR principles (data that is Findable, Accessible, Interoperable and Reusable) provide the starting point for building that consistency. But data also needs semantic enrichment, mapping raw information to standard concepts so that meaning is preserved across sources. A practical example is terminology alignment.
Systems must recognise that “GLP-1” and “glucagon-like peptide-1” refer to the same biological target. Without that kind of enrichment, AI systems cannot reliably connect evidence across datasets, which is what makes them useful in drug discovery.
What good looks like
The next stage of drug discovery will not be defined by AI acting as a research assistant. It will be defined by AI systems that can work through multi-step reasoning, coordinate parts of the discovery workflow and bring forward evidence-backed recommendations for expert review.
Used well, AI can help teams make better use of evidence, surface connections earlier and focus experimental effort on the most viable candidates.
The goal is not to remove scientific or human judgement from the process. It is to give researchers better evidence, better options and more time to apply that judgement where it matters most.
The ambition is for AI in drug discovery to be not only transparent and traceable –but also active, integrated and capable of helping scientists move from insight to action faster.
