Why AI Drug Discovery Depends on More Than Just Data

AI is changing how researchers approach drug discovery, but having access to billions of data points does not automatically lead to better treatments. The challenge is understanding what that data means, identifying relationships that might otherwise go unnoticed and establishing whether the resulting insights stand up to scientific scrutiny.

For Romain Clément, founder and CEO of ArcaScience, these questions are personal as well as professional. After being diagnosed with a brain tumour in 2014, he began researching pharmaceutical development and clinical data, spending three years building his understanding of an industry he had previously known little about.

That research eventually led him to co-found ArcaScience in 2018. In a recent episode of Nothing Ventured, hosted by EmergeOne founder and CEO Aarish Shah, Clément discussed the problems with pharmaceutical data, the potential for AI to identify overlooked treatment options and why building a company in this sector requires a different approach to growth.

Finding new uses for existing drugs

Drug discovery is often associated with developing entirely new compounds. However, there is another route to treatment: identifying drugs that already exist, or have been investigated previously, that could be useful for a different condition or group of patients.

This is known as drug repurposing. A medicine developed for one purpose may have properties that make it relevant to another, but identifying those possibilities requires researchers to connect evidence from different studies, clinical trials and scientific publications.

The information needed to make those connections may already exist. The difficulty is finding it, assessing its relevance and understanding what it means for a particular patient.

Clinical trials also generate publicly available information, while pharmaceutical companies may hold compounds that are no longer being developed or marketed. In some circumstances, a drug that is no longer commercially available could still be worth investigating for an individual patient.

Making this information easier to navigate could help researchers identify options that might otherwise remain overlooked. It does not remove the need for clinical assessment or establish that a treatment will work, but it can help direct attention towards possibilities worth investigating.

The problem with scientific data

AI systems are only as useful as the information they can work with. In pharmaceutical research, that creates a substantial challenge.

Scientific data is often unstructured. Relevant information can sit inside research papers, clinical documents, images and files produced by specialist equipment. Different systems use different formats, making it difficult to bring everything together in a form that computers can analyse consistently.

Even when information is available, it may not be reliable. Scientific literature includes research of varying quality, and published findings do not all carry the same weight. Any system drawing conclusions from this material needs to distinguish credible evidence from weaker or misleading sources.

This is where the volume of available information can become a distraction. Feeding more material into an AI system does not necessarily produce better conclusions. If the input is incomplete, poorly structured or unreliable, the output can be equally flawed.

Preparing scientific information for analysis is therefore a substantial part of the work. Documents need to be interpreted, relevant details identified and evidence assessed before it can contribute meaningfully to a research question.

That process may be less visible than the latest AI model, but it has a direct bearing on what the technology can achieve.

More information does not always mean better answers

As AI systems become capable of processing larger amounts of information, the question of data quality becomes more important.

A dataset can contain thousands of studies without providing a clear answer to a specific scientific question. What matters is how the evidence relates to the problem being investigated, what may be missing and whether the available information supports a credible conclusion.

There is also a risk in deciding too early which information matters. If a system only considers evidence selected in advance, it may reproduce the assumptions behind that selection rather than identify something genuinely unexpected.

In drug discovery, that could mean overlooking a useful connection between a compound and a disease simply because the relevant evidence sits outside the dataset or category researchers initially considered.

AI can help explore relationships across large bodies of research, but those relationships still need to be interpreted carefully. A pattern in the data is a starting point for investigation, not proof that a treatment is safe or effective.

The distinction matters in a field where decisions ultimately affect patients. Speed and scale are valuable, but scientific rigour remains essential.

Balancing innovation with safety

Developing AI for healthcare also raises questions about how innovation is regulated and assessed.

Pharmaceutical development operates within a framework designed to protect patients. Treatments must meet demanding standards for safety and efficacy, and the evidence required to bring a drug to market can take years to generate.

AI development often moves at a different pace, with a strong emphasis on rapid experimentation and deployment. Applying that approach directly to drug development is difficult because the consequences of an incorrect conclusion are much more serious than an unsatisfactory software output.

The challenge is to make progress while maintaining appropriate standards of evidence. Better analysis can help researchers investigate promising options more efficiently, but it cannot replace the clinical trials, regulatory scrutiny and medical judgement needed to establish whether a treatment should be used.

Europe’s experience in highly regulated industries may be relevant here. Its pharmaceutical sector has developed extensive systems for assessing risk and protecting patients. The opportunity is to combine that experience with the capabilities of AI, while ensuring that regulation supports useful innovation rather than simply slowing it down.

Building a business around scientific progress

Commercialising deep technology brings another set of constraints. In sectors such as healthcare, development can be expensive, sales cycles can be long and the route from technical capability to commercial adoption is rarely straightforward.

The conventional venture capital model often places considerable emphasis on rapid growth. That can be difficult to reconcile with businesses whose products depend on scientific validation, specialist expertise and lengthy adoption processes.

For companies working in drug discovery, commercial progress needs to account for the realities of pharmaceutical research. The technology has to solve a meaningful problem, fit into existing workflows and produce results that customers can trust. Technical sophistication alone is unlikely to be enough.

The same applies to the wider AI sector. A model’s capabilities matter, but so do the quality of the data, the reliability of its conclusions and the practical value it creates for the people using it.

AI has considerable potential to help researchers make better use of existing scientific knowledge. Realising that potential will depend on more than building increasingly powerful systems. It will require reliable evidence, careful application and a clear understanding of the problems the technology is meant to solve.

Enjoyed this article?

Subscribe to our newsletter for founder insights, practical advice on fundraising and finance, and the biggest takeaways from our latest podcast episodes, delivered straight to your inbox.

EmergeOne provides fractional CFO support to ambitious startups, with a team of experienced fractional CFOs in London and across the UK.