GenAI: The double-edged sword revolutionizing drug discovery

We all want to apply AI to improve patient outcomes. But how?

“It’s about passion,” says Dr Ivan Krstic 새 탭/창에서 열기, Senior Director of Life Sciences at Elsevier, as moderator of the webinar Generative AI: Transforming Knowledge and Drug Design 새 탭/창에서 열기. “And I believe all of us here share this passion for harnessing computational advances to transform drug discovery and improve patient outcomes.”

Of course, with passion come practicalities — and much room for improvement. Current drug discovery approaches typically require three to five years and €50 million to €100 million per new candidate, and they still have only a 10% to 20% success rate.

So it’s easy to understand why researchers in the life sciences want to adopt AI solutions to streamline this process. But first, the industry needs to face some fundamental challenges: data quality and availability, integration with existing systems, regulatory uncertainties, talent and expertise requirements and scalability concerns.

Opportunities on one side, challenges on the other

It’s tricky, but the effort is more than worth it, according to the two guest experts on the webinar — the first installment of the Foundations for Effective AI webinar series for 2025. While AI tools offer tremendous potential — such as the ability to explore vast chemical spaces beyond traditional libraries — they should be viewed as aids to human expertise rather than replacements.

“AI is not magic,” stresses Dr Brice Hoffmann 새 탭/창에서 열기, Principal Scientist and Head of Computational Chemistry at the innovative French AI company Iktos 새 탭/창에서 열기. “It’s a tool that will help you decide what to make — and that’s already good enough.”

On the future of scientific research

According to Dr Arash Hajikhani 새 탭/창에서 열기, Research Team Leader at VTT Technical Research Centre 새 탭/창에서 열기 of Finland, who studies and assesses AI’s positive and negative impact, AI is already reshaping the role of computational chemists in modern research.

On one fundamental level, GenAI can open doors and unite a community once limited by language barriers. Arash highlights how Generative AI democratizes science by breaking down language barriers 새 탭/창에서 열기 — a crucial advancement considering that 95% of the world’s population are not native English speakers.

But while GenAI may work to democratize science, you must also work to protect its integrity. Arash warns of an emerging “vicious circle” where generative AI is filling up the internet 새 탭/창에서 열기 and “polluting scientific discourse and literature.”

"We are faced with the critical task of addressing this self-sustaining loop of misinformation," he says, pointing to a trend where AI is now being used to peer-review papers that may be AI-generated. “This creates a potentially dangerous echo chamber that could further undermine scientific integrity.”

In short, researchers must be vigilant and apply strategies to identify AI-generated content, and the peer review process must evolve to address these challenges.

On emerging technologies

Meanwhile, GenAI has already proven its worth. Arash cites the practical solution of combining Large Language Models (LLMs) with knowledge graphs. Through its ability to “understand” context and make semantic connections, this integration helps researchers navigate the ever-growing volume of scientific literature more effectively while minimizing GenAI’s proclivity to hallucinate when not trained on the correct information.

For instance, when searching for “breakthroughs in AI for cancer diagnosis,” the integrated system can prioritize recent publications while identifying relevant connections across research papers, datasets and clinical trials.

Meanwhile, Brice offers his company’s efforts as a compelling case study on how AI-driven approaches can help pharmaceutical companies accelerate their Design-Make-Test-Analyze (DMTA) drug development research cycle.

Iktos’s end-to-end drug discovery platform 새 탭/창에서 열기 integrates AI-driven design and synthesis prediction to cut development time in half. Brice shares a concrete example of AI-assisted drug design for a C1s inhibitor 새 탭/창에서 열기. The Iktos system generates molecules that achieve a 10-fold improvement in binding affinity through strategic molecular structure optimization.

“By combining multi-parametric optimization with AI-driven retrosynthesis, we can design molecules that are very fast to synthesize and test,” Brice explains. “With this rapid turnaround, we can accelerate the DMTA cycle.”

An emerging trend such as autonomous AI agents, where an AI can cast its net across a variety of AIs and data sources, will only work to further transform — and accelerate — research methodology and decision-making. Improved robotics will undoubtedly also accelerate DMTA’s “T” for testing.

However, we must also work to understand how these increasingly complex systems generate their outputs.

On validation and quality control

“We need to ensure all innovations follow human-centric design principles and are validated by peer review,” Brice emphasizes, bringing the discussion back to Arash’s opening concerns about scientific integrity. Systems also need to include their own checks and balances. For instance, with the Iktos two-pronged approach, one AI works to make a molecule, and another evaluates whether the suggestion is feasible — an approach that fortunately can also filter for greenness and cost.

Arash also recommends that researchers not limit themselves to one LLM, pointing out that there are 15,000 available, each trained for different purposes. However, he stresses the importance of data quality, illustrated by the 10:10 clock phenomenon 새 탭/창에서 열기: When asked to create a clock, LLMs invariably set it to 10:10 because that’s how clocks appear in their training data (i.e., the internet). As Arash notes:

On implementation and ethics

Regarding overall implementation, Arash believes the key to success lies in individual responsibility:

We are a community benefiting from this solution. If we don't use it responsibly and pollute, we will suffer from that same solution.

Arash also warns that over-reliance on these tools could impact researchers’ cognitive capabilities. “When we write, we also think,” he notes, suggesting that outsourcing too much of the scientific process to AI could potentially diminish human analytical skills: “I try to use them as mentors to elevate my capabilities, for instance, with my writing. Instead of asking to proofread, I ask the LLM to give me hints and tips so I can learn gradually along the way.”

The discussion concludes with a powerful reminder from Ivan:

The path forward

To summarize, while AI presents opportunities and challenges for the scientific community, success lies in finding the right balance: leveraging AI’s capabilities while maintaining human oversight, accelerating processes while ensuring accuracy, and embracing innovation while preserving scientific integrity.

As research becomes increasingly AI-enabled, this balanced approach will be crucial in realizing AI’s potential to transform scientific discovery and improve patient outcomes — further fueling the passion that drives researchers in the life sciences community forward.