Professor Carlos Gershenson-Garcia, AI expert and professor in the School of Systems Science and Industrial Engineering at Binghamton University, State University of New York, and
Professor Weizi Li, professor of informatics and digital health, and deputy director of Informatics Research Centre, Henley Business School, University of Reading
Estimated reading time: 6 minutes
DeepSeek AI has been touted as a potential game-changer in terms of transforming drug repurposing, reducing research and development (R&D) costs and speeding up development deadlines for new treatments. But how does this new kid on the block differentiate from its competitors, and what could this potentially mean for the rare disease field?
Changing the game
“DeepSeek-R1 is a reasoning-focused large language model (LLM) trained mainly via reinforcement learning rather than huge amounts of human-labelled data, with strong performance across STEM (science, technology, engineering and mathematics) and medical benchmarks,” explains Professor Weizi Li, professor of informatics and digital health, and deputy director of Informatics Research Centre, Henley Business School, University of Reading.
The R1 model’s open weights parameters, combined with its permissive MIT software license means it is engineered to run efficiently on cheaper hardware, and this has advantages in healthcare, she says.
“The arrival of China’s DeepSeek-R1 is more about strong, low-cost, and open(ish) reasoning models for AI in medicine,” Professor Li continues. “Hospitals, academic medical centres and small and medium-sized enterprises (SMEs)can self-host a high-end model behind their own firewall, rather than having to send protected health information to a third-party cloud.”
Added to this, early studies1 have shown “DeepSeek models perform on par with or better than some proprietary LLMs for clinical reasoning tasks, including rare and complex cases,” she adds. “Open weights make it easier to retrain and fine-tune under data protection rules, and make it easier for regulators to demand transparency.”
In terms of its immediate impact, DeepSeek is in some respects trailblazing but it also (at least currently) has its limitations, Professor Li highlights.
“DeepSeek AI has changed the economics and accessibility of high-end models, more than it has (as yet) changed the fundamentals of drug discovery biology,” she explains.
“DeepSeek-R1 changes who can play the AI-supercharged drug discovery game and at what price point. That’s significant, but it’s an enabler rather than a singular breakthrough in biology. The real test will be whether those cheaper, more widely distributed tools translate into more successful first-in-human assets over the next 5-10 years.“
The arrival of China’s DeepSeek-R1 is more about strong, low-cost, and open(ish) reasoning models for AI in medicine
AI healthcare innovation before DeepSeek
Prior to DeepSeek’s arrival, much of the momentum in AI-driven healthcare innovation had been led by major technology companies such as Google DeepMind and Isomorphic Labs. Google Health had spearheaded advances in medical imaging such as retinal, breast and chest diagnostics, alongside electronic health record (EHR) models and protein-structure prediction through AlphaFold, with Isomorphic Labs going onto develop an AI-driven drug-design platform.
“Microsoft, working closely with Nuance Communications, was transforming clinical documentation through ambient scribing technologies such as Dragon, and expanding hospital-facing AI tools through its Microsoft Azure OpenAI service,” Professor Li explains. “Open AI, Anthropic, Google Gemini and Amazon Bedrock were providing general-purpose LLMs that were being integrated into EHR systems, imaging pipelines and pharmaceutical R&D workflows.”
The technology company Nvidia has been a critical enabler of the entire ecosystem, she adds, through its graphics processing units (GPUs), and its healthcare-specific software development kits such as Nvidia Clara and Nvidia BioNeMo.
“Nvidia has also had deep partnerships with biotechnology and pharmaceutical companies to build AI-optimised supercomputers for drug discovery and biomedical research.”
Could DeepSeek drive innovation?
“China’s growing activity in AI health technology is likely to encourage faster innovation across the sector,” Professor Li suggests.
“Models like DeepSeek have shown that strong performance can be achieved with efficient training methods and cost-effective, open weight releases, which increases expectations around affordability and capability for all developers.”
This could prompt technology companies in the United States and Europe, for example, to improve value, optimise model efficiency and accelerate development. The trend also highlights the importance of flexibility and control over data in healthcare, encouraging organisations to consider locally deployable or open weight models that can be adapted to their own settings, Professor Li explains.
“Established players still retain strengths in areas such as biomedical research, clinical partnerships and regulatory experience, and competition is likely to drive further progress in specialised clinical models, deeper integration in electronic health records and imaging systems, and better alignment with healthcare standards and safety requirements,” she says.
“Overall, China’s rapid progress adds healthy competitive pressure, helping to push the entire field toward more advanced, practical and affordable applications.”
It’s natural that competition would arise, both nationally and internationally, following China’s arrival in the healthcare AI space, says Professor Carlos Gershenson-Garcia, an AI expert and professor in the School of Systems Science and Industrial Engineering at Binghamton University, State University of New York.
“Of course, DeepSeek is not the final word—there will still be more improvements coming soon. The fact that China has developed DeepSeek AI makes the race international, but it should come as no surprise. After all, China announced in 2017 a $150bn plan to be AI leaders by 2030.”
What could DeepSeek’s arrival mean for rare diseases?
Professor Li says there are strong reasons to be optimistic, while Professor Gershenson-Garcia is more sceptical of the potential for progress in the rare disease field.
“DeepSeek-R1 could be particularly useful in rare disease research because its strength lies in reasoning across sparse, messy and highly fragmented data,” Professor Li explains. “Early evaluations1 suggest it can perform clinical reasoning at a level similar to some proprietary models, making it helpful as a second reader for rare disease clinics.
“It’s long-context capabilities allow it to mine case reports, small studies and registry data, cluster similar presentations and surface potential mechanisms or treatment ideas. It may also support target identification, drug repurposing hypotheses, and the design of small, adaptive trials.”
One of its benefits is it supports smaller scale research and development when resources are stretched, Professor Li adds.
“Importantly, its low cost and accessibility means researchers in under-resourced settings can also adopt it, potentially widening global participation. While risks such as hallucinations and evolving regulation remain, R1 aligns well with the needs of rare disease, where each condition has limited data but the overall evidence base is vast and scattered.”
Professor Gershenson-Garcia expresses caution with optimism but hopes to be proved wrong as AI innovation in healthcare and rare diseases continues to improve in the future.
“Rare diseases mean that there is little or no data to train any AI algorithms with,” he says. “While LLMs are very good at generalising from lots of data, the problem is a very different one for rare diseases. They might be useful to assist humans in some ways, but I don’t see them making a breakthrough. It would be great if I’m mistaken, but I do not see how these limitations could be overcome any time in the next 15 years.”
DeepSeek-R1 could be particularly useful in rare disease research because its strength lies in reasoning across sparse, messy and highly fragmented data
References [1] Sandmann, S., Hegselmann, S., Fujarski, M. et al. Benchmark evaluation of DeepSeek large language models in clinical decision-making. Nat Med 31, 2546–2549 (2025). https://doi.org/10.1038/s41591-025-03727-2
innovation spotlights the future of RARE through a technological lens. To access more innovation articles click below.
