A group of scientists at the Technical University of Denmark proved that a quantum computer can make generative AI drug discovery models more accurate and far reaching. They did it using their weekends and money left over from other projects.
The team ran their generative AI model for predicting proteins alongside a printer sized quantum computer built by British startup Orca Computing. That machine sped up the AI by linking quantum hardware with traditional processors. The researchers used the hybrid approach to generate new peptides, short chains of amino acids, that can bind to specific proteins in the body. That binding is a critical step in vaccine development.
A side project born from frustration
Professor Timothy Patrick Jenkins, who led the project, said the team worked weekends and pooled unspent funds from other projects because most innovative science is too scary for foundations. The researchers then made those peptides in the lab and tested whether they would actually bind to the target proteins. The model produced more successful peptides than its classical counterpart, especially where training data was scarce.
The team believes the machine could speed up the development of personalized immunotherapies and vaccines. It could also improve drug effectiveness for understudied groups, such as populations in Asia and Africa that have been left out of most medical research.
Jenkins said they needed to prove that their predictions connect to the real world in order to convince skeptics. Quantum computing is still a young field and faces heavy scrutiny because of the technical hurdles involved in building and applying these machines to real problems.
From quantum skeptic to believer
Jenkins himself was initially reluctant to explore the technology. He admitted he was a huge quantum skeptic and thought any application to his work would be decades away. His team normally uses big data and AI to discover proteins that could lead to cheaper and faster immunotherapies, often funded by the Novo Nordisk Foundation.
A particular challenge for his group has been the lack of data covering the full variety of human genetic information, since most medical research has focused on Western populations. That makes it difficult to develop peptides that work on understudied groups. The team hypothesized that embedding a quantum computer into their workflow could generate a more diverse set of peptides, especially for targets where they had less data. They got the idea after learning that quantum machines had a similar effect in generating images.
The process is not ready to transform research yet. Quantum computers remain too small to run full scale cutting edge AI models, so better results could still be achieved on a classical computer. PhD student Jonathan Funk noted that quantum is still not very powerful, so the level of complexity they could encode was not a normal sized antibody, which is what they usually work with. And finding a peptide that binds to a specific gene is only one step in vaccine development. It would not alone produce successful drugs.
Orca Computing CEO Richard Murray said it is no surprise that many industrial companies think quantum is hazy and far away, partly because the technology has never had really clear near term examples of usefulness. He added that this study is novel because it shows a near term commercial application for quantum. His company is also working with oil major BP on chemistry and with carmaker Toyota on making its design process more efficient.
The DTU team plans to test the workflow with more advanced models and larger proteins. Jenkins said they needed an easy way to validate that they now have a real shot at moving the needle substantially. He noted that generative AI workflows are especially valuable for neglected diseases that receive little research money. He is also looking at using a quantum computer to enhance his generative AI method for designing synthetic antidotes for snakebite venom.
For a broader look at how artificial intelligence timelines are evolving, see our analysis of AGI timelines.
