New AI Molecular Prediction Model Could Speed Up Drug Discovery

Innovative research conducted in collaboration between a graduate student here and the global biopharmaceutical leader AstraZeneca has led to an important advancement in the field of AI-assisted drug discovery.

As recently published in Nature Communications, a novel AI approach has been developed that outperforms existing methods to predict the key properties of molecules and identify promising drug candidates.

In his PhD research, conducted in collaboration with AstraZeneca, David Buterez developed and trained new graph-based AI models on large, real world drug discovery datasets to see if they could make improved predictions about key properties of molecules, which inform their potential in therapeutic applications.

The newly published results, which build on earlier research, advance AI methods in drug discovery. David hopes that in using AI, he can simplify and accelerate scientific decision making and the identification of novel drug candidates.

Partnership with AstraZeneca
The University of Cambridge has had a partnership with AstraZeneca for a decade, during which time AstraZeneca has supported more than 160 PhD candidates, who have gained industrial as well as academic experience.

In David's case, for the last four years he and his supervisors – Pietro Liò, Professor of Computational Biology here, and Dino Oglic, Senior Director of Machine Learning & AI and Jon Paul Janet, Director of Molecular Design, at AstraZeneca – have been researching graph-based AI models, which have emerged as promising tools for molecular prediction.

Graph-based AI models
Graph-based AI models are particularly well-suited for analysing molecular structures as they represent molecules as graphs, where atoms are nodes and chemical bonds are edges. This allows the AI to learn and predict molecular properties based on the structure and connectivity of the molecule.

It was exciting to see that our architecture sets new state-of-the-art results on several molecular benchmarks.
- David Buterez

Drug discovery starts with primary screenings of huge libraries of millions of compounds. These screenings take one single measurement of the strength of interactions between molecules and their protein targets and the resulting 'hits' are then selected to undergo a second round of more detailed – and therefore more expensive and resource-intensive – screening.

The initial goal of this research was to see if the AI could transfer what it learned from these abundant measurements to the smaller, more targeted datasets. But it quickly became apparent that standard graph neural networks could not learn from this data.

To overcome this, David led the development of a new class of AI components called adaptive readout functions, allowing his model to perform up to eight times better than had previously been possible and unlocking the transfer learning potential for graph-structured data.

The initial results, published in 2024 in Nature Communications, were exciting. They showed ways that resource-intensive drug discovery processes could be optimised to run more efficiently and even, David says, to "suggest new active compounds that would likely be missed by traditional drug development techniques".

Using attention-based AI models
And in the latest work published this year, David has developed a new architecture for graph-based learning called 'Edge Set Attention', leveraging a mechanism called attention.

Adapting attention to graphs was a natural next step for David. "Attention is widely used in language modelling and also now in areas like computer vision, where it is being used with great success for generating images. Its application in this research was a natural step forward after last year's work on adaptive readouts for graph neural networks, since one of our most successful readout functions was attention-based."

Attention-based models are AI models that consider sequences and allocate weights to their component parts to understand which are the most important. In language processing, for example, an attention-based model will look for the key words in a sentence. "And in molecular prediction tasks," David says, "the model may try to find the atom, or group of atoms, that is the most relevant to the particular task."

So he designed a new 'Edge Set Attention' model, which was evaluated across a range of tasks. These included molecular property prediction, but also vision graphs and social networks. "We wanted to develop a general-purpose graph learning algorithm, meaning that it would apply to any graph type, not only molecules," he adds, "but it was exciting to see that our architecture sets new state-of-the-art results on several molecular benchmarks."

As the paper published in Nature Communications this year reports, "despite its simplicity, the model outperforms [other] methods across more than 70 tasks, including challenging long-range benchmarks. It also scales much better than alternatives with a similar performance level or expressive power." According to Dino Oglic, "This could accelerate the transition from traditional wet lab work to more sophisticated in silico methods."

I think that this opens the door to new models and applications – which are now seeing a lot of attention – generating molecules from scratch, or with a particular desired property.
- David Buterez

Collaboration helps push the boundaries of science
The collaborative research showcases how AI can help us better understand and manage the complex process of developing new medicines and highlights the critical role of academic partnerships and early-career talent in helping us push the boundaries of science.

In David's case, the partnership with AstraZeneca gave him valuable access to the company's datasets, which also motivated him to search for, curate and release similar datasets from the public domain. He also benefited from the deep scientific expertise and years of experience of his AstraZeneca colleagues.

"I think that this opens the door to new models and applications – which are now seeing a lot of attention – generating molecules from scratch, or with a particular desired property," David says. "One of the innovations of our paper is learning over the edges in a graph, instead of its nodes (for molecules, bonds instead of atoms) – which as of right now is uncharted territory in graph learning."

• David Buterez, Jon Paul Janet, Dino Oglic & Pietro Liò, 'An end-to-end attention-based approach for learning on graphs', Nature Communications, 5 June 2025.
• David Buterez, Jon Paul Janet, Dino Oglic & Pietro Liò, 'Transfer learning with graph neural networks for improved molecular property prediction in the multi-fidelity setting', Nature Communications, 26 February 2024.