An AI Built to Read Biology's Hidden Geometry

A team at Cardiff University has proposed two ways of teaching AI to respect something most models quietly ignore: the shape of the data itself. The work, described on 1 September and presented at this year's International Conference on Machine Learning, takes aim at a mismatch that has held back AI in biology for years. Living systems are organised with a rich internal geometry, and the standard tools were never built to see it.

The senior author, Dr You Zhou of Cardiff's School of Medicine, puts the problem plainly.

"Biological research generates enormous amounts of complex data, but much of today's AI still struggles to understand the way biological systems are naturally organised."

That gap matters because so much of modern biology now runs on machine learning. When a model flattens the spatial and molecular relationships between cells into a plain list of numbers, it throws away exactly the structure a biologist cares about.

Two ideas, one theme

The first method, Vector Bundle Attention, folds geometric relationships directly into the attention mechanism that powers modern models, so the system weighs information according to the underlying structure rather than treating every data point as interchangeable. In tests it reached state-of-the-art results on single-cell RNA sequencing, and performed well on spatial and three-dimensional data where position carries meaning.

The second, named Dynamic Fractal Mamba, tackles a different frustration. AI models trained on small slices of data often fail when handed something much larger. This method applies the same learned rules repeatedly across ever-wider patterns, which let it analyse far bigger datasets it had never seen before without being retrained. For labs sitting on data that outgrows their models every year, not having to retrain from scratch is a practical win, not just an elegant one.

Why the plumbing is the point

Neither method will make headlines the way a chatbot milestone does. That is rather the point. Progress in AI for science tends to arrive in the plumbing, in the unglamorous business of representing messy real-world data faithfully, and it compounds quietly across every study that adopts it. Better single-cell analysis feeds work on disease, drug response and how tissues are built.

It also sits alongside a run of results showing AI moving into the harder, more rigorous corners of research, from formal mathematics to the structure of proof itself. The common thread is a shift in ambition. The interesting question is no longer whether a model can produce a plausible answer, but whether it can respect the real structure of the thing it is reasoning about. On that measure, geometry is a good place to start.