Cardiff researchers develop AI models for complex biological data

Cardiff University researchers have developed two AI models intended to help scientists interpret complex biological data, from individual cell behavior to the structure of whole tissues. The work targets two persistent challenges in AI: representing the geometry and relationships embedded in data, and recognizing patterns that appear across very different scales.

Vector Bundle Attention, or VBA, modifies the attention mechanism used in transformer-based systems so it can account for geometric relationships in biological data. The model aligns information from cells according to underlying geometry before comparing it, a capability designed for contexts where cells operate within molecular and spatial networks. Tests using single-cell RNA sequencing and spatial transcriptomics showed state-of-the-art performance in single-cell RNA sequencing tasks, strong performance in spatial transcriptomics and competitive results on 3D datasets.

A second model, Dynamic Fractal Mamba, or DF-Mamba, focuses on scale. Inspired by physics concepts for describing behavior across scales, it repeatedly applies learned rules as it moves from smaller patterns to larger contexts. The model can learn from smaller-scale data and analyze much larger datasets it has not previously encountered without retraining. The studies were presented at the 2026 International Conference on Machine Learning.