1.Background:
Artificial intelligence has transformed the way researchers analyse complex biomedical data, with Transformer models emerging as one of the most powerful approaches. Their attention mechanism allows the model to learn which parts of an input are most relevant to a task, a feature that has driven major advances in language and image analysis. Yet neuroscience has not fully benefited from this innovation. Many current neuroimaging approaches treat brain regions as independent features and only add biological knowledge after training. This neglects the fact that the brain is an interconnected and hierarchical system in which function emerges from patterns of communication across structurally and chemically organised circuits.
2. Novelty and importance:
This project introduces Neurobiologically Optimized Transformers, or NeuroBOTs, a new generation of models that embed biological knowledge directly into the architecture. By incorporating fixed attention maps derived from receptor density atlases, structural connectivity, and metabolic gradients, NeuroBOTs can prioritise biologically plausible interactions. This approach makes the models more interpretable and aligned with known brain organisation, while also supporting generalisation across populations. The work addresses a key limitation of current health AI, which often achieves predictive accuracy but provides little mechanistic insight. By bridging machine learning with systems neuroscience, NeuroBOTs will support the development of biomarkers that are not only statistically robust but also biologically meaningful.
3. Aims and objectives:
The project aims to design and evaluate these models for neuroimaging applications ranging from brain age prediction and disease classification to mapping clinical symptoms in psychiatric and neurological disorders. Their interpretability will be assessed against established brain networks, and their outputs will be refined in dialogue with clinical professionals, service users, and people with lived experience. Ultimately, the project will deliver open-source models, code, and documentation, contributing to reproducible science and advancing the role of artificial intelligence in precision brain health.
(https://www.preprints.org/manuscript/202506.0858/v2#sec1-preprints-172505).