Self-Supervised Foundation Model for Calcium-imaging Population Dynamics

A team of researchers has introduced CalM, a self-supervised foundation model designed to analyze functional calcium imaging data, a key method for recording neural activity. Unlike previous task-specific approaches, CalM is trained solely on neuronal calcium traces and can be adapted to multiple downstream objectives. Its core innovation is a pretraining framework featuring a high-performance tokenizer that maps single-neuron traces into a shared discrete vocabulary and a dual-axis autoregressive transformer. This architecture uniquely models dependencies along both the neural population axis and the temporal axis, capturing the complex dynamics of brain activity.

Evaluated on a large-scale, multi-animal, multi-session dataset, CalM demonstrated superior performance. It outperformed strong specialized baselines in the task of forecasting neural population dynamics. Furthermore, when equipped with a task-specific head, the model successfully adapted to behavior decoding, achieving results superior to supervised models. Beyond raw predictive accuracy, linear analyses of CalM's learned representations revealed interpretable functional structures within the neural data. This work establishes a novel, effective paradigm for self-supervised pretraining on calcium traces, paving the way for more scalable and broadly applicable tools in functional neural analysis. The team has stated that code for CalM will be released soon.