Title: CalM: A Self-Supervised Foundation Model for Population Dynamics in Calcium Imaging Data

Abstract:

Emerging research indicates that large-scale, multi-animal modeling can substantially enhance the analysis of neural recordings. Despite this potential, current methods for processing functional calcium traces are largely restricted to specific tasks, hindering their applicability across broader neuroscience goals. To overcome these limitations, we introduce CalM, a self-supervised neural foundation model trained exclusively on neuronal calcium traces. CalM is designed to be versatile, supporting a variety of downstream applications such as forecasting and decoding.

The core innovation of our work lies in a novel pretraining framework. This framework features a high-efficiency tokenizer that converts individual neuron traces into a unified discrete vocabulary, alongside a dual-axis autoregressive transformer. This transformer captures dependencies along both the neural and temporal dimensions. We assessed CalM using a comprehensive dataset comprising multiple animals and sessions. In the realm of neural population dynamics forecasting, CalM demonstrated competitive results against robust specialized baselines following the pretraining phase.

Furthermore, by integrating a task-specific head, CalM successfully adapted to behavior decoding tasks, outperforming standard supervised decoding models. Additionally, linear analyses of CalM’s internal representations uncovered interpretable functional structures that extend beyond mere predictive accuracy. Collectively, our study presents a new and effective self-supervised pretraining paradigm for foundation models based on calcium traces, establishing a foundation for scalable pretraining and wide-ranging applications in functional neural analysis. The code is available at https://github.com/TSuXinH/CalM.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC