POCO: Scalable Neural Forecasting through Population Conditioning

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Predicting future neural activity is crucial for understanding brain dynamics and developing neurotechnologies.
• Existing models often focus on interpretability or decoding, leaving neural forecasting, especially across multiple sessions of spontaneous calcium recordings, underexplored.

Purpose of the Study:

• To introduce POCO, a unified model for accurate and generalizable neural forecasting in calcium imaging data.
• To capture both neuron-specific and brain-wide dynamics for improved prediction accuracy.

Main Methods:

• Developed POCO, a unified forecasting model combining a lightweight univariate forecaster with a population-level encoder.
• Trained and validated POCO across five diverse calcium imaging datasets from zebrafish, mice, and C. elegans.
• Analyzed factors influencing performance, including context length, session diversity, and preprocessing.

Main Results:

• POCO achieved state-of-the-art accuracy at the cellular resolution for spontaneous behaviors.
• The model demonstrated rapid adaptation to new recordings with minimal fine-tuning after pre-training.
• Learned unit embeddings in POCO recovered biologically meaningful structures, like brain region clustering, without anatomical labels.

Conclusions:

• POCO offers a scalable and adaptable approach for cross-session neural forecasting in calcium imaging.
• The findings provide insights for designing future neural forecasting models.
• POCO facilitates the development of adaptive neurotechnologies and large-scale neural foundation models.