Evaluating linear model frameworks for directed predictive influence estimation from hypothalamic calcium imaging

Understanding neural interactions from calcium imaging requires statistical tools that can extract directed relationships from high-dimensional, noisy signals. Here, we evaluate linear modeling frameworks for estimating functional connectivity from hypothalamic MCH neuron activity recorded via deep-brain calcium imaging in freely behaving mice. Rather than inferring anatomical connectivity, we benchmark the performance of linear regression models - implemented as pairwise and multi-predictor formulations - with and without L1 (lasso) regularization. We emphasize that these analyses benchmark statistical methods for estimating directed predictive influence (DPI) - defined as predictor-to-target regression coefficients in zero-lag models - from fluorescence signals; they do not constitute experimental measurements of anatomical or synaptic connectivity. Here, direction refers to the predictor-to-target orientation of coefficients in a zero-lag regression model, not temporal or causal direction. These models are compared against traditional correlation-based approaches to assess their ability to capture directed dependencies among neurons. Cross-validation and hold-out testing were used to assess internal predictive consistency and coefficient stability across neurons (with the caveat that random time-sample splits can yield optimistic performance in autocorrelated calcium traces). Simulations with predefined network structure further delineate the conditions under which such models recover functional relationships, with performance compared against null distributions from 500 shuffled networks per condition. Analyses were performed on both z-scored fluorescence signals and smoothed spike-rate estimates obtained from MLSpike deconvolution, demonstrating consistent inference across preprocessing methods. Cross-dataset validation across eight recordings (N=11-25 neurons, four mice, two behavioral states) confirms robust generalization of key findings. Together, these results define a transparent, scalable, and interpretable modeling pipeline for estimating DPI (regression-based functional coupling) from calcium imaging data, providing a benchmark for evaluating statistical approaches to neuronal network inference.