P154 Comparison of derivative-based and correlation-based methods to estimate effective connectivity in neural networks
Niklas Laasch1, Wilhelm Braun1,2, Lisa Knoff1, Jan Bielecki2, Claus C. Hilgetag1,3
1Institute of Computational Neuroscience, Center for Experimental Medicine, University Medical Center Hamburg-Eppendorf, Martinistraße 52, 20246, Hamburg, Germany
2Faculty of Engineering, Department of Electrical and Information Engineering,Kiel University, Kaiserstrasse 2, 24143, Kiel, Germany
3Department of Health Sciences, Boston University, 635 Commonwealth Avenue, Boston, MA, 02215, USA
E-Mail:niklas.laasch@posteo.de
Introduction
Inferring effective connectivity in neural systems from observed activity patterns remains a challenge in neuroscience. Despite numerous techniques being developed, no universally accepted method exists for determining how network nodes mechanistically affect one another. This limits our understanding of neural network structure and function. We focus on purely excitatory networks of small to intermediate size with continuous dynamics to systematically compare different connectivity estimation approaches, aiming to identify the most reliable methods for specific network characteristics.Methods
We used the Hopf neuron model with known ground truth structural connectivity to generate synthetic neural activity data. Multiple connectivity inference algorithms were applied to reconstruct the system's connectivity matrix, including lagged cross-correlation (LCC) [1], derivative-based covariance analysis (DDC) [2], and transfer entropy methods. We varied parameters controlling bifurcation, noise, and delay distribution to test method robustness. Forward simulations using estimated connectivity matrices were performed to evaluate each method's ability to recreate observed activity patterns. Finally, we applied promising methods to empirical data fromC. elegans.Results
In sparse non-linear networks with delays, combining LCC with DDC analysis provided the most reliable connectivity estimation. LCC performed comparably to transfer entropy in linear networks but at significantly lower computational cost. Performance was optimal in small sparse networks and decreased in larger, denser configurations. With the Hopf model, LCC-based connectivity estimates yielded higher trace-to-trace correlations than derivative-based methods for sparse noise-driven systems. When applied toC. elegansneural data, LCC outperformed more computationally expensive methods, including a reservoir computing approach.Discussion
Our findings demonstrate that a comparatively simple method - lagged cross-correlation - can reliably estimate directed effective connectivity in sparse neural systems despite spatio-temporal delays and noise. This has significant implications for biological research scenarios where only neuronal activity, but not connectivity or single-neuron dynamics, is observable. We provide concrete suggestions for effective connectivity estimation in such common research scenarios. Our work contributes to bridging the gap between observed neural activity and underlying network structure in neuroscience.
Acknowledgements
The authors would like to thank Kayson Fakhar, Alexander Schaum, Fatemeh Hadaeghi, Arnaud Messé, Gorka Zamora-López and Heike Siebert for useful comments.
References
[1] 10.1038/s41598-025-88596-y
[2]10.1073/pnas.2117234119
