2026 Theses Doctoral

Computational methods for optical interrogation of neural circuits

Antin, Benjamin

This thesis describes the development of computational tools to enable precise characterization of neural circuits. The common theme is that developing computational approaches in tandem with experimental approaches enables experiments which would be difficult or impossible otherwise. Computational approaches therefore have the potential to inform not just data analysis, but data collection itself.

Chapter 1 deals with monosynaptic connectivity mapping using optical stimulation. Two-photon (2p) optogenetics allows direct excitation of possible presynaptic neurons, while postsynaptic responses are measured using patch clamp. Our lab developed a technique, inspired directly by compressed sensing techniques, which vastly accelerates the speed at which mapping can take place by stimulating ensembles of several neurons simultaneously. In a companion project, which is the subject of Chapter 1, we developed an algorithm which allows mapping to proceed even in the presence of direct photocurrent artifacts.

Chapter 2 deals with reconstructing dendritic fluorescence signals. With few exceptions, microscopy techniques which yield 3D signals (e.g volumetric 2p imaging) are too slow to record signals from many dendritic branches at high speed (>1Khz). By contrast, widefield imaging techniques acquire many pixels simultaneously, but do not have intrinsic 3D resolution; branches of the tree which overlap in space may have their signals mixed together. Our method, termed DENDRO, overcomes this tradeoff by combining information from two data sources: an anatomical description of the dendrite, along with a wide-field functional recording. Using the known anatomy of the tree, we show that it is possible to reconstruct 3D fluorescence signals even though the recorded functional movies do not include 3D information.

Finally, Chapter 3 presents unpublished work on automated construction of biophysical models using voltage imaging data. Though not fully realized here, we predict that it may soon be possible to automatically fit parameters of Hodgkin-Huxley style multicompartment models so that they produce the same patterns observed in data.

Taken together, these projects demonstrate that computational methods can be used to turn unruly optical datasets, which may be corrupted by noise or artifacts, into useful observables for neuroscience.