P125 Modeling the biophysics of computation in the outer plexiform layer of the mouse retina
Kyra L. Kadhim*1, 2, Ziwei Huang1, Michael Deistler2, 3, Jonas Beck1, 2, Jakob H. Macke1, 2, 3, Thomas Euler4, Philipp Berens1, 2
1Hertie Institute for AI in Brain Health, University of Tübingen,Tübingen,Germany
2Tübingen AI Center,University of Tübingen,Tübingen, Germany
3Machine Learning in Science, University of Tübingen,Tübingen, Germany
4Institute for Ophthalmic Research, University of Tübingen,Tübingen,Germany
*Email: kyra.kadhim@uni-tuebingen.de
Introduction
The outer retina is a complex system of neurons that processes visual information, and it is experimentally accessible for the collection of multimodal data. What makes this system complex and nonlinear are mechanisms such as the phototransduction cascade, specialized ion channels, ephaptic feedback mechanisms, and the ribbon synapse [1]. These mechanisms are not typically included in network models that either fit neural data or perform tasks. In particular, optimizing the parameters of such models is computationally challenging with current modelling approaches, which do not include gradient-based optimization methods. However, ignoring such mechanisms limits the ability to capture the computations performed by the retina.
Methods
We developed a fully-differentiable, biophysically-detailed model of the outer plexiform layer of the mouse retina and optimized its parameters with gradient descent. We implemented our model using the new software library Jaxley [2] which inherits functionality from the state of the art machine learning library JAX. In our model, we have so far implemented the phototransduction cascade [3], ion channels [4], and ribbon synapse [5], and we fit their parameters to electrophysiology and neurotransmitter release data. We then optimized the synaptic conductances of the model to classify images with different contrast levels and global luminance levels and analyzed the trained parameter parameter distributions.
Results
We successfully trained our model of a single photoreceptor with gradient descent and found phototransduction cascade parameters that fit the electrophysiology data from Chen and colleagues [3], as well as parameters of the ribbon synapse model that fit glutamate release data from Szatko, Korympidou, and colleagues [6]. We then built a network of photoreceptors with these trained parameters and a horizontal cell, and we trained the network’s 200 synaptic conductances to classify images in which contrast and global luminance levels were distorted. The model was able to classify these images despite these distortions, providing further evidence that the structure of the outer retina facilitates contrast normalization.
Discussion
Biophysical models are capable of implementing a variety of computations that are often attributed to larger neural networks higher in the sensory processing hierarchy. For instance, the fitted model of the phototransduction cascade enables a layer of photoreceptors to adapt to drastically different global luminance levels [3] while at the same time regulating glutamate release consistent with data. Our model, fit to multimodal data, can also classify images with different contrasts using very few trainable parameters. This small but biophysically-inspired network may support many other computations as well, broadening our appreciation of the outer retina.
Acknowledgements
Hertie Stiftung, DFG, ERC Starting Grants NextMechMod and DeepCoMechTome)
References
● https://doi.org/10.1016/C2019-0-00467-0
● https://doi.org/10.1101/2024.08.21.608979
● https://doi.org/10.7554/eLife.93795.1
● https://doi.org/10.1016/j.visres.2009.03.003
● https://doi.org/10.7554/eLife.67851
● https://doi.org/10.1038/s41467-020-17113-8
