Periodic Reporting for period 1 - NextMechMod (Next generation mechanistic models of retinal interneurons)
Période du rapport: 2023-01-01 au 2025-06-30
Ever since the work Hodgkin and Huxley, models of neurons have been essential for our understanding of neural computations. Such models have been developed at diverse levels of “realism”, from linear-nonlinear cascade or black-box models to detailed compartmental models. While these approaches are commonly viewed as incompatible, they have attractive strengths from an epistemic point of view. In this project, I propose to develop a new generation of “hybrid” mechanistic models that reconcile these levels of modelling: they will consist of a compartmental model for the neuron of interest with inputs approximated by black-box models. I will leverage the power of these hybrid models to tackle one of the most challenging questions in visual neuroscience: the staggering diversity of amacrine cells, a major class of inhibitory interneurons in the vertebrate retina. Despite their diversity, they are the least understood class of neurons in the retina, in stark contrast to the remaining circuitry. I will build on the latest advances in machine learning to develop a framework for efficiently inferring the parameters of a hybrid mechanistic model. To constrain the model parameters, we will acquire two-photon calcium and voltage imaging data during natural stimulation. Further, we will extend our framework to incorporate transcriptomic information about gene expression collected via patch-seq into the inference procedure, allowing us to map the amacrine cells to genetically defined types. Thus, in this project, I propose to develop a toolset to systematically uncover the role of retinal amacrine cells during natural visual computations, and link it to its mechanistic basis, providing a path forward to solving one of the key remaining mysteries of visual neuroscience.
We worked on algorithms to make inference for the mechanistic component of the hybrid AC models more efficient. To this end, we evaluated different strategies for inference of Hodgkin-Huxley models and proposed diffusion tempering, a novel regularization technique for probabilistic numerical methods. In addition, we implemented a new framework for simulation of detailed biophysical neuron models —Jaxley—which makes use of automatic differentiation and GPU acceleration. Jaxley opens up the possibility to efficiently optimize large-scale biophysical models with gradient descent. We showed that Jaxley can learn parameters of biophysical neuron models with several hundreds of parameters to match voltage or two photon calcium recordings, sometimes orders of magnitude more efficiently than previous methods. Also, we started recording two-photon imaging data from dendrites of GABAergic AC types using genetic calcium sensors. Using these initial datasets, we identified 25 functional types with distinct chromatic and achromatic properties. We used pharmacology and a biologically inspired circuit model to explore how inhibition and excitation shape the properties of functional types. We also contributed to the development of a new method for extracting most discriminative visual stimuli from functional clusterings (Burg et al. 2024). We started creating a library of ion channels, synapses and other mechanisms including many specifically useful for the retina, available at https://github.com/jaxleyverse/jaxley-mech(s’ouvre dans une nouvelle fenêtre).
We introduced Jaxley, a new computational framework for differentiable simulation of neuroscience models with biophysical detail. Unlike previous biophysical simulation toolboxes, Jaxley enables automatic differentiation through its differential equation solver, making it possible to use backprop to compute the gradient with respect to virtually any biophysical parameter. We think this has the potential to be a breakthrough for biophysical modelling as automatic differentiation and computational frameworks which provide efficient, scalable, and easy-to-use implementations, have been key to the deep learning revolution. We demonstrated that gradient descent allows to fit biophysical models of neural dynamics to large datasets of experimental voltage and calcium recordings and that it enables training biophysical models to perform physiologically meaningful computations, with as many as 100k parameters. We expect that Jaxley will enable a range of new investigations in neuroscience: It will make it possible to efficiently optimize detailed single-cell models. This will allow insights into cellular properties across cell types and their relationship with, for example, transcriptomic measurements. We are currently working on building a community around jaxley and are talking to international colleagues who are looking into adapting it to their needs.