Periodic Reporting for period 1 - DeepCoMechTome (Using deep learning to understand computations in neural circuits with Connectome-constrained Mechanistic Models)
Período documentado: 2023-07-01 hasta 2025-12-31
Recent advances in experimental neuroscience now make it possible to map complete wiring diagrams, or connectomes, of neural circuits in model organisms such as the fruit fly (Drosophila melanogaster). These datasets provide an unprecedented opportunity to understand how the brain’s structure gives rise to its function. Yet, knowing the connectivity alone does not reveal how groups of neurons process information and generate behaviour. Achieving this requires models that combine the anatomical and biophysical detail of mechanistic neuroscience with the adaptability and optimisation strengths of deep learning methods from artificial intelligence.
Currently, the two modelling traditions are largely separate. Mechanistic models faithfully incorporate anatomy and biophysics, but methods for optimising them to perform complex computations are lacking. Deep learning models can be trained to perform such computations efficiently, but they often lack mechanistic interpretability and biological plausibility. The DeepCoMechTome project addresses this gap by developing a unified machine learning framework that integrates the strengths of both approaches.
Our aim is to algorithmically identify “deep mechanistic models” that respect anatomical and biophysical constraints, reproduce experimental neural data, and perform computations relevant for behaviour. This work sits at the interface of neuroscience and AI: by constraining deep learning models with biological data, we can uncover how biological intelligence achieves highly robust and energy-efficient computation—capabilities that remain out of reach for most artificial systems.
We apply this framework to neural circuits that process visual information and use it to guide behaviour, for example in visually driven locomotion. Using detailed connectome and physiological data, we build models that can be optimised for these tasks, analysed to reveal underlying computational principles, and used to make testable predictions about neural tuning and behavioural output. This creates a virtuous cycle in which models guide experiments and experimental results refine models.
Although our initial focus is on the fruit fly, the methodology is general and can be applied to a wide range of neural circuits and species. By linking structure, dynamics, and behaviour, DeepCoMechTome will provide new insights into the fundamental principles of biological intelligence and inform the design of artificial systems that share the brain’s robustness and energy efficiency.
Currently, the two modelling traditions are largely separate. Mechanistic models faithfully incorporate anatomy and biophysics, but methods for optimising them to perform complex computations are lacking. Deep learning models can be trained to perform such computations efficiently, but they often lack mechanistic interpretability and biological plausibility. The DeepCoMechTome project addresses this gap by developing a unified machine learning framework that integrates the strengths of both approaches.
Our aim is to algorithmically identify “deep mechanistic models” that respect anatomical and biophysical constraints, reproduce experimental neural data, and perform computations relevant for behaviour. This work sits at the interface of neuroscience and AI: by constraining deep learning models with biological data, we can uncover how biological intelligence achieves highly robust and energy-efficient computation—capabilities that remain out of reach for most artificial systems.
We apply this framework to neural circuits that process visual information and use it to guide behaviour, for example in visually driven locomotion. Using detailed connectome and physiological data, we build models that can be optimised for these tasks, analysed to reveal underlying computational principles, and used to make testable predictions about neural tuning and behavioural output. This creates a virtuous cycle in which models guide experiments and experimental results refine models.
Although our initial focus is on the fruit fly, the methodology is general and can be applied to a wide range of neural circuits and species. By linking structure, dynamics, and behaviour, DeepCoMechTome will provide new insights into the fundamental principles of biological intelligence and inform the design of artificial systems that share the brain’s robustness and energy efficiency.
During the reporting period, the DeepCoMechTome project has advanced on all major scientific and technical fronts. We developed and integrated methods for building, training, and analysing large-scale mechanistic neural network models constrained by measured connectomes, neural recordings, and biophysics.
For model construction, we made significant progress on a machine learning framework for identifying ensembles of connectome-constrained mechanistic models that can solve behaviourally relevant tasks, including the development and dissemination of powerful simulation-based inference methods for estimating model parameters.
For simulation infrastructure, we extended our connectome-scale simulation framework for simplified neuron models to more detailed biophysical models. This led to the development of Jaxley, a GPU-accelerated, differentiable simulator for detailed neuron and synapse models (Deistler et al, BiorXiv 2024). Jaxley enables efficient gradient-based optimisation of multi-compartment models under both data and task constraints.
We applied our modelling framework to the fruit fly visual system, culminating in the first large-scale mechanistic model of the Drosophila motion detection pathway constrained by both connectomic structure and a behavioural goal (motion detection).
This model accurately reproduced neural activity across multiple independent datasets (Lappalainen et al., Nature 2024), demonstrating that anatomical connectivity combined with task constraints can yield precise and testable neuron-level predictions.
The project has thus delivered both major methodological advances—particularly in differentiable biophysical simulation and simulation-based inference—and a flagship application that demonstrates the feasibility and power of our approach. These results provide a foundation for applying our unified framework to more complex connectomes and behaviours in the next reporting period.
For model construction, we made significant progress on a machine learning framework for identifying ensembles of connectome-constrained mechanistic models that can solve behaviourally relevant tasks, including the development and dissemination of powerful simulation-based inference methods for estimating model parameters.
For simulation infrastructure, we extended our connectome-scale simulation framework for simplified neuron models to more detailed biophysical models. This led to the development of Jaxley, a GPU-accelerated, differentiable simulator for detailed neuron and synapse models (Deistler et al, BiorXiv 2024). Jaxley enables efficient gradient-based optimisation of multi-compartment models under both data and task constraints.
We applied our modelling framework to the fruit fly visual system, culminating in the first large-scale mechanistic model of the Drosophila motion detection pathway constrained by both connectomic structure and a behavioural goal (motion detection).
This model accurately reproduced neural activity across multiple independent datasets (Lappalainen et al., Nature 2024), demonstrating that anatomical connectivity combined with task constraints can yield precise and testable neuron-level predictions.
The project has thus delivered both major methodological advances—particularly in differentiable biophysical simulation and simulation-based inference—and a flagship application that demonstrates the feasibility and power of our approach. These results provide a foundation for applying our unified framework to more complex connectomes and behaviours in the next reporting period.
DeepCoMechTome has shown how combining simulation-based inference and gradient-based optimisation with detailed neural circuit data can yield powerful, testable models of biological intelligence. By integrating structural connectivity, neural activity, and behaviour within a single modelling framework, the project has demonstrated a general approach for extracting mechanistic insight from complex biological systems.
The methodological advances—particularly in SBI and differentiable simulation—are not limited to neuroscience. They can be applied wherever mechanistic models are combined with experimental data, from physics and climate science to systems biology and engineering. In neuroscience, these tools make it possible to turn large-scale connectome datasets into predictive, interpretable models; in other domains, they provide a flexible way to identify and refine models under real-world constraints.
Future progress will benefit from extending the framework to more complex connectomes, improving interoperability with diverse datasets, and continuing open-source development to ensure adoption across disciplines. Strengthening interdisciplinary collaborations will be essential for exploiting the full potential of these methods.
By the end of the project, DeepCoMechTome has delivered both proof-of-principle applications—most notably in the Drosophila visual system—and a robust set of computational tools. These outcomes form a foundation for future work using SBI and gradient-based methods to bridge mechanistic modelling and data analysis in a wide range of scientific fields.
The methodological advances—particularly in SBI and differentiable simulation—are not limited to neuroscience. They can be applied wherever mechanistic models are combined with experimental data, from physics and climate science to systems biology and engineering. In neuroscience, these tools make it possible to turn large-scale connectome datasets into predictive, interpretable models; in other domains, they provide a flexible way to identify and refine models under real-world constraints.
Future progress will benefit from extending the framework to more complex connectomes, improving interoperability with diverse datasets, and continuing open-source development to ensure adoption across disciplines. Strengthening interdisciplinary collaborations will be essential for exploiting the full potential of these methods.
By the end of the project, DeepCoMechTome has delivered both proof-of-principle applications—most notably in the Drosophila visual system—and a robust set of computational tools. These outcomes form a foundation for future work using SBI and gradient-based methods to bridge mechanistic modelling and data analysis in a wide range of scientific fields.