Simulation-based inference for efficient identification of generative models in computational connectomics

Jan Boelts, Philipp Harth, Richard Gao, Daniel Udvary, Felipe Yáñez, Daniel Baum, Hans-Christian Hege, Marcel Oberländer, Jakob H. MackePLOS Computational Biology 2023 · DOI

TL;DR

This paper introduces Simulation-Based Inference (SBI) as a powerful tool to efficiently identify the underlying principles (wiring rules) that govern how neurons connect in the brain. By leveraging deep learning and Bayesian inference, SBI can sift through vast numbers of potential wiring rules to find those that best explain empirical connectivity data, offering a more systematic and less labor-intensive approach than traditional methods.

What Problem Does This Paper Address?

Understanding how the brain's intricate network of neuronal connections forms is a fundamental question in neuroscience. Connectomics aims to map these connections, but simply collecting data isn't enough. We need ways to test hypotheses about the "wiring rules" that dictate these connections. A common approach is to build computational models that simulate these rules and then compare the simulated data to real-world measurements. However, these rules often have unknown parameters that need to be "tuned" to match the data. This manual tuning process is tedious and can be challenging because many different parameter settings might produce similar results, making it hard to pinpoint the true underlying principles. This paper addresses the need for a more efficient and systematic way to identify these data-compatible wiring rules and their parameters.

Methods

The core of this paper is the application of Simulation-Based Inference (SBI). Instead of directly calculating the probability of observing data given model parameters (the likelihood, which is often intractable for complex simulations), SBI uses simulations to learn this relationship.

Here's a breakdown of their approach:

  • Parametrization of Generative Models: They take existing generative models for neuronal connectivity and equip them with parameters. These parameters represent variations in the hypothesized wiring rules.
  • Bayesian Inference: They frame the problem as inferring a posterior distribution over these parameters. This distribution represents all parameter combinations that are "data-compatible."
  • Simulation-Based Inference (SBI): Since the likelihood is hard to compute, they use SBI. This involves:
  • Generating Training Data: They simulate many pairs of parameter values and the resulting connectivity data from their generative model.
  • Neural Network Training: They train artificial neural networks (specifically, Neural Spline Flows) to approximate the posterior distribution directly from these simulated parameter-data pairs. This is known as Neural Posterior Estimation (NPE).
  • Amortized Inference: Once trained, the neural network can quickly infer the posterior for new, unseen empirical data without retraining.
  • Validation: They rigorously validate their SBI approach using simulated data with known ground-truth posteriors and through techniques like Simulation-Based Calibration (SBC) to ensure the inferred uncertainties are reliable.

Datasets

The primary dataset used is empirical connectivity measurements from the rat barrel cortex. Specifically, they use seven measured population connection probabilities between the ventral posterior medial nucleus (VPM) of the thalamus and different cell types in the barrel cortex, drawing from existing literature ([34, 35]).

They also generate extensive simulated connectivity data from their generative model for training the SBI algorithms and for validation purposes.

Key Insights and Ideas

  • SBI as a "Sieve" for Wiring Rules: SBI acts like a sophisticated sieve, efficiently filtering through a vast space of potential wiring rule parameters to identify those that are consistent with empirical data. This moves beyond finding a single "best fit" to characterizing all plausible solutions.
  • Parameter Interactions are Biologically Meaningful: The inferred posterior distributions reveal correlations between parameters. These correlations aren't just mathematical artifacts; they highlight how different structural features (like presynaptic bouton density and postsynaptic spine density) interact in biologically plausible ways to influence connectivity.
  • Predicting the Unseen: By inferring a posterior distribution, SBI allows for predictions of connectivity patterns that haven't been directly measured. This is a powerful tool for generating new, testable hypotheses.
  • Quantitative Ruling Out of Hypotheses: SBI provides a rigorous, quantitative method to test and potentially discard existing hypotheses, such as Peters' rule, by comparing their predictions to those derived from the data-informed SBI posterior.

Key Results

  • Identification of Data-Compatible Rules: SBI successfully identified a broad range of parameter configurations for the Dense Structural Overlap (DSO) wiring rule that were compatible with the measured VPM-barrel cortex connection probabilities. This suggests that multiple wiring principles could explain the observed data.
  • Parameter Correlations: They found significant correlations between the parameters of the DSO rule, particularly a strong negative correlation between the influence of presynaptic boutons and postsynaptic targets, and positive correlations involving the normalization factor. These correlations were much stronger than in the prior distribution, indicating that the data strongly constrains parameter relationships.
  • Accurate Predictions: SBI-inferred posteriors accurately predicted held-out connectivity measurements for most cell types, demonstrating the method's predictive power. The accuracy was generally high, with classification accuracy for cross-validation below 0.55 for most cell types.
  • Peters' Rule Disproven: For both neuron-level and synapse-level proximity-based wiring rules, SBI demonstrated that Peters' rule (which posits connectivity based solely on axo-dendritic proximity) predicts connection probabilities that are too high and inconsistent with empirical data, especially when compared to the SBI-inferred rules that incorporate subcellular features.

Limitations and Future Work

  • Model Misspecification: A key limitation is that SBI assumes the generative model is "well-specified" – meaning it can, in principle, generate data similar to the observed data. If the model is fundamentally flawed, the inferred posterior can be biased. The authors recommend prior predictive checks to mitigate this.
  • Computational Cost: While SBI is more efficient than manual tuning, training the neural networks can still be computationally intensive, especially for complex models or high-dimensional parameter spaces. The paper notes that SBI's performance depends on simulation runtime and the number of parameters.
  • Approximation: The inferred posterior distributions are approximations. The accuracy depends on the amount of training data and the flexibility of the neural network architecture.
  • Future Work: The authors suggest that SBI is applicable to a wide range of generative models in connectomics, including those based on dense reconstructions of brain tissue. They also highlight the potential for combining SBI with Bayesian model comparison to simultaneously select the best wiring rule and infer its parameters. Further exploration of using neural networks to automatically learn informative summary features from high-dimensional raw data is also a promising direction.

Summary updated

BibTeX
@article{boelts2023simulation,
  author = {Jan Boelts and Philipp Harth and Richard Gao and Daniel Udvary and Felipe Yáñez and Daniel Baum and Hans-Christian Hege and Marcel Oberländer and Jakob H. Macke},
  title = {{Simulation-based inference for efficient identification of generative models in computational connectomics}},
  journal = {PLOS Computational Biology},
  year = {2023},
  doi = {10.1371/journal.pcbi.1011406},
}
Figure