Predicting Neuron Morphology from Single-Cell Gene Expression

Abstract

The intricate relationship between gene expression and neuronal morphology is fundamental to understanding brain function and development. This dissertation introduces deep learning approaches that bridge the gap between transcriptomics and neuronal structure. We present 2D- and 3D- MorphNet, which leverages state-of-the-art generative models to predict cell morphology from single-cell gene expression profiles.

2D-MorphNet combines VAE and GANs to generate realistic 2D morphological images of neurons and nuclei from gene expression data. The model demonstrates robust performance on spatial transcriptomic datasets, including MERFISH and Patch-seq, achieving a FID of 15.82 on held-out test data. Notably, 2D-MorphNet generalizes to predict morphologies across diverse brain regions, including the motor cortex, isocortex, and BNST. To overcome limited training data, we employ ADA, enabling high-quality image generation with as few as 1,000 training samples.

3D-MorphNet significantly extends these capabilities to full three-dimensional neuron morphologies, utilizing an adapted DiT architecture to generate complete structural representations in the standardized SWC format. This innovation is particularly impactful as it allows for the direct integration of predicted morphologies into existing computational neuroscience workflows and databases. 3D-MorphNet outperforms existing methods in capturing a wide range of morphological features, achieving superior performance in 20 out of 22 calculated morphometrics compared to state-of-the-art alternatives. The model demonstrates remarkable fidelity in reconstructing complex neuronal structures, with a Chamfer distance of 40.43 (plus/minus 20.43) on the validation set, significantly outperforming baseline methods.

Our models exhibit the ability to perform meaningful latent space interpolation, enabling in silico experimentation to explore the effects of gene expression manipulations on neuronal morphology. This capability allows researchers to predict potential morphological changes resulting from genetic perturbations, offering a powerful tool for hypothesis generation in neurodevelopmental and neurodegenerative disease research.

Furthermore, 3D-MorphNet shows promise in reconstructing full morphologies from partial data, addressing a critical challenge in neuroanatomical studies where complete reconstructions are often unavailable. The model successfully predicts missing axonal structures while maintaining the integrity of existing dendritic trees, demonstrating its potential to enhance and complete partial neuron reconstructions in large-scale brain mapping projects.

To facilitate broader adoption and impact, we have developed a web tool that allows users to predict neuron morphologies from their own scRNA-seq data. This tool democratizes access to advanced morphological prediction capabilities, enabling researchers without extensive computational expertise to generate hypotheses about cellular morphology based on transcriptomic data.

The MorphNet framework represents a significant advancement in computational neuroscience, providing powerful tools for augmenting large-scale single-cell atlases with 2D and 3D morphologies. By bridging the gap between molecular profiles and cellular architecture, this work opens new avenues for investigating the molecular basis of neuronal diversity and function. The ability to generate accurate, diverse, and biologically plausible neuronal structures from gene expression data paves the way for a deeper understanding of brain organization, circuit formation, and the impact of genetic variations on neuronal morphology.

By leveraging advanced machine learning techniques and biological insights, MorphNet provides a robust platform for exploring the complex interplay between genetic programs and structural phenotypes in the nervous system. This work not only advances our understanding of brain architecture but also provides valuable tools for the broader neuroscience community, potentially accelerating discoveries in both fundamental and translational neuroscience research.