Deep Learning Analysis on Images of iPSC-derived Motor Neurons Carrying fALS-genetics Reveals Disease-Relevant Phenotypes

Installation

This project uses pixi to install and manage dependencies. To install pixi, run:

curl -fsSL https://pixi.sh/install.sh | bash

You can then set up the project using the setup make target.

make setup

Overview

This repo contains the code to reproduce the figures from the paper "Deep Learning Analysis on Images of iPSC-derived Motor Neurons Carrying fALS-genetics Reveals Disease-Relevant Phenotypes". The code is organized into the following directories:

• src/fals: contains utility functions and scripts for data loading, preprocessing, and visualization
• notebooks: stand-alone notebooks for reproducing the figures in the paper

Usage

To run JupyterLab and explore notebooks:

make notebook

Data

Data to reproduce the figures from the paper is available from AWS s3 in the s3://2025-fals bucket. The config file refers to this path and can be overridden by setting the ROOT_DATA_PATH environment variable if you chose to download the data locally.

Figures

Below is a summary of the key figures reproduces by the notebooks, illustrating important findings on phenotypic differences between mutuant and wild-type cells lines via statistical analysis, ML classification, and RNA imputation.

Figure 2

• 2.E: dose-response curves of TDP-43 mislocalization in four stressor conditions and three seeding densities
• 2.E, F (supp.): STMN2 intensity within the soma and neurites

Figure 3

• 3.A: cell density vs. TDP-43 mislocalization correlation
• 3.B: TDP-43 mislocalization mask ratio coefficients using a linear model accounting for density across all wells
• 3.D: WT/VCP-R115C het regression coefficients
• 3.F: C9ORF72/corrected regression coefficients
• 3.A (supp.): density matching-based well replicates selection schematic
• 3.B (supp.): TDP-43 mislocalization heatmap with density-matched comparisons
• 3.D (supp.): STMN2 intensity vs. cell density
• 3.E (supp.): STMN2 intensity coefficients heatmap from a linear model accounting for donor and live cell density
• 3.F, G (supp.): soma and neurite STMN2 expression intensity under basal conditions

Figure 4

• 4.A: correlation between predicted and actual TDP-43 mislocalization values derived from morphology embeddings
• 4.D, E: measured vs. imputed change of expression for ALS genes and all genes
• 4.A (supp.): density histogram of the TDP-43 localization as predicted from DAPI and TUJ1
• 4.D (supp.): RNA imputation from morphology-derived embeddings of each familial ALS mutant

Figure 5

• 5.A: classifier accuracy differentiating mutant from WT cells using different feature sets
• 5.C: mean accuracy across mutants in a donor hold-out regime with different feature combinations
• 5.E: classifier accuracy using the TDP-43 mislocalization ratio versus all features
• 5.F: power analysis of phenotypic reversion in a simulated C9ORF72 mutant screen