A Neural-Network-Free Calibration Matches or Beats Deep Single-Cell Perturbation Response Models Across Four Datasets

Area of Science:

• Computational Biology
• Machine Learning
• Single-cell Genomics

Background:

• Deep generative models like scGen are standard for predicting cellular responses to perturbations.
• Assessing model performance under leave-one-group-out (LOCO) cross-validation is crucial for generalization.
• The advantage of deep models over simpler baselines in this context requires further investigation.

Purpose of the Study:

• To determine if the performance advantage of deep generative models can be replicated without neural networks.
• To develop a computationally efficient and robust method for predicting cellular responses.
• To compare the novel method against existing deep learning and baseline approaches.

Main Methods:

• Decomposition of scGen's predictive capability into interpretable components: response magnitude and direction.
• Introduction of an affine moment-matching step for calibration.
• Development of AMM-SimWMag, a neural network-free, CPU-only method.
• Evaluation across 4 datasets, 3 biologies, and 2 LOCO axes using a 9-metric panel and dataset-stratified testing.

Main Results:

• AMM-SimWMag demonstrates competitive or superior performance compared to scGen across 8-9 metrics on all 4 datasets.
• The method significantly outperforms scGen on most metrics, particularly in robustness across different biologies.
• AMM-SimWMag surpasses other modern baselines (biolord, CPA, scPRAM) on multiple metrics.
• The affine moment-matching component enhances distributional similarity metrics.

Conclusions:

• AMM-SimWMag successfully replicates the performance of deep generative models without employing neural networks.
• The proposed method offers a robust, efficient, and reproducible alternative for predicting cellular responses.
• AMM-SimWMag shows greater robustness across diverse biological datasets compared to single deep learning models.