LazyNet: Interpretable ODE Modeling of Sparse CRISPR Single-Cell Screens Reveals New Biological Insights

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Single-cell CRISPR screens are powerful for dissecting gene function.
• Analyzing large-scale CRISPR A/I data requires efficient and interpretable models.
• Existing models may struggle with low-data regimes or lack mechanistic interpretability.

Purpose of the Study:

• To introduce LazyNet, a compact one-step neural ordinary differential equation (neural-ODE) model for single-cell CRISPR A/I.
• To enable direct analysis of two-snapshot (pre- and post-perturbation) measurements.
• To yield model parameters with clear mechanistic interpretations.

Main Methods:

• Developed LazyNet with a core log-linear-exp residual block to model multiplicative gene effects.
• Applied LazyNet to a large neuronal Perturb-seq dataset (53k cells x 18k genes).
• Compared LazyNet's performance against transformer and state-space models trained under identical low-data, short-time budget conditions on CPUs and GPUs.

Main Results:

• LazyNet achieved competitive prediction error (genome-wide r ≈ 0.67) and strong ranking performance on CPUs within a 1-hour budget.
• LazyNet matched or exceeded transformer and state-space models in predictive performance under strictly controlled low-data conditions, using fewer parameters and resources.
• The model successfully identified gene regulatory networks, enriched against known biological resources, and recovered known regulators of ferroptosis.

Conclusions:

• LazyNet offers a practical, resource-efficient solution for analyzing CRISPR A/I data, especially in low-data scenarios without extensive pretraining.
• The model's architecture allows for mechanistic interpretation of gene interactions.
• LazyNet demonstrates strong generalization capabilities and potential for discovering novel biological insights.