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

## TL;DR

LazyNet is a new tool that helps labs analyze gene-editing experiments with limited resources, revealing how genes work together and linking findings to known biological processes.

## Contribution

LazyNet introduces an interpretable ODE-based model for sparse CRISPR screens, enabling accurate and resource-efficient analysis of gene interactions.

## Key findings

- LazyNet achieved strong predictive performance on a large Perturb-seq dataset using only CPU resources and limited training time.
- The model identified gene networks consistent with public resources and proteomic data, including known regulators of ferroptosis.
- LazyNet outperformed transformer and state-space models in low-data conditions with fewer parameters and computational resources.

## Abstract

Most labs study how genes change when a single gene is switched on or off, but they rarely have the budget to collect long, repeated measurements or to standardize pipelines across groups. We built a tool that works directly on a lab’s own “before and after” gene-editing experiments and turns them into clear, mechanism-level readouts of cause and effect. In tests, our method produced accurate predictions under tight time and hardware limits and revealed when genes act together rather than one at a time. The networks it learned agreed with outside evidence from large public gene resources and independent protein measurements, and they recovered many known regulators in ferroptosis. Because the same workflow can be rerun on other datasets and on standard CPU-only hardware (no GPU needed), small teams can analyze their own data, compare to public studies, and plan sharper follow-up experiments. Our results are therefore most relevant for labs that must train models from scratch on their own data; when large pretrained models can be fine-tuned on massive public datasets, they remain powerful complementary options.

We present LazyNet, a compact one-step neural-ODE model for single-cell CRISPR activation/interference (A/I) that operates directly on two-snapshot (“pre → post”) measurements and yields parameters with clear mechanistic meaning. The core log–linear–exp residual block exactly represents multiplicative effects, so synergistic multi-locus responses appear as explicit components rather than opaque composites. On a 53k-cell × 18k-gene neuronal Perturb-seq matrix, a three-replica LazyNet ensemble trained under a matched 1 h budget achieved strong threshold-free ranking and competitive error (genome-wide r ≈ 0.67) while running on CPUs. For comparison, we instantiated transformer (scGPT-style) and state-space (RetNet/CellFM-style) architectures from random initialization and trained them from scratch on the same dataset and within the same 1 h cap on a GPU platform, without any large-scale pretraining or external data. Under these strictly controlled, low-data conditions, LazyNet matched or exceeded their predictive performance while using far fewer parameters and resources. A T-cell screen included only for generalization showed the same ranking advantage under the identical evaluation pipeline. Beyond prediction, LazyNet exposes directed, local elasticities; averaging Jacobians across replicas produces a consensus interaction matrix from which compact subgraphs are extracted and evaluated at the module level. The resulting networks show coherent enrichment against authoritative resources (large-scale co-expression and curated functional associations) and concordance with orthogonal GPX4-knockout proteomes, recovering known ferroptosis regulators and nominating testable links in a lysosomal–mitochondrial–immune module. These results position LazyNet as a practical option for from-scratch, low-data CRISPR A/I studies where large-scale pretraining of foundation models is not feasible.