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

Single-cell CRISPR activation/interference screens offer a direct route to causal gene-regulatory maps, yet existing deep-learning pipelines are GPU-intensive and yield hard-to-interpret latent factors. We introduce LazyNet, an explicitly Euler-integrated neural ODE whose paired log-linear-exp layer collapses multiplicative transcript interactions into a compact, mechanistically interpretable weight matrix. Training a three-replica ensemble on a 55k-cell, 30k-gene Perturb-seq dataset completes on a single CPU in <1 h, running 3 to 4 folder faster than transformer (scGPT) or state-space (RetNet) baselines while lowering global RMSE by ≈ 25 % and raising genome-wide Pearson r to 0.67. Averaged Jacobians, expanded in a 32*4 breadth-first search around seven ferroptosis seeds, recapitulated 15 of 27 benchmark regulators (56 % recall) within a 4 611 gene, 11 676 edge subgraph; 26.6 % of edges show ARCHS4 co-expression r ≥ 0.2 versus 5 % expected at random, and 523 overlap STRING interactions (hypergeometric p = 1.2e-5). Elasticity ranks uncover a previously unrecognized lysosomal-mitochondrial-immune module linking PSAP-mTOR, MFN2-TLR4 and ADCY10-SIRT3, generating experimentally testable hypotheses. By combining state-of-the-art predictive accuracy, laptop-level resource demands and one-to-one parameter interpretability, LazyNet democratizes causal network discovery from sparse two-snapshot screens, enabling small laboratories to move from large-scale perturbation data to mechanistic insight without GPUs or external pathway priors.