Datasets for: Architectural biomimicry and motif diversity enhance generalization in CRISPR-Cas9 on-target activity prediction

The accurate prediction of CRISPR-Cas9 single-guide RNA (sgRNA) on-target cleavage efficiency is critical for the design of reliable genome editing experiments. Recently, predictive modeling has gravitated toward massive parameter scaling and exhaustive feature engineering, utilizing dense multi-model ensembles and foundational genomic language models. While these approaches achieve high accuracy on familiar datasets, their excessive parameterization and tendency to mask underlying data sparsity often result in poor generalization on out-ofdistribution (OOD) sequences. In this study, we apply the principle of Occam's Razor to CRISPR-Cas9 on-target activity prediction, hypothesizing that architectural biomimicry and training data motif diversity are critical determinants of generalization, obviating the need for massive architectural scaling. To systematically evaluate this, we construct a rigorous ladder of architectural complexity, benchmarking massive language models against lightweight, biomimetic neural networks. We introduce a highly efficient, sequence-only CNN-GRU architecture that structurally parallels Cas9 target interrogation. By processing the target sequence toward the PAM, our puresequence model utilizes only ~269,000 parameters, yet achieves predictive power and bench utility on diverse external benchmarks that is statistically indistinguishable from massive foundational models when trained under high motif diversity (IID) conditions. Furthermore, we demonstrate that when strict out-of-distribution (OOD) generalization is required, integrating exactly one established biophysical constraint, the thermodynamic binding energy (𝛥𝐺𝐵), acts as a necessary compensatory mechanism when training sets fail to represent the broader biological landscape. However, we demonstrate that when training data exhibits high motif diversity, the pure sequence model, devoid of auxiliary thermodynamic constraints, is optimal due to its structural parsimony. Ultimately, we demonstrate that the simplest, thermodynamically unaided architecture architecturally approximates the directional mechanics of Cas9, establishing that future predictive robustness relies on improving motif diversity rather than architectural complexification.