#13 · ML / AI methods

Benchmarks and deep models for CRISPR editing, single-cell methylation and scGPT tuning

The dominant thread is benchmarking and optimization of existing computational frameworks against curated experimental data. Several items evaluate methods directly: five scDNA methylation imputation methods across 13 datasets, CROP across 18 curated datasets, and active learning strategies across protein and DNA regulatory libraries. Others tune or repurpose models, including Bayesian TPE hyperparameter optimization of scGPT fine-tuning on multi-omics data and VirEvo's use of AlphaGenome-based virtual dual-luciferase scoring with genetic evolution for promoter engineering. CRISPR editing is a recurring domain, spanning DeepCC9's interpretable sgRNA-based efficiency prediction, CROP's frameshift prediction, and the CRISPRko versus CRISPRi Perturb-seq comparison, which benchmarks experimental screen modalities rather than a computational method. What is new is the emphasis on quantitative comparison and interpretability within specific genomic and single-cell regimes, rather than new model architectures.