Comparison of semisupervised scNym to other single-cell classification methods and ablated scNym variants. (A) We assign each method a rank order (rank 1 is best) based on performance for each benchmark task. scNym is the top-ranked method across tasks and ranks highly on all tasks. A support vector machine (SVM) baseline is the next best method, consistent with a previous benchmarking study (Abdelaal et al. 2019). (B) Ablation experiments comparing simplified supervised scNym models (Base) against the full scNym model with semisupervised and adversarial training (SSL + Adv.). We found that semisupervised and adversarial training significantly improved scNym performance across diverse tasks (all tasks shown, Wilcoxon rank sums, P < 0.05).
