Table 1.
Results from models tested in this work on Kaggle OpenVaccine public leaderboard, private test set, and orthogonal mRNA degradation results. 1Bootstrapped Spearman correlation of degradation rate (resampled from experimental error) to half-life.
Public test set (400 constructs, 27200 nucleotides) | Private test set (1801 constructs, 162316 nucleotides) | mRNA degradation prediction from ref. 6 (188 constructs) | |
---|---|---|---|
Metric | MCRMSE | MCRMSE | Spearman Correlation |
Experimental error | 0.12491 | 0.10571 | 0.88 |
Single Model (blind prediction) | |||
DegScore | 0.39219 | 0.47297 | 0.36 |
DegScore-XGBoost | 0.35854 | 0.43850 | 0.42 |
Nullrecurrent | 0.22758 | 0.34198 | 0.43 |
Kazuki2 | 0.22756 | 0.34266 | 0.48 |
Ensembled models (post hoc) | |||
Genetic algorithm (10 of top 100 selected) | 0.2237 | 0.3397 | |
Ensemble top 2 models | 0.2244 | 0.33788 | 0.47 |
Genetic algorithm on private test set | 0.3382 | -- |