Balanced Training Sets Improve Deep Learning-Based Prediction of CRISPR sgRNA Activity

Figures

• 

  Figure 1. DeepGuide performance with imbalanced CRISPR-Cas12a training data sets. (a, d) Normalized cutting score (CSnorm) distributions of the Cas12a training data set imbalanced by removing 0, 25, 50, 75, and 90% high- and low-activity sgRNAs along with the total number of sgRNAs (n) in every training set. (b, e) Performance of DeepGuide models on the sgRNA test sets (Pearson’s r), and high- and low-activity Cas12a sgRNAs from individual phenotype screening experiments (TPR, 1-FPR). Bars represent mean values of Pearson’s r, TPR, and 1-FPR across five independent runs (n = 5). Error bars indicate one standard deviation, and data points represent values from each individual run. (c, f) Mean predicted CS of high-activity (KO efficiency ≥ 50%) and low-activity (KO efficiency < 50%) sgRNAs from individual phenotype screening experiments in Y. lipolytica when DeepGuide was trained on imbalanced data sets with 90% high-activity and 90% low-activity sgRNA removed. Data points represent mean values of predicted CS for experimental sgRNA with a given KO efficiency across five independent runs (n = 5), and error bars indicate one standard deviation. Dashed line represents average predicted CS…

• 

  Figure 2. DeepGuide performance with imbalanced CRISPR-Cas12a training data sets augmented with synthetic sgRNAs. (a, d) Pie charts showing change in composition of imbalanced training sets skewed toward low- and high-activity sgRNAs after adding synthetic (a) high-activity and (d) low-activity sgRNAs. (b, e) Performance of DeepGuide models on the test set of sgRNAs (Pearson’s r), and high-activity and low-activity Cas12a sgRNAs from individual phenotype screening experiments (TPR, 1-FPR), when trained using the original training set, imbalanced training sets obtained after removing 50% and 90% (b) high- and (e) low-activity sgRNAs, and rebalanced training sets obtained after adding synthetic (b) high- and (e) low-activity sgRNAs. Bars represent mean values of Pearson’s r, TPR and 1-FPR across five independent runs (n = 5), error bars indicate one standard deviation, and data points represent values from each individual run. (c, f) Mean predicted CS of high-activity (KO efficiency ≥ 50%) and low-activity (KO efficiency < 50%) sgRNA from individual phenotype screening experiments in Y. lipolytica when DeepGuide was trained on balanced data sets containing synthetic (c) high- and (f)…

• 

  Figure 3. Composition of the Y. lipolytica (top) and K. phaffii (bottom) CRISPR-Cas9 training sets and DeepGuide performance with the two data sets. (a, e) Normalized cutting score (CSnorm) distributions of the original Cas9 training data sets for Y. lipolytica and K. phaffii. (b, f) Pie charts showing the proportion of high- and low-activity sgRNAs in the original Cas9 training sets containing a total of 19953 sgRNA for Y. lipolytica and 27821 sgRNA for K. phaffii. (c, g) Performance of DeepGuide on the test set of sgRNA for Y. lipolytica and K. phaffii when trained on the respective original training sets and rebalanced training sets obtained after adding synthetic low-activity sgRNA to the original sets. Bars represent mean Pearson’s r across five independent runs (n = 5), error bars indicate one standard deviation, and data points represent values from each individual run. (d, h) DeepGuide performance on high- and low-activity Cas9 sgRNAs from individual experiments for Y. lipolytica and K. phaffii. Bars represent mean values of TPR and 1-FPR across five independent runs (n = 5), error bars indicate one standard deviation, and data points represent values from each individual run.