Fig. 3: Prediction accuracy of deep neural networks.

A Architecture of the convolutional neural network (CNN) employed in this paper; the output is the predicted sfGFP fluorescence in relative units. The CNN architecture was designed with Bayesian optimization³⁵ to find a single architecture for all mutational series; our strategy for hyperparameter optimization can be found in the Methods, Supplementary Fig. S3, and Supplementary Tables S4–S5. B Accuracy of the CNN in panel A trained on all mutational series. R² values were computed on held-out sequences (10% of total) and averaged across 5 training repeats; bars denote the mean R². C Prediction accuracy of CNNs against random forest (RF) and multilayer perceptrons (MLPs) on all 56 mutational series using binary one-hot encoding. The CNNs yield more accurate predictions with the same training data. Violin plots show the distribution of 56 R² values for each model averaged across 5 training repeats; R² values were computed on held-out sequences (10% of sequences per series). For all violins, the white circle indicate the median, box edges are at the 25th and 75th percentiles, and whiskers show 95% confidence interval. Inset shows predictions of a CNN trained on 75% of the mutational series with a right-skewed phenotypic distribution (Fig. 1B) computed on held-out test sequences. The CNNs are more complex than the shallow MLPs (2,702,337 vs 58,801 trainable parameters, respectively), but we also found that the CNNs outperform MLPs of comparable complexity (Supplementary Fig. S8); this suggests that improved performance is a result of the convolutional layers acting as a feature extraction mechanism. Details on CNN training can be found in the Methods and Supplementary Fig. S7. D Average R² scores for each model across all 56 mutational series using 75% of sequences for training.