Improving ΔΔG Predictions with a Multitask Convolutional Siamese Network