In this repo, we provide source data, trained models, and NN-derived mutationally-determined phenotype data as a supplement to the paper
Bo Wang & Eric R. Gamazon*.
*Correspondence to:
ericgamazon@gmail.com
A We evaluated three biochemical phenotypes generated from Deep Mutational Scanning (DMS) experiments: binding affinity, protein expression, and antibody escape. These biochemical phenotypes may determine the evolutionary trajectory of the coronavirus, viral infection, and maladaptive host response. B We quantified the extent to which state-of-the-art neural network methodologies can predict these phenotypes, using sequence mutation data. The predicted biomedical phenotypes are mutation- or genetically-determined traits, which are therefore of intrinsic interest because they are determined by the (amino-acid) sequence data. We encoded the protein sequence input using AAindex, consisting of 566 intrinsic (data-independent) physicochemical properties of the amino acids, such as hydrophobicity and long-range non-bonded energy per atom, and one-hot encoding. We then trained neural networks on the mutation-phenotype data from the DMS experiments. C We trained sequence convolutional neural networks and graph convolutional neural networks (along with [fully connected] multilayer perceptrons and linear regression) on the DMS data. Shown are the convoluted sequences and (3-dimensional) structural motifs, differentiated by color, for the two types of architecture, respectively.
Source datasets are curated and fed into a customized neural network engine with four architectures including: CNN-2D, GCN-AVE, MLP and LR.
A joint Random Forest Regression model has been implemented to evaluate the feature importance of the convolutional neural network derived (i.e., estimated, mutation-mediated) antibody-escape phenotype for each of the 10 antibodies in predicting binding affinity towards the ACE2 receptor. Optimized parameters are achieved after grid search.
Available under the MIT License:
Neural Network engine nn4dms, paper Copyright © 2020 Sam Gelman, Philip Romero, Anthony Gitter
GCN-AVE pipgcn, paper Copyright © 2020 Alex Fout.
gast=0.2.2
python=3.6.12
matplotlib=3.3.1
numpy=1.16.4
conda-forge::jupyterlab=2.2.8
pyyaml=5.3.1
pandas=1.1.1
seaborn=0.10.1
tensorflow=1.14
tqdm=4.48.2
scikit-learn=0.23.2
scipy=1.5.2
networkx=2.5
conda-forge::biopandas=0.2.7
umap-learn=0.4.0
pynndescent=0.4.8
statsmodels
MASS=7.3-47
LSD=4.1-0
colorRamps=2.3
ggplot2=3.3.1
pheatmap=1.0.12
RColorBrewer=1.1-2
igraph=1.1.2
ggcorrplot=0.1.3
cowplot=1.1.1
tidyr=1.1.3