Predicting future high-risk SARS- CoV-2 variants with deep learning
Item Type Poster
Authors Chen, NingNing
Download date 2023-11-02 02:13:52
Link to Item http://hdl.handle.net/10754/678112
● We developed an accurate multi-task deep learning model for predicting ACE2/antibody binding specificity and synthesized high risk variants of Sars-CoV-2 that may cause concens in the future.
● One limitation is that we only focus on the RBD sequences, while many mutations occur outside the region.
● This work might be used to study co-evolution of virus- antibodies in patients in the future, it can also be extend to study other virus evolution
* WenKai Han and NingNing Chen contribute equally to this work 1.Computational Bioscience Research Center, King Abdullah University of Science and Technology (KAUST) 2.Computer, Electrical and Mathematical Sciences and Engineering Division, King Abdullah University of Science and Technology (KAUST)
3. Biological and Environmental Science & Engineering Divison, King Abdullah University of Science and Technology (KAUST)
WenKai Han
1,2,*, NingNing Chen
1,3*, Xin Gao
1,2Predicting future high-risk SARS-CoV-2 variants with deep learning
Introduction
● The global pandemic caused by SARS-CoV-2 has been continuously invades our daily live since 2019.
● The SARS-CoV-2 virus evolves fast throughout the pandemic and cause repeated outbreaks.
● To step ahead of the virus, we need to understand the effects of mutations and be prepared for the high-risk variants.
● Here, we develop a novel in silico approach combing the antibody/ACE2 structure modeling and a multi-task deep neural network, to successfully model the fitness landscape of the spike RBD, by accurately predicting the effects of mutations on ACE2 binding and antibody escape.
● Develop and train a multi-task deep neutral network model predicting the variant RBD binding specificity towards the ACE2 and antibodies from four classes. (upper)
● Training dataset: Deep mutational scanning of mutated SARS-CoV-2 RBD on human ACE2 binding affinity and escape ability assessment of SARS-CoV-2 RBD to 8 human antibodies (left)
● Using genetic algorithm to search for the high risk variants (right upper) Method
Overview of predicting potential SARS-CoV-2 variants by deep learning
Model Performance
● Our model outperforms the other SOTA methods in predicting the effects of mutations in all the nine tasks. (our model is colored with green)
● Based on the embedding plot (bottom), our model captures binding specificities for all nine targets.
(Only show 4 here)
Model embeddings colored by antibody classes Model metrics comparison
Evolution velocity Analysis
MDS plot of sequence space Escape map on RBD
Seqlogo of top sites
Binding scores to LY-CoV555 Heatmap of docking binding scores
Evolution velocity Analysis
Spearman correlation in different time
● The existing variants between Dec 2019 to Mar 2022 were collected from GISAID website and the sampling time are indicated by color in the left figure.
● The evolution velocity trajectory are based on the average prediction scores of 9 tasks provided by our model (vector fields of all figures ).
● Our model's prediction is quite consistent with the real virus evolution trend.
Result - Evolution velocity analysis on existing variants validate the prediction accuracy of our model
Result - Synthetic variants expand the mutation landscape and capture key mutations for ACE2 binding and antibodies escape
● Frames on RBD means contact sites of RBD to ACE2 and four classes antibodies.
● The intensity of red and height in seqlogo can be regared as model preference
Result - In-silico Docking simulation shows that all the top20 synthetic variants have increased immune escape to all 4 classes antibodies
● These figures reveals the binding scores of top20 synthetic variants to all four classes antibodies are higher than wt, with some even higher than omicron.
● From the stream, we can see how the Sars-CoV2 might evolve in the future.
Result - Evolution velocity analysis on both synthetic and existing variants
Result - Spearman correlation of the model prediction score and the sampling time indicate trend of virus evolution
● The Sars-CoV2 tends to evolve to escape antibodies rather than increase transmissibility under the pressure of vaccination
Conclusion & Discussion