Persistent Laplacian projected Omicron BA.4 and BA.5 to become new dominating variants

Guo-Wei Wei, Jiahui Chen, Yuchi Qiu, Rui Wang, Guo-Wei Wei 0001Computers in Biology and Medicine 2022 · DOI

TL;DR

This paper introduces "persistent Laplacian," a novel topological data analysis technique that captures both the shape and topology of protein-protein interactions. This method is used to predict the infectivity of SARS-CoV-2 variants, accurately forecasting the rise of Omicron BA.4 and BA.5 as dominant strains due to their increased transmissibility.

What Problem Does This Paper Address?

The emergence of new SARS-CoV-2 variants poses a continuous threat, necessitating accurate prediction of their transmissibility and potential to evade immunity. While previous topological methods have shown success, they struggle to account for subtle shape variations in protein-protein binding, which are crucial for viral evolution. This paper addresses this limitation by developing a method that can capture both topological and geometric features of these interactions, leading to more accurate predictions of variant infectivity.

Methods

The core innovation is the persistent Laplacian, an extension of persistent homology that incorporates spectral graph theory. Unlike traditional persistent homology, which focuses solely on topology, the persistent Laplacian analyzes the "harmonic and non-harmonic spectra" of q-combinatorial Laplacian matrices derived from simplicial complexes representing protein structures. This allows it to capture both the topological structure and the geometric shape of data.

The authors then integrate this persistent Laplacian representation with machine learning models, specifically gradient boosting trees (GBT) and deep neural networks (Net). They also explore various combinations of these models and features, including pure persistent Laplacian features and those augmented with other topological and geometric information.

Datasets

The study utilizes several datasets for training and validation:

  • AB-Bind dataset: Contains 1,101 mutational data entries for experimentally determined binding free energy (BFE) changes in antibody-antigen complexes.
  • SKEMPI dataset: A larger dataset with 3,047 entries of BFE changes from protein-protein heterodimeric complexes.
  • SKEMPI 2.0 dataset: An updated version of SKEMPI with more mutations and complexes.
  • SARS-CoV-2 related datasets: These include deep mutational enrichment ratio data for ACE2 binding to the S protein RBD, as well as data on RBD binding to ACE2 and CTC-445.2. These were used to calibrate models for predicting real-world experimental results.

Key Insights and Ideas

  • Beyond Topology: The paper highlights that pure topological analysis (like standard persistent homology) can miss crucial geometric nuances in protein interactions, which are vital for understanding viral evolution.
  • Shape Matters: The persistent Laplacian effectively bridges this gap by incorporating spectral information, allowing it to capture both the "shape" and "topology" of protein binding interfaces.
  • Predictive Power: This enhanced representation, when combined with machine learning, leads to significantly improved predictions of mutation-induced BFE changes and, consequently, variant infectivity.
  • Omicron BA.4/BA.5 Dominance: The model accurately predicts that Omicron subvariants BA.4 and BA.5 are approximately 36% more infectious than BA.2 and are poised to become the next dominant strains due to natural selection.

Key Results

  • State-of-the-Art Performance: The proposed persistent Laplacian-based models (particularly TopLapGBT and TopLapNetGBT) achieve state-of-the-art performance on benchmark datasets like AB-Bind, SKEMPI, and SKEMPI 2.0, outperforming existing methods in Pearson correlation coefficients (Rp) and RMSE. For instance, on the AB-Bind S645 set, their best model achieved an Rp of 0.89, significantly higher than other methods.
  • Accurate Infectivity Forecast: The models correctly identified Omicron BA.2.11, BA.2.12.1, BA.3, BA.4, and BA.5 as more contagious than BA.2.
  • BA.4/BA.5 Prediction: The study specifically predicts BA.4 and BA.5 to be about 1.36 times as infectious as BA.2, making them strong candidates for new dominant variants.
  • Correlation with Experimental Data: For SARS-CoV-2 RBD binding to ACE2, the predicted BFE changes showed a strong correlation (Pearson correlation of 0.69) with experimental deep mutational enrichment data.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, the complexity of generating persistent Laplacian features and training deep learning models suggests potential computational demands.
  • Future Work: The authors suggest that their approach can be extended to analyze other biological systems and protein-protein interactions. The accurate forecasting of emerging variants also implies potential for guiding future vaccine and therapeutic development.

Summary updated

BibTeX
@article{wei2022persistent,
  author = {Guo-Wei Wei and Jiahui Chen and Yuchi Qiu and Rui Wang and Guo-Wei Wei},
  title = {{Persistent Laplacian projected Omicron BA.4 and BA.5 to become new dominating variants}},
  journal = {Computers in Biology and Medicine},
  year = {2022},
  doi = {10.1016/j.compbiomed.2022.106262},
}
Figure