scGImpute: A hybrid BiLayer multi-head graph attention-based imputation framework for zero dropout in single-cell sequencing datasets.

Single-cell sequencing (SCS) is a robust high-throughput sequencing technology used to measure RNA and DNA molecules, revealing hidden insights into the multi-ome profiles of humans, plants, animals, and microorganisms. Recently, advances in SCS technology have been applied on a large scale in biomedical fields for the improvement of disease diagnosis and treatment at the molecular level. However, a number of technical challenges cause significant noise in SCS data even with advancements in experimental procedures. This noise leads to corruption and hinders the better outcome of SCS for various aspects. Among these, one of the serious technical issues is the high frequency of zero read counts or dropout events in the SCS datasets. To address this challenge, this study proposes a novel hybrid two-layer multi-head graph attention-based neural network framework (scGImpute) for preserving biological zeros while recovering technological zeros in omics SCS datasets. The proposed method was made generic for three omics scRNA-seq, scATAC-seq, and CITE-seq datasets. The scGImpute method achieved the lowest Root Mean Square Error (RMSE) of 0.0039 and Mean Absolute Error (MAE) of 0.0032 for the raw vs imputed scRNA-seq dataset. Similarly, it obtained the best performance for the ground truth vs imputed scRNA-seq dataset, with an RMSE of 0.3301 and MAE of 0.2924. The evaluation metrices results indicated that the proposed scGImpute method outdid existing imputation techniques and the other proposed methods. This proposed method can reduce the technical noise from the SCS datasets for better downstream analysis to find the biological realm.