CaDenseNet: a novel deep learning approach using capsule network with attention for the identification of HIV-1 integration site

The human immunodeficiency virus (HIV) is one of the many variants of the retrovirus and it affects the human immune system. HIV integration site (IS) is a key determinant in the entire infection process and in finding a cure for the fatal disease owing to the fact that it is critical in determining the entire latent viral reservoir formation process. The ISs are specific to the retrovirus and in the case of HIV, the integration may take place at any stage of the cell cycle. The extent and process of the rebound of the virus when the antiretroviral therapy is disturbed is also determined by the IS. IS databases are mainly populated by lab-based observations/experiments which are performed for their isolation and analysis of some 'other' related characteristics in an attempt to influence the integration process. This work proposes a unique Deep Learning (DL) approach, using the recently developed Capsule Network along with attention, for the prediction of IS. Retrovirus Integration Database has been used as it is one amongst the few publicly available gold standard databases of retrovirus IS in the human genome and has the cells transfected under lab conditions. This work achieves a performance much better than the existing state-of-the-art method (both with and without additional genomic markers), presented here in terms of the metrics AUC-ROC, AUC-PR and F-score. It looks beyond lab-based studies for HIV IS analysis and suggests that some advanced DL- based automated architecture can help in progressing towards a cure to the disease.

Automated summary

The HIV virus affects the human immune system and its integration site is crucial in the infection process and finding a cure. This study proposes a deep learning approach to predict the integration site, which outperforms existing methods and can contribute to finding a cure for HIV.