| Enabling inexpensive, rapid and effective detection of drug-target interactions and repurposing [8], [51], [71], [72]. |
Still requires inputs from relevant experts to ensure its reliability in predicting drug candidates [69], [70], [71]. |
| The potential to aid in designing and discovery of therapies by processing vast datasets and complex pattern recognition capacity involving genomics, proteomics, microarray data and clinical trial data [66], [68], [69], [71], [72]. |
Due to potential biases in different encoding models and dataset utilized, deep learning models are likely to provide different drug predictions [8], [4], [36]. |
| The ability to identify drugs that target several COVID-19 targets such as variational autoencoders (VAE) which allows the generation of specific molecules with greater diversity, overcoming the limitation of ligand-based designs [58], [69], [71]. |
Lacks capacity to predict or repurpose synergistic drug combinations as potential treatment for COVID-19, such as those seen in certain clinical studies [25]. |
| Combined with transfer learning technique, it could predict, repurpose, design and discover new COVID-19 therapies using predictive models [69], [71]. |
Despite strong applicability, it is likely a black box which could provide predictions that are not well understood. Hence, there is a need for more insight into interpretability. |
Natural language processing (NLP) can transform unstructured texts into structured data, which can be analyzed appropriately to gain new insights. Various text mining-based tools have been developed:
-
•
PISTON, a tool that can predict drug side effects and drug indications, using NLP and topic modeling.
-
•
STITCH, text mining-driven database, which contains information on interactions between proteins and chemicals/small molecules [69], [70], [71].
|
Most approaches adopted for repurposing COVID-19 drugs used SMILES strings and protein sequences as inputs [4], [8], [36], [51]. The usage of molecular graphs as molecular topology with nodes and edges representing atoms and chemical bonds are common in most research studies for drug discovery and prediction of drug side effects [71], [73]. Models using SMILES representation of molecular structures as an input for drug-repurposing may have a low rate of valid molecules due to lack of topological information [47]. |
| Ability to predict side effects of new drugs using biological, chemical, and semantic information extracted from structured and unstructured big data [71], [73]. |
No deep learning approaches have been deployed to design therapies against mutated COVID-19 strains. |
