Table 2.
Classification of articles related to drug discovery and DL
| References | Publishing year | Method | Advantages | Drawbacks |
|---|---|---|---|---|
| Drug–target interactions (DTIs) | ||||
| Yang et al. (2019) | 2019 | A graph convolutional model that consistently equals or outperforms models with constant molecular descriptors and earlier graph neural designs on both open-source and closed-source data sets | The model performs equally well or better on 12 of the 19 open data sets | The proposed model is underperforming other models when either (1) the 3D information from the other models, (2) the data set is especially small, or (3) the classes are particularly imbalanced |
| Hirohara et al. (2018) | 2020 | A brand-new CNN for assessing data about chemical compounds. The CNN processes picture data similarly to traditional CNNs using a SMILES-based feature matrix | The standard fingerprint method used for the virtual screening of chemical compounds performed worse than the CNN based on SMILES string. Additionally, the use of motif detection with learnt filters allowed the detection of undiscovered functional group substructures in addition to significant recognised substructures like protein-binding sites | The model performed better than DNN using simply ECFP in the TOX 21 Challenge 2014 experiment, but somewhat worse than models utilising both ECFP and DeepTox features |
| Shin et al. (2019) | 2019 | A novel representation of a molecule employing the self-attention method that has been pre-trained using massive data on chemicals that is available to the public | Regarding four evaluation metrics, the model performs better than all other techniques now in use. Furthermore, the strategy is successful in including all the existing EGFR medications in the top-30 promising prospects, as demonstrated by the example study of discovering drug candidates targeting a cancer protein (EGFR) | The protein was ignored by the model. One explanation is because, on average, a protein sequence is 10 times longer than a molecule sequence, which requires a significant amount of computation time. Another justification is the requirement of a protein dataset with adequate data to pre-train the model |
| Lee et al. (2019) | 2019 | A new DTI prediction model by identifying regional residue trends from the complete sequences of the target protein via CNN | Other protein descriptors, such as CTD and SW scores, do not perform as well as the local properties of protein sequences that have been found. Compared to techniques requiring 3D structures, it can be more broadly utilized to predict DTIs | To attain enhanced performance, more detailed chemical characteristics are used for DTI forecasting. Chemical elaboration can be replaced by the consideration of 3D structure information |
| Wang et al. (2020b) | 2020 | A novel method MDADTI to predict DTIs based on MDA | MDA lowered the size of the characteristic of medications and objectives, which accelerate the training of MDADTI | Did not consider predicting the binding affinity scores for drug–target pairs |
| Wan et al. (2019) | 2019 | DeepCPI is a new and scalable approach that predicts innovative CPIs by combining data-driven representation learning and DL (DTIs) | To effectively utilize the enormous amount of data on the interactions between compounds and proteins accessible from expansive databases, including PubChem and ChEMBL, the synthesis of the efficient feature embedding methodologies with the potent DL model is especially helpful | Better prediction results may be reached by combining other accessible data, like as gene expression and protein structures, into the suggested DL model |
| Vazquez et al. (2020) | 2020 | Model for defining hybrid LB (ligand-based) + SB (structure-based) computational schemes in VS (Virtual screening) studies | VS has been a powerful alternative to high-throughput screening assays. The performance of VS campaigns is significantly impacted by the inherent weaknesses of both LB and SB procedures while being the most popular way for exposing innovative hit compounds in the early phases of drug discovery | Although the progress is encouraging, it can be assumed that two primary elements will determine if these integrated tactics are adopted. Initially, a thorough benchmarking of the various combination tactics Secondly, convenient access to screening of focused chemical library searches should be made possible by the capacity to integrate the combined LB and SB techniques into automated modelling platforms |
| Xie et al. (2018) | 2018 | A framework that offers better prospects for inferencing and for DTI prediction based on L1000 database transcriptome data from pharmacological perturbation and gene knockdown experiments | The outcomes showed that our system can find DTIs that are more trustworthy than those discovered by previous techniques | investigating a possibility to discover an internal characteristic for assessing fresh DTI potential |
| Drug sensitivity and response | ||||
| Dincer et al. (2018) | 2018 | A low-dimensional representation (LDR) is encoded using the DeepProfile framework, which learns a variational autoencoder (VAE) network using tens of thousands of publicly available gene expression samples. This network is then used to predict complicated disease symptoms | DeepProfile successfully disentangles data inconsistencies and discovers a useful LDR that precisely foretells complicated phenotypes of many malignancies | The work needs some future directions which include: (1) training DeepProfile using samples from various cancers and (2) expanding the application of DeepProfile by utilizing multi-omics data to produce more insightful embeddings for cancer |
| Rampášek et al. (2019) | 2019 | ‘Dr.VAE’, the first unified machine learning method for the semi-supervised learning for drug response prediction that successfully exploits the prior knowledge of drug-induced transcriptomic perturbations | First, the flexibility of This paradigm made it possible to incorporate the effects of transcriptional disruption into the framework for predicting medication reaction in a special way Second, neural networks with the ability to simulate complex non-linear interactions are employed to parametrize all conditional distributions that make up their model, in addition to variational posteriors employed by Dr. VAE to make a rough inference | The study could be limited in numerous ways. As it has repeatedly been demonstrated to offer the greatest predictive potential in several prior research on treatment response, they first just took the gene expression modality into consideration. Second, they modeled CMap-L1000v1 perturbations following a 6-h treatment period. Since many feedback regulatory processes take 6 h to appear, it may be argued that these studies alone do not give a complete picture of the transcriptome response |
| Ahmed et al. (2020) | 2020 | Two neural network models based on graphs for predicting drug sensitivity as well as a network-based feature selection approach | To predict drug sensitivity, using a network to pick features. first identifies additional illustrative characteristics based on the network of gene co-expression. Second, Random Forest performs better than all the other established prediction techniques., Third, compared to DNN, the graph-based neural network models exhibit higher drug response prediction capability., Fourth, the performance of the prediction is reliant on the drug and could be related to the drug's mode of action (MOA) | In this study, there was little improvement for DNNs using graphs. Multi-omics data can be used for analysis to further improve drug prediction accuracy. In addition to the extensive biological features, data acquisition using multi-omics the genomic, epigenomic, and transcriptomic traits of each cell line in the cohort and offer more precise molecular fingerprints to forecast medication reaction than single-omics data alone |
| Ren et al. (2022) | 2022 | DeepGRMF, which demonstrated enhanced capability for predicting drug sensitivities | model can predict new responses to drugs or medications not previously seen, which It makes it possible to use already FDA approved medications in new ways to both cure cancer and maybe find new molecules that fight cancer | the DeepGRMF model only uses data from gene expression profiling, incorporating genetic changes further and epigenetic data will probably help the model perform even better |
| Drug–drug interactions (DDIs) side effect | ||||
| Jae et al. (2018) | 2018 | DeepDDI that accurately forecasts DDI types for medication combinations as well as drug-food component pairings using as inputs names and chemical structures | According to the seven common performance criteria, DeepDDI's feature vector of pharmaceuticals, which combines an improved DNN and SSP, demonstrated great accuracy (84.8–93.2%) | Only two medications are included in the best DDI database provided by DrugBank. The DNN can be updated by training DeepDDI for numerous medications and food ingredients once DDI data for multiple drugs and/or food ingredients becoming accessible |
| Xiang et al. (2020) | 2020 | An overview of various graph embedding approaches is provided, along with an assessment of how well they perform on link prediction and node categorization, two significant biological tasks | Through comprehensive testing, they discover that current graph embedding techniques can compete favorably with state-of-the-art techniques or even outperform them in a variety of biological prediction tasks | The idea of pre-training followed by fine-tuning becomes increasingly intriguing as there are more entities that may have pre-trained embeddings as biomedical data volume increases. The relationships between the freshly created network propagation (diffusion) methods and the graph embedding techniques may also have an impact on future research |
| Xifan et al. (2020) | 2020 | A multimodal deep learning system called DDIMDL uses deep learning and a variety of pharmacological characteristics to forecast DDI events | The experimental findings demonstrate that DDIMDL performs better than the approaches that were examined in terms of efficiency and accuracy | to enhance the forecast of DDI events. First, there needs to be consideration of unique strategies for dealing with the unbalanced dataset due to the extreme imbalance in the amount of DDIs for various occurrences. Second, certain events have insufficient numbers of interactions, making it easier for the DL approach to underfit, hence techniques to increase the size of the event dataset can be investigated |
| Karim et al. (2019) | 2019 | A new ML approach for forecasting DDIs using a variety of sources | The incorporation of drug data from many sources using knowledge graphs. This candidate will get comprehensive background information on medications, illnesses, processes, proteins, enzymes, chemical structures, etc | The approach's inability to offer reasons for the projected DDIs is one of its limitations |
| Bongini et al. (2022) | 2022 | Graph Neural Networks (GNNs) are exploited to predict DSEs | A GNN-based predictor aids in foreseeing the incidence of adverse effects. Additionally, its use with novel candidate medications would assist drug discovery studies save time and money while avoiding health problems for the volunteers taking part in clinical trials | An exciting future path is the creation of a prediction based on a GNN that It has the capacity to analyze molecular structural formulae, which are displayed as graphs. Features could be added to these molecular graphs to improve them by way of genes and links between genes and drugs |
| Iorio et al. (2010) | 2010 | A broad method to identify previously unknown uses for well-known medications as well as to forecast the molecular actions and MOA of new compounds | Using information buried in the gene expression profiles after drug treatment to identify pharmacological MoA similarities | The approach's primary drawback is the network's small number of chemicals. If a chemical is not like any of the medications in the network, no inferences about its MoA or its biological effects can be made because the approach is predicated on comparing how similar two drugs are |
| Wang et al. (2019) | 2019 | An approach using networks to find clinically effective drug combinations for disorders | Provides a network-level analysis of the comparative efficacy and harmful interactions of treatment combinations | The applicability of their findings to different disorders must be investigated in further research |
| Heba et al. (2021) | 2021 | The ML-based on similarity for drug interaction prediction (SMDIP) framework, which incorporates popular ML models utilizing internal similarity models based on products to choose the characteristics and benefit from more sparse feature space to enhance DDI prediction efficiency | creating a generalizable machine learning framework based on similarities to reveal innovative DDIs having excellent forecasting abilities; and (2) Showcasing superior predictor elements in comparison to the body of existing research | SMDIP exhibits some restrictions. SMDIP is first and foremost a binary classifier framework with DDI results. It cannot, however, predict how serious a DDI will be. Second, SMDIP cannot make recommendations for the best course of action to be done for the management of DDI since it cannot forecast the unfavorable effects brought on by the interaction |