Table 1.

Deep learning approaches and examples in predicting drug treatment efficacy.

Model	Tools*	Input Data	Purpose	Ref
Drug-Target Interactions
DBN	DeepDTIsi	Drug-target pairs information, drug structure, and protein sequence	The probability of interaction for any provided drug-target pair was inferred by DeepDTI based on external, experimental drug-target pairs. Among the top ten predicted drug-target interactions, four had been previously reported, and one was found to have a low binding affinity to the glucocorticoid receptor.	[17]
DNN	DeepCPIii	Drug structure and protein sequence for drug-target pairs	Drug-target interactions were predicted by DeepCPI. The inferred interactions between small molecules and glucagon-like peptide-1 receptor, glucagon receptor, and vasoactive intestinal peptide receptor were experimentally validated.	[18]
DNN	deepDTnetiii	Drug-target pairs information, drug similarity, and target similarity	deepDTnet can identify targets of known drugs using a heterogeneous drug–gene–disease network embedding 15 types of chemical, genomic, phenotypic, and cellular network profiles. A new, direct inhibitor of human retinoic-acid-receptor-related orphan receptor-gamma t, topotecan is predicted by deepDTnet, and then experimentally validated by authors.	[26]
CNN	DeepConv-DTIiv	Raw protein sequences and drug-target pair data	Local residue patterns from proteins in drug-target interactions were captured by DeepConv-DTI via the convolution on various lengths of amino acid subsequences. This model achieved higher accuracy than DBN-based DeepDTI and CNN-based DeepDTA.	[20]
CNN	DEEPScreenv	2-D structure of compounds and protein structure	Drug-target interactions were predicted based on 704 target proteins and the 2-D structure of compounds. JAK proteins were predicted by the model as new targets of drug cladribine and experimentally validated in vitro.	[84]
CNN	AtomNet	3-D structure of target proteins and small molecules	Applying local convolutional filters to extract the target’s structural information, AtomNet successfully predicted new active molecules for targets like wee1 and 1qzy which previously had no known modulators.	[21]
GCN	DeepChemvi	Compound structure	GCN and long short-term memory were employed to optimize small molecule-based drug discovery by predicting the toxicity and bioactivity of candidate drugs using their structural data.	[23]
RNN CNN GCN DTL	DeepAffinityvii	Raw protein sequences and compound sequences (2-D information)	Bidirectional RNN was utilized to capture nonlinear joint dependencies among either protein residues or compound atoms that are sequentially distant. DeepAffinity unified RNN-CNN/RNN-GCN to predict drug-target interactions. The model outperformed conventional models in achieving relative error in the half-maximal inhibitory concentration within five-fold for test cases and 20-fold for protein classes not included in training.	[19]
DNN	Deep-AmPEP30viii	Genomic sequence data and known AMP sequences	Antimicrobial peptides to treat a variety of diseases such as cancer and infections were identified based on known AMP sequences. One peptide (FWELWKFLKSLWSIFPRRRP) the model produced proved to have the same anti-bacteria efficacy as ampicillin.	[24]
Drug Repurposing
DNN	-	Transcriptional response to drug exposure	The therapeutic categories of drugs were exclusively identified from transcriptional profiles. 26,420 drug perturbation samples were analyzed for three cell lines and then assigned to one of twelve therapeutic categories.	[29]
VAE	deepDRix	Drug-disease pairs and drug’s chemical information	Fourteen candidates of the top 20 drug candidates to treat Parkinson’s disease predicted by deepDR were validated by previous studies.	[28]
Drug Response
DNN	RefDNNx	Drug’s structure and gene expression data prior to drug exposure	RefDNN learned representations for a high-dimensional gene expression vectors and a molecular structure vectors of drugs, to predict drug response, then labeled and identified biomarkers contributing to drug resistance. Among the top ten genes identified by RefDNN, six (high expression patterns of MYOF, UBC, NQO1, and LGALS3 and low expression patterns of RACK1 and RPS23) were experimentally proven to be associated with nilotinib resistance.	[30]
DNN VAE	DeepDR	Genomic and transcriptomic profiles before and after drug treatments	The drug response of tumors was predicted from integrated genomic profiles. Specifically, DeepDR improves the prediction of drug response and identification of novel therapeutic options. The model was applied to predict drug response in 9059 tumors from 33 cancer types. The resulting predictions include known therapies, such as EGFR inhibitors in non-small cell lung cancer and tamoxifen in breast cancer, as well as novel drug targets, such as vinorelbine for TTN-mutated tumors.	[33]
MLP RNN CNN	-	HIV genome sequence and drug sensitivity data	HIV-related drug resistance was predicted by three deep learning models. Of the 20 most important features predicted by the models, 18, 9, and 16 known drug resistance mutations positions were identified by using CNN, MLP, and RNN models, respectively.	[32]
VNN	Dcellxi	Genotype data	The phenotypic resulting from individual gene perturbations in eukaryotic cells was transparently simulated. During the simulation, 80% of the importance of growth prediction is captured by 484 subsystems.	[31]
Drug-Drug Interactions
DNN	DeepSynergyxii	Drug chemical data and transcriptional data	DeepSynergy distinguished different cancer cell lines and found specific drug combinations to maximal efficacy on a given cancer cell line through the incorporation of genomic information with compound information.	[34]
GCN	Decagonxiii	Drug-drug interaction, protein-drug interaction, protein-protein interaction, and side effects	Decagon constructed a large two-layer multimodal graph of protein-protein interactions, drug-target interactions, and drug-drug interactions to predict the potential side effects of drug pairs. Decagon accurately predicted polypharmacy side effects, outperforming baselines by up to 69%. It had the best performance in modeling side effects with strong apparent molecular underpinnings; for example, in Mumps, Carbuncle.	[85]
Other
CNN	DeepMACTxiv	3D image data	DeepMACT performed image recognition to track the biodistribution of antibody-based agents. Trained on an MDA-MB-231 cancer cell-based tumor model, DeepMACT has 80% accuracy to detect metastasis.	[86]

^*The superscript numbers refer to the websites in the Resource section.