Page 10 - GPD-2-1

Gene & Protein in Disease AI-based drug repositioning

Table 2. Feature‑based approach better performance than support vector machine model.
Therefore, this deep neural network model was used for
Methods Features References drug repurposing. In addition, they proposed that a deep
KNN No need to estimate parameters and train [18] neural network confusion matrix can be used for drug
drug target data repositioning [26] . Segler et al.’s [27] method based on deep
SVM Requires a relatively small number of drug [20] learning combined with Monte Carlo algorithm is simple
target samples and efficient and has been affirmed by professionals.
DT Can handle both continuous and discrete [22] Hughes et al. developed the first model capable of
drug target data fast screening of compounds using deep learning
LR The weights of the target features can see [24] models [28] . Turk et al. extracted matched molecules
how different features affect the final results
KNN: K-nearest neighbor, SVM: Support vector machine, from the ChEMBL database as a dataset for deep
. Zhang et al. proposed an extraction
learning models
[29]
DT: Decision tree, LR: Logistic regression.
strategy based on multisource features to construct a
In 2006, Guengerich used machine learning algorithms protein-ligand interaction prediction model using an
to reveal the role of P450 enzymes arising in drug integrated learning approach, which outperformed
metabolism and toxicity . In 2011, Dr. Feixiong Cheng the single classifier prediction model in terms of
[21]
proposed a method to predict P450 enzymes using sensitivity and Youden index and could effectively
[22]
traditional classifiers such as KNN, DT, and SVM . solve the data imbalance problem [30] . Chen et al.
A large number of new algorithms to predict human proposed a multisimilarity fusion-based drug relocation
cytochrome P450 enzymes were published immediately recommendation algorithm to address the shortcomings
afterward. Napolitano et al. applied non-linear SVMs to of traditional drug relocation recommendation
the classification of drug efficacy . Gottlieb et al. used algorithms. First, disease similarity was calculated based
[23]
logistic regression algorithm for drug repositioning . on drug-disease data sources. Then, three similarities
[24]
Gönen used Bayesian algorithm in machine learning for were calculated based on drug-chemical structure, drug-
drug-target protein prediction to find new drug-target target protein, and drug side effect data sources and were
protein association relationships. First, drug and side effect fused into drug similarity. Finally, the predicted values of
information, drug chemical structure information, and drug-disease correspondence were calculated using two
disease and gene-related information were integrated, and similarities and fused into the final predicted values by
then, the training data were obtained by feature selection the prediction fusion method [31] . Zhang et al. obtained
and feature extraction. Then, suitable machine learning knowledge associations from PubMed, DrugBank, CTD,
algorithms were selected to train them, and finally, and other databases, constructed semantic knowledge
the trained algorithm models are used to obtain drug graphs by knowledge fusion, attribute definition, and
[25]
repositioning results . used drug repositioning as empirical evidence to
reason about new uses of drugs in tumor therapy by
Nowadays, the technology of drug repositioning is two methods: Path search and link prediction [32] . Zeng
influenced by deep learning, of which feature learning is et al. investigated a deep learning drug repositioning
the main embodiment of deep learning technology. Deep (deepDR) based on a network deep learning approach
learning builds computational models by simulating using multimodal deep self-encoder and variational self-
the human brain, which can not only extract features encoder models to discover drug-disease associations,
automatically but also obtain effective feature information which is shown in the steps in Figure 3. They combined
at different levels. Based on these advantages, deep learning many drug association data phases into one dataset
has also been applied in drug repositioning. Deep learning (drug-disease association, drug-target association, drug-
technique is a concept closely related to artificial neural drug association, and drug side effects) and then used
network. this dataset to train a multimodal deep autoencoder
To predict the pharmacological characteristics of and then define advanced drug features [33] . Next, to
a drug, Aliper et al. fully used connected deep neural identify new indications based on features that can be
networks to make predictions. The drug characteristics identified, they used a variational autoencoder to encode
were also used to calculate the therapeutic potential as and decode combinations of advanced drug features
well as new drug indications. They constructed a deep and disease-drug associations in clinical reports. These
neural network model using gene expression signature studies have been tested on datasets with common
data and pathway data. This model has a high accuracy disease-drug associations and have shown better results
in prediction for drug indication classification and has than previous machine learning models.

Volume 2 Issue 1 (2023) 4 https://doi.org/10.36922/gpd.v1i3.201