Machine-Learning-Based Prediction of Cell-Penetrating Peptides and Their Uptake Efficiency with Improved Accuracy

Abstract

Cell-penetrating peptides (CPPs) can enter cells as a variety of biologically active conjugates and have various biomedical applications. To offset the cost and effort of designing novel CPPs in laboratories, computational methods are necessitated to identify candidate CPPs before in vitro experimental studies. We developed a two-layer prediction framework called machine-learning-based prediction of cell-penetrating peptides (MLCPPs). The first-layer predicts whether a given peptide is a CPP or non-CPP, whereas the second-layer predicts the uptake efficiency of the predicted CPPs. To construct a two-layer prediction framework, we employed four different machine-learning methods and five different compositions including amino acid composition (AAC), dipeptide composition, amino acid index, composition-transition-distribution, and physicochemical properties (PCPs). In the first layer, hybrid features (combination of AAC and PCP) and extremely randomized tree outperformed state-of-the-art predictors in CPP prediction with an accuracy of 0.896 when tested on independent data sets, whereas in the second layer, hybrid features obtained through feature selection protocol and random forest produced an accuracy of 0.725 that is better than state-of-the-art predictors. We anticipate that our method MLCPP will become a valuable tool for predicting CPPs and their uptake efficiency and might facilitate hypothesis-driven experimental design. The MLCPP server interface along with the benchmarking and independent data sets are freely accessible at www.thegleelab.org/MLCPP .

A combined drug discovery strategy based on machine learning and molecular docking

“…Twenty‐eight clusters were obtained. Among each cluster, the variable with the smallest ratio of 1‐R (Manavalan et al., ) was selected as a representative variable. Results of principal component analysis were showed in Figure , where values of the 28 variables for the training set and test set as well as the validation set occupied similar chemical space and they were significantly different from each other, indicating that the effectiveness of variable selection.…”

“…First, methodological evaluation was carried out to identify appropriate methods. Specifically, eight machine learning algorithms (Ericksen et al, 2017;Friedman, 2002;Manavalan et al, 2018;Siramshetty, Chen, Devarakonda, & Preissner, 2018) including DT, kNN, SVM, RF, ERT, AdaBoost, GBT, and XGBoost were evaluated and comprehensively compared through an application case of discriminating ACC inhibitors from decoys. Then, machine learning methods and traditional structure-based drug discovery were organically combined to construct a robust strategy for the discovery of ACC inhibitors.…”

Interpretable Deep Learning for De Novo Design of Cell-Penetrating Abiotic Polymers

“…(3-5, 26-28) These previous predictors were mostly trained using physicochemical descriptors, with datasets obtained from non-standardized experiments and containing only canonical residues. (6,(29)(30)(31) Inclusion of chemically diverse unnatural moieties is challenging because such physicochemical descriptors may not be readily available. The ability to encode for unnatural residues, however, would greatly expand the chemical search space, and Peptide sequences are represented as row matrices comprised of residue fingerprints.…”

DHSpred: support-vector-machine-based human DNase I hypersensitive sites prediction using the optimal features selected by random forest

“…Results from the compositional analyses suggested that integrating the amino acid preference information would be helpful for differentiating between DHSs and non-DHSs, and so, we used these as input features for ML methods to improve classification. The major advantage of ML methods is their ability to consider multiple features simultaneously, often capturing hidden relationships [16][17][18][19][20][21][22][23].…”

“…In the second step of the previous section, we used three different ML-based methods instead of SVM, including, RF, ET, and k-NN. A detailed description of the development of prediction models using these methods was provided in our recent studies [21,23]. For each ML-based method, we generated 33 prediction models using different sets of features, including individual composition, hybrid models, and features based on FIS cut-off.…”