Deep-learning based repurposing of FDA-approved drugs against Candida albicans dihydrofolate reductase and molecular dynamics study

Tanuja Joshia#, Hemlata Pundirb# and Subhash Chandraa

ABSTRACT

Candida albicans causes the fatal fungal bloodstream infection in humans called Candidiasis. Most of the Candida species are resistant to the antifungals used to treat them. Drug-resistant C. albicans poses very serious public health issues. To overcome this, the development of effective drugs with novel mechanism(s) of action is requisite. Drug repurposing is considered a viable alternative approach to overcome the above issue. In the present study, we have attempted to identify drugs that could target the essential enzyme, dihydrofolate reductase of C. albicans (CaDHFR) to find out potent and selective antifungal antifolates. FDA-approved-drug-library from the Selleck database containing 1930 drugs was screened against CaDHFR using deep-learning, molecular docking, X-score and similarity search methods. The screened compounds showing better binding with CaDHFR were subjected to molecular dynamics simulation (MDS). The results of post-MDS analysis like RMSD, RMSF, RG, SASA, the number of hydrogen bonds and PCA suggest that Paritaprevir, Lumacaftor and Rifampin can make good interaction with CaDHFR. Furthermore, analysis of binding free energy corroborated the stability of interactions as they had binding energy of 114.91 kJ mol1, 79.22 kJ mol1 and 78.52 kJ mol1 for Paritaprevir, Lumacaftor and Rifampin respectively as compared to the reference (63.10 kJ mol1). From the results, we conclude that these drugs have great potential to inhibit CaDHFR and would add to the drug discovery against candidiasis, and hence these drugs for repurposing should be explored further.

KEYWORDS
Candida albicans; deeplearning; FDA-approveddrugs; molecular docking & dynamics simulation; dihydrofolate reductase (CaDHFR)

Introduction

Candida albicans is the major systemic fungal pathogen which are often life-threatening and associate with high mor which is responsible for the fatal bloodstream infection in tality (mortality rate: 20–40%) in immunocompromised hosts humans called candidiasis, a disease that continues to be a (Lai et al., 2008; Park et al., 2009). In immunocompromised people, for example, patients with AIDS, or patients undergoing organ transplants or anticancer chemotherapy, Candida causes irritation and infections that can range from thrush in immunocompetent colonized hosts, to hazardous systemic diseases (Brown et al., 2012; Brzankalski et al., 2008). The expense of treating bloodstream Candida infections is $2–4 billion per year in the US (Wilson et al., 2002). In the US alone, yearly instances of systemic candidiasis are around 70,000 per year with a death rate of about 30–40%, even after antifungal therapy treatment (Perlroth et al., 2007). This situation is particularly increasing in cancer patients. The instances of Candida infection in all cancer patients are about 40–88% (DiNubile et al., 2005; Fujitani et al., 2006) and the mortality rate reaches very high (75%) among the Filipino and Pacific Islanders (Fujitani et al., 2006).
Clinically, several antifungals are used to treat life-threatening IFIs including amphotericin B (AmB), azoles (e.g. fluconazole, itraconazole and voriconazole), echinocandins (e.g. caspofungin, micafungin and anidulafungin) and 5-fluorocytosine (normally utilized as adjunctive therapy) (Odds, 2005; Odds et al., 2003). AmB makes communication with ergosterol on the fungal cell membrane (Volmer et al., 2010). Although AmB is considered for the treatment of some severe infections (Ostrosky-Zeichner et al., 2003), it also has serious nephrotoxicity and numerous other side effects (Fanos & Cataldi, 2000). Azoles are inhibitors of CYP51 in the fungal cell membrane and are generally used in antifungal therapy (Odds, 2005). However, most antifungals have moderate to severe side effects as well as their antifungal potency has been greatly decreased due to severe drug resistance (Pfaller, 2012; Sanglard & Odds, 2002). Most of the Candida isolates (about 30%) are resistant to all antifungals used to treat them (Fisher et al., 2011) particularly; the resistance of C. albicans toward azoles is getting progressively common (Fisher et al., 2018; Liu et al., 2015). Drug resistance over the last 20 years has posed a vital public health issue and restricted the use of these antifungal drugs. Accordingly, the development of new antifungals is highly required. The drug resistance arises, due to the mutation in the target enzyme of the fungal cell. To address this problem, we utilized a new emerging target dihydrofolate reductase of C. albicans (CaDHFR).
Dihydrofolate reductase (DHFR) is a crucial enzyme in the folic acid metabolic pathway which plays a major role in the RNA, DNA, as well as protein biosynthesis (Askari & Krajinovic, 2010). DHFR is a conventional anti-cancer target, however, in recent years; targeting DHFR has also proved to be a successful technique for antimicrobial drug development. There are major variations in the binding site of human and Candida species, which form the basis for the advancement of novel and specialized inhibitors. While some fungal DHFR inhibitors have been reported (Otzen et al., 2004), there is still much-renewed interest in the development of possible DHFR inhibitors as the next-generation antifungal medications.
Drug repurposing has been demonstrated to be a promising way in the drug discovery process, not just in distinguishing new uses for old medications but also in offering new lead compounds (Pushpakom et al., 2019). Several ‘nonantifungal’ drugs with antifungal action were identified through drug-repurposing of marketed drugs (Pappas et al., 2016) for example, haloperidol, a butylbenzoic antipsychotic drug, was reported to show inhibitory activity against C. albicans (Stylianou et al., 2014). Recently, it was reported that bromperidol derivatives had synergistic effects with triazole antifungals to treat Candidasis (Holbrook et al., 2017).
Deep-learning is a machine-learning approach that employs advanced biologically brain-inspired algorithms called artificial neural networks. It consists of various artificial neurons and hidden layers to imitate the workings of the human brain (Rusk, 2016). Recently, a variety of modern machine learning methods, in particular, deep neural networks were used for drug discovery and development. Previous studies have suggested that deep learning techniques have demonstrated superior results to machine learning algorithms. There are clear distinctions between deep learning and conventional machine learning methods, as traditional machine learning methods use sparse representations to explain input data, and the associated learning-task features are further derived from representations that require substantial domain knowledge and time involved and may lose some essential information in the process; whereas the deep-learning representation method allows distributed dataset representations, it automatically extracts features that can derive conceptual higher-level features and eventually produce more precise prediction outcomes (LeCun et al., 2015).
A deep-learning algorithm is more efficient, quicker and cheaper for drug development since the deep layers of the model can learn more about the data and make accurate predictions (Zhang et al., 2020). We, therefore, believe that the combined implementation of such machine learning techniques as a drug development pipeline has applications in therapeutic drug targeting. Other than the field of drug discovery, deep learning approaches are also being utilized in the field of the scoring function. Jimenez et al. constructed a general-purpose scoring function KDEEP via 3D-convolutional neural networks (Jimenez et al., 2018). Deep-learning has also proven effective in numerous fields such as natural language processing (NLP), clinical investigation, image-processing, computer vision, computer games, self-driving vehicles and so on (Esteva et al., 2019; He et al., 2018). Healthcare, clinical issues and drug development also benefit tremendously from deep learning methods due to the exponential rise in biomedical data (Esteva et al., 2019; He et al., 2019). Deep-learning models for medical diagnosis have achieved physician-level and presented a positive application outlook (Esteva et al., 2017; Lin et al., 2018; Long et al., 2017).
However, the use of deep-learning approaches in drug discovery and chemical biology is restricted due to the scarcity of data, data analysis and lack of user-friendly deep-learning software. Recently, Liu et al. developed a user-friendly web server, DeepScreening (http://deepscreening.xielab.net) with the integration of a deep learning algorithm, which utilizes either the public or user-provided dataset to help biologists or chemists to perform virtual screening either the chemical probes or drugs for a specific target of interest (Liu et al., 2019). Using DeepScreening web server, we could create a deep-learning model and generate the target-focused de novo libraries. Constructed classification or regression models may then be used for virtual screening against de novo libraries or different chemical libraries offered on the server.
Therefore, considering the urgent need for C. albicans inhibitors and enough previous evidence of using DHFR inhibitors, the current work is attempted by adopting a deep learning algorithm. To continue the efforts of computational drug repurposing, we used FDA-approved-drug-library containing 1930 drug molecules for screening against CaDHFR to identify some new drugs against C. albicans and candidiasis. This study presents the outcomes obtained by deeplearning technique, virtual screening and molecular dynamic simulation and summarizes the existing FDA-approved drugs which could be utilized against C. albicans. This research will help to get new antifungals against C. albicans and help humans against candidasis. The workflow of the overall methodology is shown in Figure 1.

Material and methods

Receptor preparation

The X-ray crystal structure of C. albicans enzyme dihydrofolate reductase (CaDHFR) (PDB ID: 4HOF and resolution: 1.76 Å) attached with its inhibitor (18H) was obtained from RCSB Protein Data Bank (PDB) (https://www.rcsb.org) (GDayanandan et al., 2014) (Figure 2(A)). The 18H (5-[3-(2methoxy-4-phenylphenyl)but-1-yn-1-yl]-6-methylpyrimidine2,4-diamine) was used as reference molecule in this study. For receptor preparation, all the water molecules and existing ligand (except NDP) were removed from the protein molecule with the help of the PyMOL software. After that, hydrogen atoms were added to the receptor molecule by using Autodock Tool 1.4 (ADT) (Goodsell et al., 1996).

Predictive modeling by deep learning

Predictive modeling was performed by using the deep-learning online server (http://deepscreening.xielab.net) (Liu et al., 2019) which uses a deep-learning algorithm to develop the predictive model. The CHEMBL dataset (CHEMBL2329) of CaDHFR inhibitor was used as the training set. This dataset contained IC50 (inhibition constant) values of compounds which are reported as C. albicans dihydrofolate reductase (CaDHFR) inhibitors by Kuyper et al. (1996). Kuyper et al. (1996) reported a series of 7,8-dialkylpyrrolo[3,2-f]quinazolines as inhibitors of CaDHFR by a GRID analysis of the threedimensional structure of CaDHFR. These compounds are potent inhibitors of fungal and human DHFR, with Ki values as low as 7.1 and 0.1 pM, respectively, and were highly active against C. albicans. These compounds are active in lung and brain tumor models and displayed in vivo activity against Pneumocystis carinii and C. albicans (Kuyper et al., 1996).
Regression analysis was considered for building the model. This CHEMBL dataset was preprocessed for molecular vectorization by applying PubChem fingerprint which generates 881 fingerprints using PaDEL software (Yap, 2011). By applying deep recurrent neural networks (RNN), these PubChem fingerprints were used to construct a regression model.
The deep learning algorithm needs usually two types of parameters to get the best model: model parameters and hyperparameters. Model parameters are internal to the neural network like neuron weights. They are estimated or learned automatically from train dataset and also used to make predictions by model, whereas hyperparameters are external parameters set by the user of the neural networklike, activation function, epoch, batch size, learning rate, number of nodes, number of neurons and hidden layers, etc. used in training. Hyperparameters shape how the network functions, and they have a huge impact on the accuracy of a neural network (Akl et al., 2019). The hyperparameters optimization involves finding the values of each hyperparameter which will help the model to provide the most accurate predictions. In deep-learning, there are various ways to optimize hyperparameters, from manual optimization to sophisticated algorithmic methods like grid search, random search, Bayesian optimization, (Mai et al., 2019) etc.

Manual hyperparameter optimization

The simplest way to select the best hyperparameters for a neural network model is ‘manual optimization’ – in other words, trial and error (Hutter et al., 2015). Here, hyperparameters were optimized manually by trial and error. It is commonly done to achieve very high accuracy for deep learning models.

Grid search

Grid search is slightly more sophisticated than manual optimization. It involves systematically testing multiple values of each hyperparameter, by automatically retraining the model for each value of the parameter. For example, we can perform a grid search for the optimal batch size by automatically training the model for batch sizes between 10 and 100 samples, in steps of 20. The model will run 5 times and the batch size selected will be the one that yields the highest accuracy. The main disadvantage of grid search is that it can be slow to run for large numbers of hyperparameter values

Random search

Here, hyperparameters are optimized or tested in a randomized manner and it is more effective than manual search or grid search (Bergstra & Bengio, 2012). In other words, instead of testing systematically to cover ‘promising areas’ of the problem space, it is preferable to test random values drawn from the entire problem space. It provides higher accuracy with few training cycles, for problems with high dimensionality. The major drawback to random search is that results are unintuitive, difficult to understand ‘why’ hyperparameter values were chosen.

Bayesian optimization

Bayesian optimization is a technique that tries to approximate the trained model with different possible hyperparameter values (Shahriari et al., 2016). Similar to sampling methods in statistics, the algorithm ends up with a list of possible hyperparameter value sets and model functions, from which it predicts the optimal function across the entire problem set. In Bayesian optimization and random search, results are not intuitive and difficult to improve on, even by trained operators.
In addition, there are various other algorithms and optimizers used in numerous engineering field like – spotted hyena optimizer (Dhiman & Kumar, 2017), emperor penguin optimizer (Dhiman & Kumar, 2018), seagull optimization algorithm (Dhiman & Kumar, 2019), STOA (Dhiman & Kaur, 2019), ESA (Dhiman, 2021) and other research fields (Dehghani et al., 2019; Dehghani, Montazeri, Dhiman, et al., 2020; Dehghani, Montazeri, Givi, et al., 2020; Dhiman et al., 2020, 2021; Kaur et al., 2020).
In this study, several regression models were created by manual optimization of hyperparameters to select the best model. During manual optimization, based on the choice of hyperparameters and their regression score, we change a part of them, train the model again and check the difference in the score to reach the best set of parameters for making a model having a good performance. The ReLU activation function (y ¼ (max (0, 1)) was applied in all the hidden layers, while the sigmoid function was used in the output layer.

Model evaluation and virtual screening

In this study, regression modeling of the dataset was carried out for developing the deep-learning model. Several deeplearning models were built and evaluated the model efficacy in terms of regression score (R squared (R2), mean squared error (MSE), root MSE (RMSE) and mean absolute error (MAE)). A good performance model should have a high R2 value and low Loss, MSE and RMSE values.
The best regression model was deployed on the SelleckFDA-approved-drug-library (Library id: L00011) of Selleck database which contains 1930 FDA-approved-drugs for virtual screening. The model predicted 500 screened hits by virtual screening.

Molecular docking

The screened compounds by deep-learning-based virtual screening were enriched by Molecular docking. Before the docking, the reference ligand (18H) was docked in the active site of CaDHFR to re-produce the same conformation similar to co-crystallized 18H. For docking, the grid center was set into X ¼ 3.34, Y ¼ 3.96 and Z ¼ 32.46, and dimensions of the three-dimensional grid box were set as 25 34.13 34.86 Å for X, Y and Z coordinates respectively. The number of exhaustiveness was set to eight for predicting the accurate result. After that, molecular docking simulation was performed with 500 screened drugs and reference with target protein by using PyRx open-source software with AutodockVina (Trott & Olson, 2010). AutodockVina proved to be more efficient and presented an accurate algorithm. The virtual screening of drugs was conducted by rigid molecular docking in the active site of receptor CaDHFR keeping ligand molecules flexible. Finally, the result in the form of binding energy was extracted from the software. The conformations of the drugs which had lower binding energy as compared to the reference molecule were chosen for further analysis.

Rescoring

The screened compounds obtained by molecular docking were rescored using X-score (Wang et al., 2002). Three different kinds of scoring functions viz. HPScore (hydrophobic pair), HMScore (hydrophobic match) and HSScore (hydrophobic surface) were used in X-score. Vander Waals interaction, H-bond interaction, rotatable bonds were denoted by VDW, H-bond, RT respectively.

Similarity search and visualization

All the hit compounds were further screened by similarity search using the ‘ChemmineR’ package of R (version 3.4.3). Those compounds that have the highest similarity with the reference molecule were chosen for molecular interaction analysis and molecular dynamics simulation (MDS). The Pymol and Ligplot þ v.1.4.5 software were used to visualize the 3D and 2D molecular interactions of docked (CaDHFRdrug) complexes. Ligplot program was used to identify and depicts the hydrophobic bonds, hydrogen bonds and the Hbond lengths of hit compounds at the active site of CaDHFR by building a 2D figure (Ligplot).

Molecular dynamics simulation

The highlighted one is the best regression model in term of R2, MSE, RMSE, MAE and loss.
The MDS was used to examine the insight behavior of the CaDHFR and CaDHFR-drug complexes under physiological conditions. MDS was executed by using GROMACS 5.0.7 (Pronk et al., 2013) on a workstation with configuration Ubuntu 16.04 LTS 64-bit, 4 GB RAM, IntelVRCoreTM i5-6400 CPU processor. The following protocol was used to perform the MDS. The ‘pdb2gmx’ script and CGenFF server were used for generating topology for protein and ligand respectively by applying CHARMM 36 force field (Vanommeslaeghe et al., 2009). Then the protein-ligand complex was formed by assembling these topologies. The solvation of the complex was achieved in the presence of explicit water molecules with the TIP3P water model in the dodecahedral boundary box condition. The distance between the complex and the box wall was kept as 10 Å, to prevent interaction of the complexes with its periodic boundary. After that, the neutralization of the complex was done by adding ions. Further, the energy of the complex was minimized with the steepest descent algorithm by using Verlet cut off-scheme at 10 kJ/mol by taking particle mesh Edward (PME) columbic interactions for 40,000 steps. The minimized complex was then subjected to 5000 ps simulation in NPT ensemble with 300 K temperature, and NPT ensemble with the constant pressure of 1 atm, and a time step of 2 fs. Further, equilibrated complexes were subjected to the production MD for 50 ns at a constant temperature of 300 K and a constant pressure of 1 atm using a time step of 2 fs. Finally, after MD simulation, the MD trajectories were used to calculate root mean square deviation (RMSD), root mean square fluctuation (RMSF), radius of gyration (Rg), hydrogen bonds, solvent accessible surface area (SASA), principal component analysis (PCA) and distance to analyze the stability of CaDHFR and CaDHFR-Drug complexes.

Binding free energies analysis

The binding free energy calculation by molecular mechanics Poisson–Boltzmann surface area (MM-PBSA) is widely used to analyze the stability of complexes (Kumari et al., 2014). Binding free energy consists of free solvation energy (polar and non-polar solvation energies) and potential energy (electrostatic and Vander Waals interactions). Here, binding free energy calculations of the CaDHFR-Drug complexes were done by the MMPBSA method for the last 10 ns of MD trajectory. The ‘python’ script provided in g_mmpbsa was used for average binding energy calculations.

Results and discussion

Predictive modeling, model evaluation and virtual screening

The interrelatedness between IC50 values and molecular fingerprints of CaDHFR inhibitors was modeled by a deep-learning algorithm to develop a predictive model for screening the compounds with better IC50 values than the CaDHFR inhibitors. For screening purposes, we constructed various deep-learning regression models using the ChEMBL dataset of CaDHFR inhibitors (CHEMBL2329) with the help of Pubchem fingerprint descriptors in the deep-learning server.
To build the deep-learning models, several hyperparameters were manually optimized. We have performed manual optimization 50 times and made 50 different regression models. The top 15 models selected in terms of their regression score (R2, MSE, RMSE, MAE and Loss) are listed in Table 1. A good performance model should have a high R2 value and low Loss, MSE, RMSE and MAE values. In Table 1, model number 14 although have the highest R2 value but it is showing a high loss, so this model is not considered good for prediction. From these 15 models, the 9th model was selected for prediction and screening as it has the highest R2 value as well as lowest loss, MSE, RMSE and MAE values. The best performing model was observed with R2 value (0.69), MSE (0.87), RMSE (0.93) and loss (0.87) (Figure 2).The best model was accomplished with the following hyperparameters; learning rate ¼ 0.01, Epochs ¼ 80, batch size ¼ 16, hidden layers ¼ 3, number of neuron ¼ 1500, 1000, 700 activation function ¼ ReLU, drop out ¼ 0 and output function ¼ sigmoid (Table 1).
Further, the resulting best deep-learning model was applied to perform virtual screening against FDA-approveddrug-library containing the 1930 FDA-approved-drugs. Deeplearning resulted in 500 hits with scores ranging from 9.23 to 5.7. Screened drugs were obtained from the server in a single SDF file. The single file was separated into individual SDF files using the R program ‘ChemmineR’ (version 3.4.3). The reference molecule 18 h, which was co-crystallized with CaDHFR protein were retrieved from the respective protein from Protein Data Bank in SDF format (Figure 4(A)). Using Open Babel, screened drugs, as well as reference molecule, were converted from SDF to mol2 chemical format.
Subsequently, non-polar hydrogens were merged while polar hydrogen was added and then saved in a dockable pdbqt format for the molecular docking process.

Molecular docking

G-Dayanandan et al. (2014) reported the discovery and optimization of 18H as an inhibitor of CaDHFR. 18H bind with the catalytic site residues via five hydrogen bonds, two with Glu32 and the remaining three with Ile9, Tyr118 and Ile112, and hydrophobic bond with Met25, Ile62, Thr58, Pro63, Phe66, Leu69, Phe36, Ala11 and Val10 of CaDHFR (Figure 3). All 500 drugs, screened by the deep-learning model were further screened by molecular docking against CaDHFR by using AutoDockVina. Before the screening, the docking protocol was validated by re-docking the reference 18H into the active site of CaDHFR to get the correct coordinates and docking pose. The binding pose showed that the docked 18H was superimposed to compare with the experimental 18H (Figure 4(B)). RMSD value is widely used to validate the docking protocol. Herein, the RMSD value between the docked and experimental 18H was 0.84 Å, which is perfectly acceptable. The interaction analysis revealed that both experimental, as well as docked 18H, showed interaction via hydrogen and hydrophobic bonds with the same amino acid residues as found in the CaDHFR crystal structure (Figure 3). The hydroxyl group of Tyr118 formed hydrogen bonds with N atoms of the 18H and the oxygen atom of Ile9, Ile112 and Glu32 interacts with the nitrogen atom of the 18H via Hbond in both docked and experimental one. Since this protocol produced a similar docked pose of 18H as in the crystal structure of CaDHFR, hence, it can be applied for further docking experiments.
After that, the 500 drugs were docked at the same coordinates. From molecular docking, 10 drugs were selected which showed better binding energy as compared to 18H (10.5 kcal mol1). Among the screened drugs, (S)-crizotinib showed the lowest binding energy i.e. 14.6 kcal mol1 followed by Crizotinib which had the binding energy of 13.6 kcal mol1 and Sonidegib showed the highest binding energy i.e. 10.5 kcal mol1 which was comparable to the 18H (Table 2). Then, all these 10 drugs and the 18H were further used for re-scoring.

Rescoring

The X-score was used for re-scoring to calculate the binding energy of the screened drugs toward CaDHFR. The score obtained from deep-learning, molecular docking and X-score (HP score, HM score, HS score, average score and associated binding energy) of reference molecule 18H and ten screened drugs are compiled in Table 2. The 18 h showed 10.5 kcal mol1 and 9.5 kcal mol1 binding energy obtained by molecular docking and X-score analysis respectively.
The results of X-score revealed that out of ten, five drugs viz. Paritaprevir, Lumacaftor, Rifampin, Sonidegib and Netupitant have strong affinities to CaDHFR. Among them, Paritaprevir had a deep-learning score ¼ 7.53. Paritaprevir showed kcal mol1 binding energy by X-score and 11.2 kcal mol1 binding energy by molecular docking. Lumacaftor had a deep-learning score ¼ 7.03 and 9.83 kcal mol1 and 11 kcal mol1 binding energy as obtained by molecular docking and X-score, respectively. Rifampin displayed a deep-learning score ¼ 6.64 and 9.52 kcal mol1 binding energy by X-score and 11.5 kcal mol1 binding energy by molecular docking. Sonidegib showed a 6.50 deep-learning score and binding energy 9.91 kcal mol1 and 10.5 kcal mol1 calculated by X-score and molecular docking, respectively. The compound Netupitant displayed the deep learning score ¼ 6.05 and binding energy of 12 kcal mol1 by molecular docking and 9.88 kcal mol1 by X-score.

Similarity search and visualization

A similarity search was performed using the ‘ChemmineR’ package of R (version 3.4.3). Out of five screened drugs, three shows better (more than 25% similarity) similarity with 18H (Table 2). Among these drugs, Lumacaftor showed the highest similarity (31%) followed b Paritaprevir (29% and 27% respectively).
PyMol and LigPlot þ v.1.4.5 were used to visualize CaDHFR-drug interactions. The docked poses of these three drugs with CaDHFR are shown in Figure 5. According to Figure 5(a), Ser61 formed a hydrogen bond between the N atom of the Paritaprevir and its O group, which has a bond distance of 3.10 Å. It also formed hydrophobic bonds with Met25, Ile112, Thr58, Ile62, Phe66, Leu69, Phe36 and Glu32. Lumacaftor interacted with several residues that also make a hydrogen bond, Arg72 having a bond distance of 3.15 Å. Here, the O atom of the Lumacaftor interacts with N atom of the Arg72 of CaDHFR. Besides, Ala11, Glu32, Ile9, Val110, Met25, Ile112, Ile62, Phe66, Pro70, Lys37, Phe36 and Leu69 residues were found to participate in hydrophobic interactions in the CaDHFR-Lumacaftor complex (Figure 5(b)). Rifampin formed a hydrogen bond with Met25 which has a bond distance of 3.11 Å and Lys24, Phe36, Leu69, Arg28, Ile62 and Ile112 were found to have participated in hydrophobic interactions (Figure 5(c)). The O atom of Rifampin interacts with N atom of the Met25 of CaDHFR.

Molecular dynamics simulation

The dynamics, stability and binding potential of screened drug conformation obtained as a result of molecular docking were further evaluated using MD simulation. The native protein CaDHFR and three protein-drug complexes (CaDHFRParitaprevir, CaDHFR-Lumacaftor and CaDHFR-Rifampin) were subjected to MD simulation for 50 ns. Besides, MD simulation of the CaDHFR-18H is also considered as a reference. Finally, we analyzed the MD trajectories for RMSD, RMSF, Rg, SASA, hydrogen bond formation and PCA analysis throughout the simulation. RMSD is essential to determine how the binding of the drug to the active site of the target protein affects its ability to reach an equilibrium state and also to quantify the structural stability of protein-drug complexes during the simulation. RMSD analysis of protein CaDHFR (black) and reference complex CaDHFR-18H (red) revealed that it achieves stability at around 2–3 ns and maintained its stability until 50 ns. CaDHFR-Paritaprevir (green) and CaDHFR-Rifampin (yellow) attained stability at around 10 ns and maintained that stability until the end (Figure 6(a)). CaDHFR-Lumacaftor complex shows some small fluctuations in RMSD during 15–35 ns, however, the value is decreased slightly after 38 ns and remained stable till 50 ns with the average fluctuation of 0.03 nm, suggesting that CaDHFR-Lumacaftor complex may undergo a small conformational change. Based on the RMSD result, all three drugs were very comparable to 18H, and the average values of RMSD in all compound-protein complexes comparable to the protein and protein-18H complex (Table 3). The least average value of RMSD is observed for the CaDHFR-Paritaprevir and CaDHFR-Rifampin simulation, which indicates their good stability. The fluctuation analysis of RMSD suggests that the MD trajectories are overall stable during the 50 ns simulation time for all the studied complexes.
Root mean square fluctuation (RMSF) is used to analyze how the binding of drugs affects the protein structural flexibility in the single amino acid level during MD simulation and presented in Figure 6(B). The higher value of RMSF indicates higher flexibility of protein-ligand complex. All the complexes showed comparable RMSF value to the CaDHFR and CaDHFR-18H complex, which suggests that they did not cause much fluctuation in protein during binding. In all studied complexes, the RMSF is lower than 0.2 nm, which means that the complexes are quite stable during the MD simulation. Only the CaDHFR-Lumacaftor complex demonstrates the highest RMSF value for Lys24 i.e. 0.23 nm, however, this residue is not involved in the binding site. Even though the average fluctuations of all three protein-drug complexes were very similar, the CaDHFR-Rifampin complex was found to have the highest, while CaDHFR-Paritaprevir and CaDHFR-Lumacaftor complexes had the lowest average RMSF value (Table 3). The RMSF results represented that all predicted complexes were stable, and hence, these drugs had the potential to inhibit the catalytic activity of CaDHFR.
We also analyzed other structural parameters from MD trajectories, such as radius of gyration (Rg), numbers of hydrogen bond, interaction energy, solvent-accessible surface area (SASA) and principal component analysis (PCA) based on essential dynamics (ED) approach. Based on the radius of gyration (Rg) analysis, which discloses the compactness of CaDHFR and CaDHFR-drug complexes, the CaDHFR-Paritaprevir undergoes minor compression in its 3D conformation (Figure 7(A)). Among the studied complexes, the most compressed complex was that of CaDHFR-18H and CaDHFRRifampin (Table 3).
Additionally, the total numbers of intermolecular H-bonds formed between CaDHFR-drug complexes during the MD simulations were also analyzed (Figure 7(B)). An average of four hydrogen bonds present between the CaDHFR-18H complex. Among the studied CaDHFR-drug complexes, the highest number of average hydrogen bonds is observed for CaDHFR-Lumacaftor complex i.e. three, whereas the lowest number of average hydrogen bonds is observed for CaDHFRParitaprevir and CaDHFR-Rifampin i.e. two (Table 3).
The interaction energy indicates the free energies of interaction associated with the binding of drugs with the CaDHFR structure. The average interaction energy of all the studied complexes was observed in the acceptable range of 100 to 150 kJ mol1 (Table 3). The average interaction energy calculated for the CaDHFR-18H complex was 126.73 kJ mol1 and all studied complexes show better interaction energy as compared to the reference complex. The CaDHFR-Paritaprevir complex shows the highest average interaction energy i.e. 141.58 kJ mol1 followed by the CaDHFR-Rifampin complex which has the interaction energy of 127.28 kJ mol1. Likewise, CaDHFR-Lumacaftor showed the average interaction energy value 122 kJ mol1 which is comparable to the reference complex. These interaction energy calculations were in agreement with molecular docking results and demonstrate that these drugs were favorably bound with the CaDHFR (Figure 8(A)).
Solvent accessible surface area (SASA) determines the amount of overall surface area in the protein structure which can be accessible by solvent and analyze interactions of the complex with solvent during the MD Simulation. SASA analysis indicated that the CaDHFR-Lumacaftor complex exhibited the highest value i.e. 113.06 nm2. The lowest average SASA value i.e. 111.16 nm2 is observed for the CaDHFRRifampin complex suggesting that the binding of Rifampin may reduce protein expansion. Table 3 demonstrates that CaDHFR-Paritaprevir and CaDHFR-Rifampin complex show almost a similar trend compared to CaDHFR-18H. SASA results demonstrate that the amount of overall surface area in the CaDHFR structure is decreased for solvent accessibility due to compression (Figure 8(B)).
Furthermore, the snapshot conformers show that the selected drugs (Paritaprevir, Lumacaftor and Rifampin) and 18H most of the time stayed in the binding site of the CaDHFR during the simulation run (Figure 9).
The PCA was performed to calculate concerted motions over 50 ns MD trajectories of native CaDHFR as well as the CaDHFR-drug complexes by using ED approach. For the PCA analysis, the first 40 eigenvectors were selected. Figure 10(A) depicts the first 40 eigenvectors obtained by diagonalization of the covariance matrix for all CaDHFR-drug complexes. Out of selected 40 eigenvectors, the first 10 principal eigenvectors represent 63.35%, 56.49%, 71.54% and 58.35% motions for CaDHFR-18H, CaDHFR-Paritaprevir, CaDHFR-Lumacaftor and CaDHFR-Rifampin respectively for the simulation period. The CaDHFR-Paritaprevir and CaDHFR-Rifampin complexes showed very fewer motions as compared with CaDHFR-18H. Based on PCA analysis, we concluded that CaDHFRParitaprevir and CaDHFR-Rifampin complexes did not cause significant motion and form stable complexes, also that CaDHFR-Paritaprevir showed the most stable complex due to least concerted motions. Further, ED was also used for projecting the 2D plot to understand the combined fluctuations of the most unsteady regions of CaDHFR into two variables, viz. PC1 and PC2 (Figure 10(B)). CaDHFR-Paritaprevir and CaDHFR-Rifampin complexes reside in less phase space and showed stable clusters which represent that these complexes are stable as compared to CaDHFR-18H and CaDHFRLumacaftor which had more conformational space. This PCA result was also in agreement with the above PCA result.
Additionally, Gibbs energy landscape plot for PC1 and PC2 was also created and is shown in Figure 11. This plot demonstrates energy values that for CaDHFR-18H, CaDHFRParitaprevir, CaDHFR-Lumacaftor and CaDHFR-Rifampin complexes, which range from 0 to 9.81 kJ mol1, 0 to 7.54 kJ mol1, 0 to 8.17 kJ mol1 and 0 to 8.55 kJ mol1. All three CaDHFR-drug complexes show low energy in comparison with the CaDHFR-18H complex, suggesting that these complexes follow energetically more favorable transitions between conformations. Besides, the energy minima region (blue region) is also more in these three complexes, proposing that these complexes are thermodynamically more favorable.

Binding free energies analysis

The MM-PBSA method was utilized to calculate the binding free energy between the CaDHFR-drug complexes. The binding free energies calculations were performed using the last 10 ns (40–50 ns) of MD trajectories and are shown in Table 4. From the MM-PBSA result, we found that all drugs showed a good binding affinity with CaDHFR when compared with the CaDHFR-18H complex. According to Table 4, CaDHFRParitaprevir showed the least free binding energy 114.91 kJ mol1) followed by CaDHFR-Lumacaftor (79.22 kJ mol1) and CaDHFR-Rifampin (78.52 kJ mol1), while CaDHFR-18H (reference complex) showed 63.10 kJ mol1 free binding energy during the MD simulation. The binding free energy analysis showed that all CaDHFR-drug were stable. It confirms that these drugs (Paritaprevir, Lumacaftor and Rifampin) can bind effectively at the binding site of CaDHFR protein.
Drug repurposing is a promising strategy in drug discovery of both common and rare diseases that distinguishes new therapeutic opportunities for existing medications by reducing the drug discovery and developmental time (Pushpakom et al., 2019). It potentially decreases the general expense of putting the drug for sale because the pharmacokinetic profiles and safety of the repositioned drug are already evaluated. Computational drug repurposing is vital to greatly reduce drug development time and expenses by finding novel uses for existing drugs (Park, 2019). Drug repurposing has been proven a good strategy for discovering new compounds against various bacterial, fungal and plasmodium diseases (Hoagland et al., 2016; Pazhayam et al., 2019). Here, we adopted this strategy to identify some new drugs to overcome the drug resistance against C. albicans and candidiasis disease. In this study, we conducted drug repurposing against CaDHFR target to find potential drugs against C. albicans. In numerous microorganisms, DHFR is a strong inhibitor of their growth. In different examinations, DHFR has been utilized as a drug target against pathogenic as well as non-pathogenic fungal species (Wright & Anderson, 2011). Antifolate drugs are utilized to treat malignant growth in humans by repressing DHFR resulting in depleting tetrahydrofolate (THF) and slowing the process of cell proliferation and DNA synthesis (Askari & Krajinovic, 2010).
Deep-learning techniques are a significant tool that can be utilized to build up a predictive model of an experimentally validated compound dataset and that model can be used for virtual screening of another unknown dataset. Using these methods, an antimalarial drug ‘pyrimethamine’ was identified against the enzyme dihydrofolate reductase (DHFR) and another drug BPM31510 is in a Phase 2 trial involving people with advanced pancreatic cancer (Chen et al., 2018; Fleming, 2018; Liu et al., 2019). Recently, the deep-learning algorithm was also used by some researchers to identify/ screen compounds against novel coronavirus 2019- nCov (SARS-Cov-2) by targeting 3C-like protease (Joshi et al., 2020; Zhang et al., 2020).
In this way, the current investigation was undertaken to identify some novel drugs that can be utilized against the CaDHFR by utilizing deep-learning-based virtual screening, molecular docking and dynamics simulation. This study showed that Paritaprevir, Lumacaftor and Rifampin drugs gave better activity against the CaDHFR (Figure 12). This in silico study gives the significant inhibition of CaDHFR and may empower the finding of potential therapeutic drugs from existing medications against candidiasis. Though, these drugs are already used to treat many other diseases in humans. The molecular targets, as well as modes of action of these drugs, are different. Paritaprevir (ABT-450) is a nonstructural (NS) protein 3/4A protease inhibitor used in the treatment of hepatitis C (de Leuw & Stephan, 2017). Lumacaftor is a pharmaceutical drug that acts to correct CFTR mutations in cystic fibrosis by increasing mutant CFTR (F508 del-CFTR) maturation (Kuk & Taylor-Cousar, 2015). The Rifampin is an antibiotic that is a DNA-dependent RNA polymerase inhibitor, used to treat several bacterial infections, tuberculosis, Mycobacterium avium complex, leprosy, meningitis and Legionnaires’ disease (Schonell et al., 1972). In addition, recently an in vitro study was conducted by He et al. (2017), showed that the amphotericin B-rifampin combination significantly increases the antifungal potential of amphotericin B against two fungal pathogen complex i.e. Fusarium solani species complex and Aspergillus flavus species complex isolates (He et al., 2017). Based on our results, we suggest that Paritaprevir, Lumacaftor and Rifampin drugs have the potential to be repurposed to overcome the drug resistance against C. albicans.

Conclusion

Candida albicans causes lethal infection candidiasis that costs extremely high to treat. Effective and safe drugs are required for treating this life threatening fungal diseases. Using drug repurposing techniques, we can re-utilize existing drugs against various diseases. The present study aims to identify potential drugs against C. albicans from already existing drugs using deep-learning methods. With the extremely high speed and moderately high precision, our deep-learning model for CaDHFR-ligand interaction analysis is ideally suitable to address the challenge of screening 10,000 of drugs in a short time. Our deep learning model is a data-driven model that learns CaDHFR-ligand interaction from known binding and non-binder data. The deep learning model is so fast and accurate as compared to all other molecular docking procedures. Herein, the deep-learning model predicted numerous potential drugs against CaDHFR. After that, different computational methods were applied and the outcomes indicated that three drugs viz. Paritaprevir, Lumacaftor and Rifampin could inhibit the activity of CaDHFR. The detailed analysis of Simulation and MM-PBSA results demonstrated that these drugs form a very stable complex with CaDHFR and showed excellent binding affinities as compared to the reference complex. Thus the outcome of this study showed that these drugs may represent potential treatment options against candidiasis. Since these drugs provided are on the market, therefore, screened drugs can help to facilitate the candidiasis drug development and could be used immediately. However, further in-vitro and in-vivo studies are necessary to investigate the potential anti-Candida activity of these drugs. This study can have a significant impact on the treatment of multidrug resistance fungus C. albicans.

References

Akl, A., El-Henawy, I., Salah, A., & Li, K. (2019). Optimizing deep neural networks hyperparameter positions and values. Journal of Intelligent & Fuzzy Systems, 37(5), 6665–6681. https://doi.org/10.3233/JIFS-190033
Askari, B. S., & Krajinovic, M. (2010). Dihydrofolate reductase gene variations in susceptibility to disease and treatment outcomes. Current Genomics, 11(8), 578–583. https://doi.org/10.2174/138920210793360925
Bergstra, J., & Bengio, Y. (2012). Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(2), 281–305.
Brown, G. D., Denning, D. W., & Levitz, S. M. (2012). Tackling human fungal infections. Science, 336(6082), 647. https://doi.org/10.1126/science.1222236
Brzankalski, G. E., Najvar, L. K., Wiederhold, N. P., Bocanegra, R., Fothergill, A. W., Rinaldi, M. G., Pattterson, T. F., & Graybill, J. R. (2008). Evaluation of aminocandin and caspofungin against Candida glabrata including isolates with reduced caspofungin susceptibility. The Journal of Antimicrobial Chemotherapy, 62(5), 1094–1100. https://doi.org/10.1093/jac/dkn304
Chen, H., Engkvist, O., Wang, Y., Olivecrona, M., & Blaschke, T. (2018). The rise of deep learning in drug discovery. Drug Discovery Today, 23(6), 1241–1250. https://doi.org/10.1016/j.drudis.2018.01.039 de Leuw, P., & Stephan, C. (2017). Protease inhibitors for the treatment of hepatitis C virus infection. GMS Infectious Diseases, 5, Doc08. https://doi.org/10.3205/id000034
Dehghani, M., Montazeri, Z., Dhiman, G., Malik, O. P., Morales-Menendez, R., Ramirez-Mendoza, R. A., Dehghani, A., Guerrero, J. M., & ParraArroyo, L. (2020). A spring search algorithm applied to engineering optimization problems. Applied Sciences, 10(18), 6173. https://doi.org/10.3390/app10186173
Dehghani, M., Montazeri, Z., Givi, H., Guerrero, J. M., & Dhiman, G. (2020). Darts game optimizer: A new optimization technique based on darts game. International Journal of Intelligent Engineering and Systems, 13(5), 286–294. https://doi.org/10.22266/ijies2020.1031.26
Dehghani, M., Montazeri, Z., Malik, O. P., Dhiman, G., & Kumar, V. (2019). BOSA: Binary orientation search algorithm. International Journal of Innovative Technology and Exploring Engineering, 9, 5306–5310.
Dhiman, G. (2021). ESA: A hybrid bio-inspired metaheuristic optimization approach for engineering problems. Engineering with Computers, 37(1), 323–331. https://doi.org/10.1007/s00366-019-00826-w
Dhiman, G., Garg, M., Nagar, A., Kumar, V., & Dehghani, M. (2020). A novel algorithm for global optimization: Rat swarm optimizer. Journal of Ambient Intelligence and Humanized Computing. https://doi.org/10. 1007/s12652-020-02580-0
Dhiman, G., & Kaur, A. (2019). STOA: A bio-inspired based optimization algorithm for industrial engineering problems. Engineering Applications of Artificial Intelligence, 82, 148–174. https://doi.org/10. 1016/j.engappai.2019.03.021
Dhiman, G., & Kumar, V. (2017). Spotted hyena optimizer: A novel bioinspired based metaheuristic technique for engineering applications. Advances in Engineering Software, 114, 48–70. https://doi.org/10.1016/ j.advengsoft.2017.05.014
Dhiman, G., & Kumar, V. (2018). Emperor penguin optimizer: A bioinspired algorithm for engineering problems. Knowledge-Based Systems, 159, 20–50. https://doi.org/10.1016/j.knosys.2018.06.001
Dhiman, G., & Kumar, V. (2019). Seagull optimization algorithm: Theory and its applications for large-scale industrial engineering problems. Knowledge-Based Systems, 165, 169–196. https://doi.org/10.1016/j. knosys.2018.11.024
Dhiman, G., Oliva, D., Kaur, A., Singh, K. K., Vimal, S., Sharma, A., & Cengiz, K. (2021). BEPO: A novel binary emperor penguin optimizer for automatic feature selection. Knowledge-Based Systems, 211, 106560. https://doi.org/10.1016/j.knosys.2020.106560
DiNubile, M. J., Hille, D., Sable, C. A., & Kartsonis, N. A. (2005). Invasive candidiasis in cancer patients: Observations from a randomized clinical trial. Journal of Infection, 50(5), 443–449. https://doi.org/10.1016/j. jinf.2005.01.016
Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639), 115–118. https://doi.org/10. 1038/nature21056
Esteva, A., Robicquet, A., Ramsundar, B., Kuleshov, V., DePristo, M., Chou, K., Cui, C., Corrado, G., Thrun, S., & Dean, J. (2019). A guide to deep learning in healthcare. Nature Medicine, 25(1), 24–29. https://doi.org/ 10.1038/s41591-018-0316-z
Fanos, V., & Cataldi, L. (2000). Amphotericin B-induced nephrotoxicity: A review. Journal of Chemotherapy, 12(6), 463–470. https://doi.org/10. 1179/joc.2000.12.6.463
Fisher, J. F., Sobel, J. D., Kauffman, C. A., & Newman, C. A. (2011). Candida urinary tract infections – Treatment. Clinical Infectious Diseases, 52(Suppl 6), S457–S466.
Fisher, M. C., Hawkins, N. J., Sanglard, D., & Gurr, S. J. (2018). Worldwide emergence of resistance to antifungal drugs challenges human health and food security. Science, 360(6390), 739–742. https://doi.org/10. 1126/science.aap7999
Fleming, N. (2018). How artificial intelligence is changing drug discovery. Nature, 557(7707), S55–S57. https://doi.org/10.1038/d41586-01805267-x
Fujitani, S., Ricardo-Dukelow, M., Kamiya, T., Sullivan, L., & Low, L. (2006). Ethnicity and other possible risk factors for candidemia at 3 tertiary care university hospitals in Hawaii. Infection Control & Hospital Epidemiology, 27(11), 1261–1263. https://doi.org/10.1086/508831
G-Dayanandan, N., Paulsen, J. L., Viswanathan, K., Keshipeddy, S., Lombardo, M. N., Zhou, W., Lamb, K. M., Sochia, A. E., Alverson, J. B., Priestley, N. D., Wright, D. L., & Anderson, A. C. (2014). Propargyllinked antifolates are dual inhibitors of Candida albicans and Candida glabrata. Journal of Medicinal Chemistry, 57(6), 2643–2656. https://doi. org/10.1021/jm401916j
Goodsell, D. S., Morris, G. M., & Olson, A. J. (1996). Automated docking of flexible ligands: Applications of AutoDock. Journal of Molecular Recognition, 9(1), 1–5. https://doi.org/10.1002/(SICI)1099-1352(1996 01)9:1<1::AID-JMR241>3.0.CO;2-6
He, J., Baxter, S. L., Xu, J., Xu, J., Zhou, X., & Zhang, K. (2019). The practical implementation of artificial intelligence technologies in medicine.
He, Y., Zhou, L., Gao, C., Han, L., & Xu, Y. (2017). Rifampin enhances the activity of amphotericin B against Fusarium solani species complex and Aspergillus flavus species complex isolates from keratitis patients. Antimicrobial Agents and Chemotherapy, 61(4), e02069-16. https://doi. org/10.1128/AAC.02069-16
He, Z., Dong, W., Li, Q., Qin, C., & Li, Y. (2018). Sauchinone prevents TGFb-induced EMT and metastasis in gastric cancer cells. Biomedicine & Pharmacotherapy, 101, 355–361. https://doi.org/10.1016/j.biopha.2018. 02.121
Hoagland, D. T., Liu, J., Lee, R. B., & Lee, R. E. (2016). New agents for the treatment of drug-resistant Mycobacterium tuberculosis. Advanced Drug Delivery Reviews, 102, 55–72. https://doi.org/10.1016/j.addr.2016. 04.026
Holbrook, S. Y. L., Garzan, A., Dennis, E. K., Shrestha, S. K., & GarneauTsodikova, S. (2017). Repurposing antipsychotic drugs into antifungal agents: Synergistic combinations of azoles and bromperidol derivatives in the treatment of various fungal infections. European Journal of Medicinal Chemistry, 139, 12–21. https://doi.org/10.1016/j.ejmech. 2017.07.030
Hutter, F., Lucke, J., & Schmidt-Thieme, L. (2015). Beyond manual tuning€ of hyperparameters. KI – Kunstliche Intelligenz€ , 29(4), 329–337. https:// doi.org/10.1007/s13218-015-0381-0
Jimenez, J., Skalic, M., Martınez-Rosell, G., & De Fabritiis, G. (2018). KDEEP: Protein-ligand absolute binding affinity prediction via 3D-convolutional neural networks. Journal of Chemical Information and Modeling, 58(2), 287–296. https://doi.org/10.1021/acs.jcim.7b00650
Joshi, T., Joshi, T., Pundir, H., Sharma, P., Mathpal, S., & Chandra, S. (2020). Predictive modeling by deep learning, virtual screening and molecular dynamics study of natural compounds against SARS-CoV-2 main protease. Journal of Biomolecular Structure and Dynamics. https://doi.org/10.1080/07391102.2020.1802341
Kaur, S., Awasthi, L. K., Sangal, A. L., & Dhiman, G. (2020). Tunicate swarm algorithm: A new bio-inspired based metaheuristic paradigm for global optimization. Engineering Applications of Artificial Intelligence, 90, 103541. https://doi.org/10.1016/j.engappai.2020.103541
Kuk, K., & Taylor-Cousar, J. L. (2015). Lumacaftor and ivacaftor in the management of patients with cystic fibrosis: Current evidence and future prospects. Therapeutic Advances in Respiratory Disease, 9(6), 313–326. https://doi.org/10.1177/1753465815601934
Kumari, R., Kumar, R., Open Source Drug Discovery Consortium, & Lynn, A. (2014). g_mmpbsa – A GROMACS tool for high-throughput MMPBSA calculations. Journal of Chemical Information and Modeling, 54(7), 1951–1962. https://doi.org/10.1021/ci500020m
Kuyper, L. F., Baccanari, D. P., Jones, M. L., Hunter, R. N., Tansik, R. L., Joyner, S. S., Boytos, C. M., Rudolph, S. K., Knick, V., Wilson, H. R., Caddell, J. M., Friedman, H. S., Comley, J. C., & Stables, J. N. (1996). High-affinity inhibitors of dihydrofolate reductase: Antimicrobial and anticancer activities of 7,8-dialkyl-1,3-diaminopyrrolo[3,2-f]quinazolines with small molecular size. Journal of Medicinal Chemistry, 39(4), 892–903. https://doi.org/10.1021/jm9505122
Lai, C. C., Tan, C. K., Huang, Y. T., Shao, P. L., & Hsueh, P. R. (2008). Current challenges in the management of invasive fungal infections. Journal of Infection and Chemotherapy, 14(2), 77–85. https://doi.org/10. 1007/s10156-007-0595-7
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. https://doi.org/10.1038/nature14539
Lin, H., Long, E., Ding, X., Diao, H., Chen, Z., Liu, R., Huang, J., Cai, J., Xu, S., Zhang, X., Wang, D., Chen, K., Yu, T., Wu, D., Zhao, X., Liu, Z., Wu, X., Jiang, Y., Yang, X., … Liu, Y. (2018). Prediction of myopia development among Chinese school-aged children using refraction data from electronic medical records: A retrospective, multicentre machine learning study. PLoS Medicine, 15(11), e1002674. https://doi.org/10. 1371/journal.pmed.1002674
Liu, J. Y., Shi, C., Wang, Y., Li, W. J., Zhao, Y., & Xiang, M. J. (2015). Mechanisms of azole resistance in Candida albicans clinical isolates from Shanghai, China. Research in Microbiology, 166(3), 153–161. https://doi.org/10.1016/j.resmic.2015.02.009
Liu, Z., Du, J., Fang, J., Yin, Y., Xu, G., & Xie, L. (2019). DeepScreening: A deep learning-based screening web server for accelerating drug discovery. Database, 2019, baz104. https://doi.org/10.1093/database/ baz104
Long, E., Lin, H., Liu, Z., Wu, X., Wang, L., Jiang, J., An, Y., Lin, Z., Li, X., Chen, J., Li, J., Cao, Q., Wang, D., Liu, X., Chen, W., & Liu, Y. (2017). An artificial intelligence platform for the multihospital collaborative management of congenital cataracts. Nature Biomedical Engineering, 1(2), 1–8. https://doi.org/10.1038/s41551-016-0024
Mai, L., Koliousis, A., Li, G., Brabete, A.-O., & Pietzuch, P. (2019). Taming hyper-parameters in deep learning systems. ACM SIGOPS Operating Systems Review, 53(1), 52–58. https://doi.org/10.1145/3352020.3352029
Odds, F. (2005). Genomics, molecular targets and the discovery of antifungal drugs. Revista Iberoamericana de Micologia, 22(4), 229–237. https://doi.org/10.1016/s1130-1406(05)70048-6
Odds, F. C., Brown, A. J., & Gow, N. A. (2003). Antifungal agents: Mechanisms of action. Trends in Microbiology, 11(6), 272–279. https:// doi.org/10.1016/s0966-842x(03)00117-3
Ostrosky-Zeichner, L., Marr, K. A., Rex, J. H., & Cohen, S. H. (2003). Amphotericin B: Time for a new ‘gold standard’. Clinical Infectious Diseases, 37(3), 415–425. https://doi.org/10.1086/376634
Otzen, T., Wempe, E. G., Kunz, B., Bartels, R., Lehwark-Yvetot, G., Hansel, W., Schaper, K. J., & Seydel, J. K. (2004). Folate-synthesizing enzyme system as target for development of inhibitors and inhibitor combinations against Candida albicans – Synthesis and biological activity of new 2,4-diaminopyrimidines and 40-substituted 4-aminodiphenyl sulfones. Journal of Medicinal Chemistry, 47(1), 240–253. https://doi.org/ 10.1021/jm030931w
Pappas, P. G., Kauffman, C. A., Andes, D. R., Clancy, C. J., Marr, K. A., Ostrosky-Zeichner, L., Reboli, A. C., Schuster, M. G., Vazquez, J. A., Walsh, T. J., Zaoutis, T. E., & Sobel, J. D. (2016). Clinical practice guideline for the management of candidiasis: 2016 update by the Infectious Diseases Society of America. Clinical Infectious Diseases, 62(4), e1–e50. https://doi.org/10.1093/cid/civ933
Park, B. J., Wannemuehler, K. A., Marston, B. J., Govender, N., Pappas, P. G., & Chiller, T. M. (2009). Estimation of the current global burden of cryptococcal meningitis among persons living with HIV/AIDS. AIDS, 23(4), 525–530. https://doi.org/10.1097/QAD.0b013e328322ffac
Translational and Clinical Pharmacology, 27(2), 59–63. https://doi.org/ 10.12793/tcp.2019.27.2.59
Pazhayam, N. M., Chhibber-Goel, J., & Sharma, A. (2019). New leads for drug repurposing against malaria. Drug Discovery Today, 24(1),263–271. https://doi.org/10.1016/j.drudis.2018.08.006
Perlroth, J., Choi, B., & Spellberg, B. (2007). Nosocomial fungal infections: Epidemiology, diagnosis, and treatment. Medical Mycology, 45(4), 321–346. https://doi.org/10.1080/13693780701218689
Pfaller, M. A. (2012). Antifungal drug resistance: Mechanisms, epidemiology, and consequences for treatment. The American Journal of Medicine, 125(1 Suppl), S3–S13. https://doi.org/10.1016/j.amjmed.2011. 11.001
Pronk, S., Pall, S., Schulz, R., Larsson, P., Bjelkmar, P., Apostolov, R., Shirts, M. R., Smith, J. C., Kasson, P. M., van der Spoel, D., Hess, B., & Lindahl, E. (2013). GROMACS 4.5: A high-throughput and highly parallel open source molecular simulation toolkit. Bioinformatics, 29(7), 845–854. https://doi.org/10.1093/bioinformatics/btt055
Pushpakom, S., Iorio, F., Eyers, P. A., Escott, K. J., Hopper, S., Wells, A., Doig, A., Guilliams, T., Latimer, J., McNamee, C., Norris, A., Sanseau, P., Cavalla, D., & Pirmohamed, M. (2019). Drug repurposing: Progress, challenges and recommendations. Nature Reviews Drug Discovery, 18(1), 41–58. https://doi.org/10.1038/nrd.2018.168
Rusk, N. (2016). Deep learning. Nature Methods, 13(1), 35–35. https://doi. org/10.1038/nmeth.3707
Sanglard, D., & Odds, F. C. (2002). Resistance of Candida species to antifungal agents: Molecular mechanisms and clinical consequences. The Lancet Infectious Diseases, 2(2), 73–85. https://doi.org/10.1016/s14733099(02)00181-0
Schonell, M., Dorken, E., & Grzybowski, S. (1972). Rifampin. Canadian Medical Association Journal, 106(7), 783–786.
Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., & De Freitas, N. (2016). Taking the human out of the loop: A review of Bayesian optimization. Proceedings of the IEEE, 104(1), 148–175. https://doi.org/10.1109/ JPROC.2015.2494218
Stylianou, M., Kulesskiy, E., Lopes, J. P., Granlund, M., Wennerberg, K., & Urban, C. F. (2014). Antifungal application of nonantifungal drugs. Antimicrobial Agents and Chemotherapy, 58(2), 1055–1062. https://doi. org/10.1128/AAC.01087-13
Trott, O., & Olson, A. J. (2010). AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of Computational Chemistry, 31(2), 455–461. https://doi.org/10.1002/jcc.21334
Vanommeslaeghe, K., Hatcher, E., Acharya, C., Kundu, S., Zhong, S., Shim, J., Darian, E., Guvench, O., Lopes, P., Vorobyov, I., & Mackerell, A. D., Jr. (2009). CHARMM general force field: A force field for drug-like molecules compatible with the CHARMM all-atom additive biological force fields. Journal of Computational Chemistry, 31(4), 671–690.
Volmer, A. A., Szpilman, A. M., & Carreira, E. M. (2010). Synthesis and biological evaluation of amphotericin B derivatives. Natural Product Reports, 27(9), 1329–1349. https://doi.org/10.1039/b820743g
Wang, R., Lai, L., & Wang, S. (2002). Further development and validation of empirical scoring functions for structure-based binding affinity prediction. Journal of Computer-Aided Molecular Design, 16(1), 11–26. https://doi.org/10.1023/a:1016357811882
Wilson, L. S., Reyes, C. M., Stolpman, M., Speckman, J., Allen, K., & Beney, J. (2002). The direct cost and incidence of systemic fungal infections. Value in Health, 5(1), 26–34. https://doi.org/10.1046/j.1524-4733.2002.
Wright, D. L., & Anderson, A. C. (2011). Antifolate agents: A patent review (2006–2010). Expert Opinion on Therapeutic Patents, 21(9), 1293–1308. https://doi.org/10.1517/13543776.2011.587804
Yap, C. W. (2011). PaDEL-descriptor: An open source software to calculate molecular descriptors and fingerprints. Journal of Computational Chemistry, 32(7), 1466–1474. https://doi.org/10.1002/jcc.21707
Zhang, H., Saravanan, K. M., Yang, Y., Hossain, M. T., Li, J., Ren, X., Pan, Y., & Wei, Y. (2020). Deep learning based drug screening for novel coronavirus 2019-nCov. Interdisciplinary Sciences: Computational Life Sciences, 12(3), 368–376. https://doi.org/10.1007/ s12539-020-00376-6