Identification of Suitable Drug Combinations for Treating COVID-19 Using a Novel Machine Learning Approach: The RAIN Method

Simple Summary This study follows an improved approach to systematic reviews, called the Systematic Review and Artificial Intelligence Network Meta-Analysis (RAIN), registered within PROSPERO (CRD42021256797), in which, the PRISMA criterion is still considered. Drugs used in the treatment of COVID-19 were searched in the databases of ScienceDirect, Web of Science (WoS), ProQuest, Embase, Medline (PubMed), and Scopus. In addition, using artificial intelligence and the measurement of the p-value between human genes affected by COVID-19 and drugs that have been suggested by clinical experts, and reported within the identified research papers, suitable drug combinations are proposed for the treatment of COVID-19. During the systematic review process, 39 studies were selected. Our analysis shows that most of the reported drugs, such as azithromycin and hydroxyl-chloroquine on their own, do not have much of an effect on the recovery of COVID-19 patients. Based on the result of the new artificial intelligence, on the other hand, at a significance level of less than 0.05, the combination of the two drugs therapeutic corticosteroid + camostat with a significance level of 0.02, remdesivir + azithromycin with a significance level of 0.03, and interleukin 1 receptor antagonist protein + camostat with a significance level 0.02 are considered far more effective for the treatment of COVID-19 and are therefore recommended. Abstract COVID-19 affects several human genes, each with its own p-value. The combination of drugs associated with these genes with small p-values may lead to an estimation of the combined p-value between COVID-19 and some drug combinations, thereby increasing the effectiveness of these combinations in defeating the disease. Based on human genes, we introduced a new machine learning method that offers an effective drug combination with low combined p-values between them and COVID-19. This study follows an improved approach to systematic reviews, called the Systematic Review and Artificial Intelligence Network Meta-Analysis (RAIN), registered within PROSPERO (CRD42021256797), in which, the PRISMA criterion is still considered. Drugs used in the treatment of COVID-19 were searched in the databases of ScienceDirect, Web of Science (WoS), ProQuest, Embase, Medline (PubMed), and Scopus. In addition, using artificial intelligence and the measurement of the p-value between human genes affected by COVID-19 and drugs that have been suggested by clinical experts, and reported within the identified research papers, suitable drug combinations are proposed for the treatment of COVID-19. During the systematic review process, 39 studies were selected. Our analysis shows that most of the reported drugs, such as azithromycin and hydroxyl-chloroquine on their own, do not have much of an effect on the recovery of COVID-19 patients. Based on the result of the new artificial intelligence, on the other hand, at a significance level of less than 0.05, the combination of the two drugs therapeutic corticosteroid + camostat with a significance level of 0.02, remdesivir + azithromycin with a significance level of 0.03, and interleukin 1 receptor antagonist protein + camostat with a significance level 0.02 are considered far more effective for the treatment of COVID-19 and are therefore recommended. Additionally, at a significance level of less than 0.01, the combination of interleukin 1 receptor antagonist protein + camostat + azithromycin + tocilizumab + oseltamivir with a significance level of 0.006, and the combination of interleukin 1 receptor antagonist protein + camostat + chloroquine + favipiravir + tocilizumab7 with corticosteroid + camostat + oseltamivir + remdesivir + tocilizumab at a significant level of 0.009 are effective in the treatment of patients with COVID-19 and are also recommended. The results of this study provide sets of effective drug combinations for the treatment of patients with COVID-19. In addition, the new artificial intelligence used in the RAIN method could provide a forward-looking approach to clinical trial studies, which could also be used effectively in the treatment of diseases such as cancer.


Introduction
Coronavirus 2019 (COVID- 19) is an infectious disease caused by severe acute respiratory coronavirus (SARS-CoV-2). The disease was first identified in December 2019 in Wuhan, the capital of China's Hubei Province, and has since spread worldwide [1]. The virus is mainly transmitted between people during close contact, often through small droplets produced by coughing, sneezing, or talking [2].
Today, the international community and healthcare professionals are affected by the epidemic of this coronavirus (COVID-19) [3]. Although most patients with this disease have mild infectious symptoms, about 20% of patients develop severe and acute pneumonia, which can lead to death [4]. In addition to respiratory symptoms, COVID-19 causes widespread inflammation and severe damage to other organs. The activation of immune cells, release of inflammatory factors such as interleukin and tumor necrosis factor-alpha, D-dimer, ferritin, and C-reactive protein, and activation of platelets cause inflammation and widespread vascular disorders due to this disease [5]. Damages to the heart [6], brain [7], kidney [8], and liver [9] are also observed due to severe manifestations of the disease.
Standard supportive care, corticosteroids, intravenous immunoglobulin, and experimental or alternative antiviral therapies (such as ramsedivir, ribavirin, lupinavir-ritonavir, amifenovir, and interferons, etc.) have been different treatment strategies for patients with COVID-19 [4]. To date, none of the proposed therapies have been introduced as the main treatment for this disease. However, in the face of the pressures caused by the pandemic, health care workers around the world were forced to prescribe these drugs despite inconclusive clinical evidence [10].
An effective treatment strategy worldwide is essential for the rapid treatment of patients with COVID-19. There are a huge amount of COVID-19 treatment data in healthcare clinics around the world, which demonstrates the need for advanced methods of analysis well [11]. Artificial intelligence is one of the methods that can easily track the prevalence of COVID-19, identify at-risk patients, and be useful in controlling this infection in real time [12]. Artificial intelligence techniques also make it possible to analyze patients' previous data, predict the risk of death, and help fight the virus with population screening, medical assistance, information, and infection control suggestions. This technology, as an evidence-based medical tool, can improve the planning, treatment, and reported outcomes of COVID-19 patients [13,14] Network meta-analysis is another way to find effective drugs for the treatment of a disease. Using network meta-analyses based on existing clinical trials, high-quality comparative evidence can be generated to evaluate the drugs used to treat COVID-19 [15]. In addition, a network meta-analysis model can estimate treatment effects even for treatments that have never been directly compared in a direct study [16].
Systematic review and meta-analysis articles have been widely used by researchers worldwide, especially in the COVID-19 pandemic. The international community and researchers have always tried to counteract COVID-19 by determining the target group of the disease, methods of prevention, diagnosis, treatment, and vaccine discovery, which have also had noteworthy effects. However, due to the wide range of emerging COVID-19 variations due to mutations, despite the discovery of a number of vaccines, we still have several patients who require specialist medical care. We still do not know: what drug combinations and what diets can be useful for treating patients; whether Chinese medicine and common medical treatments can be used in the treatment of this disease or not; and whether children manifest the same symptoms as adults. Therefore, researchers globally should seriously consider the need for certainty and validity of information in order to save patients' lives. In this study, we assessed the impact and application of artificial intelligence to provide responses to some of the above questions by conducting a novel systematic review and meta-analysis. This novel systematic review approach is referred to as Systematic Review and Artificial Intelligence Network Meta-Analysis (RAIN), and the research protocol was registered within the PROSPERO international prospective register of systematic reviews, (CRD42021256797).

Materials and Methods
This new protocol has three general stages, each of which entails a number of activities. The first stage includes a systematic review that entails the review, the results of the identified and selected studies, and summaries of the clinical results. The second stage includes a meta-analysis of the results, if necessary, to provide quantitative information (this section can be skipped if the researchers find that the heterogeneity among the selected studies is high and that meta-analysis is not possible). Finally, the third stage is the analysis of the results of one-on-one studies using artificial intelligence and a machine learning approach. In this stage, artificial intelligence proposes an approach using specific scenarios at the significance level of 0.05 and 0.01 by reviewing the results of one-on-one studies and adding and subtracting results and variables. The study outlined in this manuscript followed the same set of phases of RAIN:

Stage I: Systematic Review
The purpose of the systematic review is to collect all of the information obtained from studies conducted on the treatments for COVID-19 patients and to extract the reported results within the selected studies. The systematic review section was performed in accordance with the Cochrane 7-step approach, including: research question selection, defining inclusion and exclusion criteria, article identification, study selection, study quality evaluation, data extraction, and analysis and interpretation of findings.

Research Question and Keyword Determination according to PICO Instructions
According to the PICO guidelines, the research question is: "What are the most effective drugs and drug combinations for the treatment of patients with COVID-19?". The study population included patients with COVID-19 worldwide, the intervention included therapeutic measures and the use of specific drugs for the treatment of COVID-19 patients, comparison included an evaluation of the effect of specific drugs in the study with different drug therapies, and the desired outcome was explained in accordance with the outcome including the effect or not of the drug in the treatment of the disease.

Searching Articles
The keywords were extracted from the Medical Subject Headings (MeSH) browser and in accordance with the PICO instructions. For each question, keywords were selected

Identifying Articles
Initially, a large number of abstracts of articles in which COVID-19 was mentioned were discovered by two researchers in our study. Through text mining, keywords related to COVID-19 were automatically ranked (e.g., affected human genes, effective drugs, etc.) to increase the accuracy of the results obtained by the researchers, and not to miss a related keyword.
International databases of ScienceDirect, Web of Science (WoS), ProQuest, Embase, Medline (PubMed), and Scopus were searched with lower time limit of 1 December 2019, and up to 31 December 2021. Articles published in English as well as articles in other languages, yet with an abstract in English, were reviewed. The search strategy for each database was determined through the Advance Search, and was followed using all possible keyword combinations and using AND and OR operators.

Study Selections, Based on Inclusion and Exclusion Criteria
The reference lists of all articles that met the inclusion criteria were manually reviewed to access further relevant studies. To avoid errors, mistakes, and also bias, all steps of article search, study selection, quality evaluation, and data extraction were performed by two reviewers (researchers) independently. If there was a difference in opinion among the reviewers regarding the inclusion of an article in our study, a third reviewer reviewed the article. The search process was also reconstructed by artificial intelligence and the machine learning process to increase search accuracy.
The information of all of the articles found in each database was transferred into the EndNote X8 reference management software and then in Python in the form of a data frame. After searching all of the databases, duplicate articles were removed. Then, in order to avoid the risk of bias in the selection of studies, the names of the authors and the titles of the journals were omitted and a checklist was prepared based on the titles and abstracts of the studies. Studies where their full text was not found and where they did not meet the inclusion criteria were excluded from the systematic review process. The full texts of all remaining articles were then evaluated [10,11].

Inclusion and Exclusion Criteria according to the Research Question
Studies involving specific drugs or drug combinations for the treatment of patients with COVID-19 globally, including clinical trial studies, were included in our work. Observational research, including case studies, cohorts, descriptive studies, rare case reports, review studies, and meta-analyses, were all excluded.

Quality Evaluation of Studies
In order to evaluate the quality of the studies, the CONSORT checklist was adopted. This checklist is helpful in reviewing clinical trial studies, and includes 25 general items, each with sub-items (a total of 37 sub-items). The items of this checklist include: Title, Abstract, Background, Objectives, Methods, Participants, Interventions, Objectives, Consequences, Sample Size, Randomization, Participants flow, Blinding, etc. In order to rate the articles, if each article referred to the items considered within the checklist, it was given a score of 1 and, if it was not mentioned, a score of zero was allocated. The minimum and maximum scores in this checklist are 0 and 37, respectively. Studies with 75% or more of the maximum achievable score (score greater than or equal to 27) were considered as "high quality", studies with a score between 50-75% (score [18][19][20][21][22][23][24][25][26] were considered as "average quality", and studies with a score lower than 50% (score less than or equal to 17) were considered as "low quality" studies [12].

Data Extraction
After selecting the studies in the systematic review, in order to select the studies to enter the meta-analysis, the four steps of Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2009 process were considered, i.e., article identification, screening, eligibility assessment, and, finally, inclusion of articles that can also be submitted for meta-analysis ( Figure 1). maximum scores in this checklist are 0 and 37, respectively. Studies with 75% or more of the maximum achievable score (score greater than or equal to 27) were considered as "high quality", studies with a score between 50-75% (score 18-26) were considered as "average quality", and studies with a score lower than 50% (score less than or equal to 17) were considered as "low quality" studies [12].

Data Extraction
After selecting the studies in the systematic review, in order to select the studies to enter the meta-analysis, the four steps of Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2009 process were considered, i.e., article identification, screening, eligibility assessment, and, finally, inclusion of articles that can also be submitted for meta-analysis ( Figure 1).

Stage 2: Artificial Intelligence
At this stage, the results of studies obtained from the network meta-analysis were imported to be used by AI algorithm. The AI algorithm receives the effectiveness of each drug in the affected human genes by COVID-19 as input, and, in the output, provides some effective scenarios. Each scenario includes the drug combinations proposed for the treatment of COVID-19.
The entire genome sequence of COVID-19 was constructed by researchers, and the following proteins are some of the identified human genes that are mostly affected by

Stage 2: Artificial Intelligence
At this stage, the results of studies obtained from the network meta-analysis were imported to be used by AI algorithm. The AI algorithm receives the effectiveness of each drug in the affected human genes by COVID-19 as input, and, in the output, provides some effective scenarios. Each scenario includes the drug combinations proposed for the treatment of COVID-19.
The percent identity, E-value, and query cover measures of the National Center for Biotechnology Information (NCBI) find the most similar nucleotides with these proteins. The first measure is to find how similar the query sequence is to the associated nucleotides. The second one describes how many times it would be expected to match by chance in a database of that size. The third one describes how much of the query sequence is covered by the target sequence. To carry this out, we found the most similar nucleotides related to each protein, separately. Then, the results of each measure were aggregated. Table 1 shows the mean and standard deviation for each measure. The first column shows nucleotides that are the most similar to the COVID-19 proteins (Table 1).

Combining p-Values
Evolutionary biologists have long used meta-analytic tactics to combine evidence from several existing studies. If raw data cannot be summarized, p-value-based meta-analysis offers a hands-on method that can be practically as influential as combining data. Many common p-value combining methods have the same general shape. At first, the p-value for the i-th study, p i , is transformed by a function H. Then, in the new distribution, one type of mean, such as arithmetic or weighted ones, is used to calculate the combination T = ∑ w i H(p i ). Finally, the inverse transform of T is computed as the result of the combined p-value. Two common methods used to combine p-values are Fisher's methods and Stouffer's method. Fisher's method combines p-values into one test statistic (chi-squared). The boxplot of Table 1 when considering the present identity is shown in Figure 2.

Using the Formula
In the interval [0,1], the negative logarithm of this distribution reveals a quantity that follows exponential distribution. Moreover, when scaling a value that follows an exponential one by a factor of two, it will follow a chi-squared distribution with two degrees of freedom. Fisher's method computes the sum of these k-independent chi-squared values, where it follows a chi-squared distribution with 2k degrees of freedom. However, this method does not use any weighting approach, whereas, for our aim, the sizes of studies are important.
Stouffer's method, as another known work, uses the inverse normal distribution function to transform the uniform distribution, allowing for the incorporation of study weights using z-scores [21]. One of its improvements is Lipták's method, in which it considers weights in the Stouffer's method [22]. This method is also known as the weighted Z-test and the combined p-value is computed by: where i is the number of study from k studies, pi is its p-value, wi is its weight, Zi = ∅ −1 (1 − pi), and ∅ indicates the standard normal cumulative distribution function [23]. Taking into account the correct weights leads to a better combination. If only study sample sizes are available, their square roots converge for an optimal result [22]. Won et al. demonstrated that the effect size over the standard error has the optimal power [24]. We used a combination method based on the weighted version of power mean in the transformed distribution. We proposed a machine learning method that uses p-values between the COVID-19 and affected human genes, as well as p-values between those genes and drug associations as input. The output finds associations' cooperations (drug recipe in this article) that make small combined p-value with COVID-19. 2.11. Using the Formula In the interval [0,1], the negative logarithm of this distribution reveals a quantity that follows exponential distribution. Moreover, when scaling a value that follows an exponential one by a factor of two, it will follow a chi-squared distribution with two degrees of freedom. Fisher's method computes the sum of these k-independent chi-squared values, where it follows a chi-squared distribution with 2k degrees of freedom. However, this method does not use any weighting approach, whereas, for our aim, the sizes of studies are important.
Stouffer's method, as another known work, uses the inverse normal distribution function to transform the uniform distribution, allowing for the incorporation of study weights using z-scores [21]. One of its improvements is Lipták's method, in which it considers weights in the Stouffer's method [22]. This method is also known as the weighted Z-test and the combined p-value is computed by: where i is the number of study from k studies, p i is its p-value, w i is its weight, Z i = ∅ −1 (1 − p i ), and ∅ indicates the standard normal cumulative distribution function [23]. Taking into account the correct weights leads to a better combination. If only study sample sizes are available, their square roots converge for an optimal result [22]. Won et al. demonstrated that the effect size over the standard error has the optimal power [24]. We used a combination method based on the weighted version of power mean in the transformed distribution. We proposed a machine learning method that uses p-values between the COVID-19 and affected human genes, as well as p-values between those genes and drug associations as input. The output finds associations' cooperations (drug recipe in this article) that make small combined p-value with COVID-19.

Power Means
To combine sets of numbers, power means are a type of function that implements positive real numbers. The power means of p 1 , . . . , p n are calculated by [25]: M −∞ is the minimum function, M −1 is the harmonic mean, M 0 is the geometric mean, M 1 is the arithmetic mean, M 2 is the quadratic mean, M 3 is the cubic mean, and M −∞ is the maximum of p 1 , . . . , p n [26]. With k = 0, the power means summarize to the geometric mean: Considering that each p i has the weight w i , the weighted geometric mean is computed by:

The Proposed AI Models
The proposed p-nearest associations algorithm (p-NA) is a nonparametric method used to propose associations' cooperations for the target. The input consists of associations where the p-values between them and interface features are known from the network metaanalysis. The output contains associations' cooperations that have a small significance. Figure 3 shows the overall structure of p-NA.

The Forward
Step where is the combined p-value between the association and the target. Let ( ) , , … , ( ) , be a reordering of the training data such that ( ) ⋯ . Figure 4 shows an example of the p-NA algorithm. The black circle on the left is the target and the associations are embedded in the axis with their combined p-values with the target. The associations with combined pvalues less than 0.01 are shown in green (i. e. , ( ) , ( ) ) and the others that are less than p = 0.05 are blue (i. e. , ( ) , ( ) ). Other ones are shown in red ( Figure 4).  The proposed model consists of forward and backward steps. In the forward step, the proposed p-NA algorithm computes all combined p-values between associations (drugs) and the target (COVID-19). Then, the association with the smallest combined p-value is selected. The combined p-value is calculated based on interface features (human genes). In the backward step, based on the selected association(s), all p-values between interface features and the target are updated. The significance of cooperating associations in each scenario is computed by multiplying the combined p-values of those associations in that scenario. These steps are iterated until that significance becomes less than a threshold ( Figure 3).

The Forward Step
Suppose we have pairs A1, cp T A1 , . . . , An, cp T An , where c T Ai is the combined p-value between the ith association and the target. Let A (1) , cp 1 , . . . , A (n) , cp n be a reordering of the training data such that cp (1) ≤ . . . ≤ cp n . Figure 4 shows an example of the p-NA algorithm. The black circle on the left is the target and the associations are embedded in the axis with their combined p-values with the target. The associations with combined p-values less than 0.01 are shown in green i.e., A (1) , A (2) and the others that are less than p = 0.05 are blue i.e., A (3) , A (6) . Other ones are shown in red ( Figure 4).

The Forward
Step where is the combined p-value between the association and the target. Let ( ) , , … , ( ) , be a reordering of the training data such that ( ) ⋯ . Figure 4 shows an example of the p-NA algorithm. The black circle on the left is the target and the associations are embedded in the axis with their combined p-values with the target. The associations with combined pvalues less than 0.01 are shown in green (i. e. , ( ) , ( ) ) and the others that are less than p = 0.05 are blue (i. e. , ( ) , ( ) ). Other ones are shown in red ( Figure 4). In other words, if the combined p-values between associations and the target are found, the forward step simply chooses the nearest association with the viewpoint of combined p-values. Thus, the forward step is summarized as computing the combined p-values between associations and the target.
Under the null hypothesis, the p-values follow the uniform distribution on the interval [0,1]. The natural logarithm of a uniformly distributed value follows an exponential distribution. The weighted mean of that exponential distribution is ∑ ∑ , where is the p-value for association. Finally, the inverse transform of the logarithm yields a quantity that follows a uniform distribution on the interval [0,1]. It is the same as the weighted geometric mean (WGM) in the original distribution. The following formula shows the combined p-value between an association and the target: where is the combined p-value between the j-th association and the target, = 1 − is the weight between the i-th interface feature and the target, is the p-value In other words, if the combined p-values between associations and the target are found, the forward step simply chooses the nearest association with the viewpoint of combined p-values. Thus, the forward step is summarized as computing the combined p-values between associations and the target.
Under the null hypothesis, the p-values follow the uniform distribution on the interval [0,1]. The natural logarithm of a uniformly distributed value follows an exponential distribution. The weighted mean of that exponential distribution is Finally, the inverse transform of the logarithm yields a quantity that follows a uniform distribution on the interval [0,1]. It is the same as the weighted geometric mean (WGM) in the original distribution. The following formula shows the combined p-value between an association and the target: where cp T Aj is the combined p-value between the j-th association and the target, w Fi T = 1 − p Fi T is the weight between the i-th interface feature and the target, P Fi T is the p-value between the target and i-th interface feature, and P Aj Fi is the p-value between the i-th interface feature and j-th association.
As a simple case of a genome with just two genes (interface features), Figure 3 shows the construction of a combined p-value for just one association. Assume that two genes are independent and thus perpendicular to each other. p A1 F1 , p A1 F2 is the p-value between these genes and the association. The weight of gene i is w Fi T . Rotating the line segment p A1 T is shown in a red dashed curve. The measure of the combined p-value is shown red line ( Figure 5).
Therefore, the combined p-values between associations and the target are computed and the association with the smallest combined p-value is selected as the p-nearest association.
For this paper's case study, i.e., the drugs recipe for COVID-19, the p-values between drugs and affected human genes, and p-values between affected human genes and COVID-19, are available at Coremine Medical online database. It uses text mining algorithms to search and explore biomedical connections.
After obtaining the drugs suggested by clinical experts, the human genes affected by COVID-19, we collected all of the p-values between drugs and genes and between COVID-19 and genes. We separated each association and then found the combined p-value using the proposed WGM. The nearest association was then found in the forward step of p-NA. Figure 6 shows the results of p-NA's forward step for associations in the drug categories.
terface feature and j-th association.
As a simple case of a genome with just two genes (interface features), Figure 3 shows the construction of a combined p-value for just one association. Assume that two genes are independent and thus perpendicular to each other.
, is the p-value between these genes and the association. The weight of gene i is . Rotating the line segment ( ) in the direction of gene 1 is shown with a dashed green curve. A circle with its center in the gene 1 line that touches the ( ) and rotated line segment of ( ) is shown in a red dashed curve. The measure of the combined p-value is shown red line ( Figure 5). Therefore, the combined p-values between associations and the target are computed and the association with the smallest combined p-value is selected as the p-nearest association.
For this paper's case study, i.e., the drugs recipe for COVID-19, the p-values between drugs and affected human genes, and p-values between affected human genes and COVID-19, are available at Coremine Medical online database. It uses text mining algorithms to search and explore biomedical connections.
After obtaining the drugs suggested by clinical experts, the human genes affected by COVID-19, we collected all of the p-values between drugs and genes and between COVID-19 and genes. We separated each association and then found the combined p-value using the proposed WGM. The nearest association was then found in the forward step of p-NA. Figure 6 shows the results of p-NA's forward step for associations in the drug categories. We also used the proposed p-NA to find nearest associations in categories such as foods, MeSH, molecular function, gene/protein, or biological process 3. However, this paper focuses on drugs 4.

The Backward Step
In this step, the proposed p-NA algorithm updates the weights between the target and all interface features. This is computed by multiplying the weights with the p-values of the selected association and those interface features; i.e., (Figure 6): where is the selected association in the forward step. Obviously, is updated by 1 − .

Systematic Review
The articles used in this study were published in the period 2020-2021. Of the 39 con- We also used the proposed p-NA to find nearest associations in categories such as foods, MeSH, molecular function, gene/protein, or biological process 3. However, this paper focuses on drugs 4.

The Backward Step
In this step, the proposed p-NA algorithm updates the weights between the target and all interface features. This is computed by multiplying the weights with the p-values of the selected association and those interface features; i.e., (Figure 6): where A selected is the selected association in the forward step. Obviously, p Fi T is updated by 1 − w Fi T .
Common and standard treatments were performed for the control group in 11 clinical trial studies with HCQ. The results of these studies indicate that HCQ has no significant effect on reducing mortality and morbidity and improving clinical symptoms caused by COVID-19 [37,49,50,53,55,57,58,62,63]. In another work, the effect of HCQ in preventing the treatment of COVID-19 was measured, which also stated that HCQ is not effective in preventing the disease [27]. This result was also observed in the study of Mitjà O. et al. [49], in which, the authors examined healthy individuals exposed to COVID-19. Only in one work was the relative effect of HCQ on the recovery of patients with COVID-19 observed [51].
A comparison between HCQ and a placebo group was observed in four studies. Most of these studies have shown that HCQ has no effect on the recovery, reduction in symptoms, and prevention of mortality in patients with COVID-19 [29,35,36]. However, the results of one study indicate that the administration of the HCQ drug prevents the onset of severe symptoms of the disease and, in these patients, only mild gastrointestinal symptoms were observed [61].
In many other studies, HCQ was prescribed as a control drug in both groups. In the study of Ansarin et al., the effect of bromhexine with HCQ was measured. The results of this study showed that oral bromhexine has a positive effect on reducing mortality, transfer to the ICU, and the rate of patients' intubation [38]. In another study, the effect of azithromycin with HCQ was measured in patients, but no positive results were obtained in this study [55]. Another article reported the examination of the effect of arbidol with HCQ in COVID-19 patients. This study stated that arbidol has a positive effect on reducing clinical symptoms, the patient respiratory status, ICU transfer, and hospitalization time [42]. However, it is not possible to confirm the results of these studies until wider trials are implemented.
The effect of azithromycin was evaluated in two studies. In one study, a significant relationship was observed between the improvement in patients' clinical symptoms [56] and, in another study, the positive effect of azithromycin on the improvement in patients' symptoms was confirmed [45]. Two other studies measured the effect of interferon on COVID-19 patients. The results of these studies showed a reduction in mortality and clinical symptoms in patients with COVID-19 [33,39]. In four other studies performed in Iran, the effect of sofosbuvir in patients was measured. The results of three studies showed a positive effect of drug and recovery and reduction in patient mortality [40,41,44], while another study did not report any significant relationship between sofosbuvir use and patients' recovery [42]. According to the information in Table 1, other drugs were tested in different trials, though, due to the limited population, confirmation of their results is subject to the implementation of larger trials in patients. The information of these studies is fully listed in Table 2.       The SpO 2 levels at discharge were significantly higher, the respiratory rate was lower, and the duration of admission was shorter in the case group. There was no significant difference in the mortality rate between the two groups

Artificial Intelligence
We initialized the associations by selecting drugs where their p-values with COVID-19 is small. The sources of the selection of drugs that seems to be effective in treating COVID-19 are articles, clinical expert reports, and medical databases. Selected drugs include: 'Baricitinib' [66,67], 'Tocilizumab' [68,69], 'Immucillin A' [70], 'Favipiravir' [71], 'Remdesivir' [72], 'Hydroxychloroquine' [73], 'Ivermectin' [74], 'Azithromycin' [74,75] The p-NA algorithm calculates combined p-values between each of these drugs and COVID-19. Then, it selects the drug with the lowest combined p-value. After selecting the first drug, the p-NA updates weights between COVID-19 and interface-affected human genes according to the pvalues between those genes and the selected drug. Thus, all of the combined p-values between these drugs and COVID-19 are updated. The p-NA iterates to find the drug with the lowest combined p-value by updating weights. At the end, the p-NA algorithm proposed scenarios, where each scenario contains a drugs recipe, as is the aim of p-NA for making associations cooperate. Tables 3 and 4 show different scenarios proposed by p-NA for the drugs recipe. In Table 3, the threshold for stopping the algorithm is 0.05, whereas 0.01 is the threshold of p-NA in Table 4. The p-NA iterates finding drugs until the combined p-value between each scenario and COVID-19 become less than the threshold. Table 3 shows the proposed scenarios of artificial intelligence in drug combinations at a significance level of 0.05. The first scenario of Table 3 contains two drugs proposed by p-NA, which are as follows: -Therapeutic corticosteroid with a combined p-value of 0.07; -Camostat with a combined p-value of 0.02, on the condition that the therapeutic corticosteroid is selected in this scenario. Table 3. Scenarios for drugs recipe with their combined p-values with threshold of 0.05. Therefore, the combined p-value between the drugs recipe in scenario 1 and COVID-19 is 0.02.
In other words, if only 'Therapeutic Corticosteroid' is used, the significance level is 0.07, and if this drug is combined with 'Camostat' and taken at the same time, the level is significantly reduced to 0.02 and has a better effect on the treatment of COVID-19 patients. The second scenario of Table 3 states that if only 'Remdesivir' is used to treat patients, the significance level is 0.1 and, if it is used concomitantly with 'Azithromycin', the significance level is reduced to 0.05, which again has a better effect than from taking 'Remdesivir' alone. Also, in the third scenario of Table 3, if the 'Interleukin 1 Receptor Antagonist Protein' is used for patients, the level of significance is 0.1 and, if this drug is taken simultaneously with the drug 'Camostat', the level of significance to it reaches 0.02. The fourth scenario in this table suggests a combination of three different drugs so that the use of 'Chloroquine' alone has a significance level of 0.17 and, if 'Favipiravir' is taken simultaneously with this drug, the significance level is 0.08 and, if 'Camostat' should be added to the combination of the other two drugs, the significance level reaches 0.046, which is a much more effective combination than the use of 'Chloroquine' alone for the treatment of patients with COVID-19. Finally, the fifth scenario in Table 3 shows the combination of the three drugs hydroxychloroquine, azithromycin, and camostat, in which, the AI predicts a p-value of this combination with COVID-19 of 0.044.
Based on the results of Table 4, in the combination of drugs for the treatment of patients with COVID-19, the proposed scenarios for the treatment of patients with COVID-19 at a significance level of 0.01 are as follows. In the first scenario of Table 4, when only interleukin 1 receptor antagonist protein is used to treat patients, the significance level is 0.16 and, if used concomitantly with camostat, the significance level is 0.02 and, if azithromycin is added, the significance level is 0.01 and, with the addition of tocilizumab and oseltamivir, a significant level is reached at 0.008 and 0.006, respectively, which is a very effective drug combination in the treatment of patients with COVID-19. Also, in the second scenario of Table 4, if interleukin 1 receptor antagonist protein is used alone, the significance level is 0.16, if camostat is added, the significance level is 0.02, if chloroquine is added, the significance level is 0.01, and, if two drugs, favipiravir and tocilizumab, are used, they reach significant levels of 0.009 and 0.007, respectively, making this combination an effective treatment. Finally, the third scenario of Table 4 shows that, in the case of taking therapeutic corticosteroid alone, a significant level of 0.07 is reached, with the addition of camostat, a significant level of 0.02 is reached, with the addition of oseltamivir, a significant level of 0.017 is reached, and with the addition of remdesivir and tocilizumab, the significance reaches 0.011 and 0.009, respectively, where another effective drug combination is suggested.
The p-values between affected human genes and COVID-19 are shown in Table 5. The second column shows the p-value between COVID-19 and the corresponding affected human genes in the first column of this table. The column with title "Scenario 1" demonstrated the combined p-value between COVID-19 and the drugs recipe in scenario 1 from 2. The combined p-values of other scenarios in 2 are shown as columns in 4. In each scenario of p-NA, the p-values between affected human genes and COVID-19 become close to 1. However, there are some genes in Table 4 where their p-values are still far from 1.  As was known until writing this paper, there is not much information in network meta-analyses for drugs that have small p-values with them.
Whereas Figure 7A shows the p-values between affected human genes and COVID-19, Figure 7B-F demonstrate the p-values between them after implementing scenarios 1 to 5, respectively. In other words, Figure 7B-F show different circular bar plots to show the efficiencies of drugs selected using p-NA by demonstrating the p-values between COVID-19 and human genes after the consumption of selected drugs. Figure 8A-C, on the other hand, indicate p-values between affected human genes and COVID-19 after implementing three new scenarios. Each colored line shows the effectiveness of the corresponding drug in that scenario in raising the p-value. Figure 9 shows the network chart for p-values between human genes and two different drugs, namely 'Therapeutic Corticosteroid' and 'Camostat'. Their p-values with affected human genes are completely different. The network chart shows how two supplement drugs help each other to overcome COVID-19. For example, the dashed green line in this figure shows that the p-value between one of the human genes and COVID-19 is about 0.002. The p-value between that human gene and 'Therapeutic Corticosteroid' is approximately one (green dashed lines), whereas the p-value between that human gene and 'Camostat' is close to zero (yellow dashed lines). implementing scenarios 1 to 3 in 3, respectively. Each colored line shows the effectiveness of the corresponding drug in that scenario in raising the p-value.  Table 5 p-value column, (B) the circular bar plot of Table 5 scenario 2 column, (C) the circular bar plot of Table 5 scenario 1 column, (D) the circular bar plot of scenario 3 of Table 5, (E) the circular bar plot of scenario 4 of Table 5, (F) the circular bar plot of scenario 5 of Table 5.  Table 4, (B) the radar chart of scenario 2 of Table 4, (C) the radar chart of scenario 3 of Table 4. Figure 9 shows the network chart for p-values between human genes and two different drugs, namely 'Therapeutic Corticosteroid' and 'Camostat'. Their p-values with affected human genes are completely different. The network chart shows how two supplement drugs help each other to overcome COVID-19. For example, the dashed green line in this figure shows that the p-value between one of the human genes and COVID-19 is about 0.002. The p-value between that human gene and 'Therapeutic Corticosteroid' is approximately one (green dashed lines), whereas the p-value between that human gene and 'Camostat' is close to zero (yellow dashed lines).

Discussion
COVID-19 is known to be a challenging disease for healthcare systems globally. The disease spreads rapidly within the community and, during a severe clinical course, leads to fatality among a considerable proportion of patients. Since there is no definitive cure for the disease, various treatment strategies were studied in the study, most of which were based on the use of antiviral drugs. However, factors such as different doses of drugs used, the interval between the onset of symptoms and the start of treatment, the severity of the disease in different patients enrolled in different trials, the lack of standard controls,

Discussion
COVID-19 is known to be a challenging disease for healthcare systems globally. The disease spreads rapidly within the community and, during a severe clinical course, leads to fatality among a considerable proportion of patients. Since there is no definitive cure for the disease, various treatment strategies were studied in the study, most of which were based on the use of antiviral drugs. However, factors such as different doses of drugs used, the interval between the onset of symptoms and the start of treatment, the severity of the disease in different patients enrolled in different trials, the lack of standard controls, and the lack of controlled randomized trials due to long contradictory results, were reported in previous studies [76].
One of the methods that can be used to solve the problems caused by heterogeneity and limitations of studies is the use of artificial intelligence techniques. Using advanced machine learning algorithms, the integration and analysis of data related to COVID-19 is conducted with higher speed and accuracy. Accordingly, it is possible to develop effective treatments and identify new treatment approaches for COVID-19 [11].
In this article, using WGM, we implemented the proposed p-NA method for COVID-19 treatment. In this section we, at first, focus on the advantages of WGM over the previously known combining p-values method. Second, we discuss the drawbacks of p-NA. Then, we discuss alternatives for associations in treating COVID-19.
In the present study, the results of 39 interventional studies were reviewed. Our studies showed a high heterogeneity for the effectiveness of different drugs. Evidence suggests that hydroxychloroquine's efficacy in reducing the symptoms and mortality of COVID-19 patients is weak, and cannot be described as an effective treatment for COVID-19 patients. Our findings are consistent with the results of previous studies, so that, in the study of Thibault et al. [77], which was conducted as a systematic review and meta-analysis, it was observed that HCQ alone cannot be effective in treating patients with COVID-19. In addition, this study showed that the simultaneous use of HCQ and azithromycin in the treatment of patients with COVID-19 increases the mortality caused by this disease. Another cohort study in the United States found that the use of HCQ, azithromycin, or both had no positive effect on improving patients and reducing their hospital mortality [78]. However, the results of our study showed that the simultaneous use of these two drugs is more effective than the use of HCQ or azithromycin alone.
Other drugs, including tocilizumab, were also used to treat patients with COVID-19. Previously, some open-label, non-randomized studies have shown that blocking early interleukin-6 receptors helps to reduce the severity of COVID-19. The results of limited RCT studies in this field indicate that this drug alone cannot be considered an effective treatment for COVID-19 [52,79].
The scenarios proposed by artificial intelligence in this study show that the addition of tocilizumab to various drug combinations significantly increases the effectiveness of treatments. This is also true for other drugs, including remdesivir. Thus, previous studies examining the effectiveness of remedial in patients with COVID-19 showed that this drug alone could not be recognized as an appropriate treatment for COVID-19 [31].
Interferon beta, lopinavir-ritonavir, sofosbuvir, calcifediol, and colchicine were other drugs used to treat COVID-19. Despite the positive efficacy of these drugs in studies, since the standard clinical trials with a high number of participants have not been performed, the effectiveness of these drugs cannot be considered with certainty. However, the evidence from artificial intelligence in our study shows that the combination of these drugs can have far better effects.

Advantages of WGM in Combining p-Values
Fisher's test is not the best test available to combine p-values of multiple tests of the same null hypothesis [80]. Fisher's method does not use weights, while different weights can be assigned to each study according to their power [21,81]. In our study, 1 − p-values were viewed as weights of affected human genes. Stouffer's method uses the Z-transform to find the combined p-values. One generalized version uses the weights to each study. However, all of them assume that all of the null hypotheses are true. The Z-transform of 0 and 1 are negative and positive infinity, respectively. The geometric mean is one of the

Drawbacks of the p-NA
Network meta-analysis computes the pvalues between two associations. However, the effectiveness between two associations is not mentioned. For instance, the p-values between COVID-19 and both remdesivir and recombinant interleukin-6 are small in network meta-analysis. However, recombinant interleukin-6 should be decreased for patients with COVID-19. Remdesivir, on the other hand, is a good drug in treating COVID-19. Greedy algorithms find the locally optimal choice, where the solution is found. In most cases, the result is not the optimal solution. For example, in the travelling salesman problem, the city with the shortest distance is always selected, while the selected tour may be the worst tour [75].

Alternatives for Associations
Sequence similarity searching is a method of searching sequence databases by using alignment to a query sequence. By statistically assessing how well the database and query sequences match, one can infer homology and transfer information to the query sequence. The tools can be launched with different form presets using the links. These can be changed on the tool page as well.
Sequence similarity searching is a method of searching for sequence databases by aligning them with a query sequence. By statistically evaluating how closely the database and query sequences match, we can derive homology and transfer information to the query sequence. The tools can be started with various form presets via links. Related diseases: in recent years, there have been three types of coronaviruses: severe acute respiratory syndrome (SARS-CoV), Middle East respiratory syndrome (MERS), and coronavirus disease 2019 (COVID-19). The first one, SARS-CoV, spread over months in 2002 and has symptoms that include a high fever, headache, body aches, and a dry cough. The disease is transmitted through the mouth, nose, or eyes, such as kissing, touching, sharing utensils for eating and drinking, or talking with an infected person. There is no treatment for SARS. Scientists are testing treatments and vaccines. The second disease, MERS, was first reported in 2012 in a pneumonia patient. Symptoms include renal failure and severe acute pneumonia, with often fatal outcomes. Many people with the third one, COVID-19, also have pneumoniaviral infection in lungs. Chest x-rays and blood tests can help to determine the infection. However, preventing pneumonia is always better than treating it. Symptoms: COVID-19 has symptoms such as fever, cough, and shortness of breath (dyspnea). However, other coronaviruses have a sore throat and headache as their other symptoms.

Conclusions
We have proposed a new machine learning algorithm, namely p-NA, to find a combinations of drugs (associations) for COVID-19 treatment. The drugs that are examined in p-NA are the ones where the p-values between them and COVID-19 are smallest in the network meta-analysis. The p-NA, on the other hand, uses interface features (human affected genes in this work) to compute the p-values between the target COVID-19 and associations (drugs). The p-values between drugs and these human genes, and the weights (1 − p-values) between human genes and COVID-19, lead to computing a combined p-value between drugs and COVID-19. Since p-values follow the uniform distribution, the combination method is calculated by a weighted geometric mean. Sorting in ascending order of the combined p-values between COVID-19 and drugs, p-NA selects the first drug. After selecting this drug, the weights between human genes and COVID-19 are updated; consequently, the combined p-values are updated. The process of selecting the drug with the smallest combined p-value and updating the combined p-value iterates until the p-NA becomes less than the threshold. Two different thresholds in this work are suggested as 0.05 and 0.01, i.e., the combined p-values between COVID-19 and drugs. The p-NA implemented with