Prediction of Drug Targets based on In Vitro Activity Profiles Toward Drug Repurposing for Rare Diseases

Abstract

Over 300 million people are suffering from rare diseases, most of which have limited treatment options. Therefore, discovering new treatments for rare diseases is imperative. Drug repurposing, which identifies new uses for approved drugs, is considered one of the viable and risk-managed strategies for disease treatments. To promote the drug repurposing process, we introduced a prediction model to uncover novel relationships between gene targets and chemical compounds. In our previous study, we identified enriched genes for compounds from the Toxicology in the 21st Century program (Tox21) 10K library, to extend that study for enriched gene target prediction, we developed machine learning (ML) models including Support Vector Machine; K-Nearest Neighbors; Random Forest; and extreme gradient boosting (XGBoost), by using Tox21 bioassay screening data. All four models perform well with f1_score over 0.7, and XGBoost has the best performance with four different multi-label prediction embedding algorithms, including Binary Relevance; Label Powerset; Classifier Chain; Multi-Output Classifier. Our study explored a reliable method to predict potential gene targets from in vitro activity profile data toward drug repurposing.

I. Introduction

Thirty million people, which equate to 10% of the U.S. population, are suffering from rare diseases.[1] Only about 5% of these diseases have treatment options. Thus, it is urgent to identify drugs for them to improve their quality of life. Drug repurposing as one of the economized and time-saving methods, is the process of finding alternative use of approved or investigational drugs. One way to promote drug repurposing is based on the scientific fact that a target involved in a certain disease often influences the function of genes that are associated with other pathogenetic mechanisms of rare diseases [2]. Thus, understanding the relationship between compounds and gene targets is crucial for rare disease treatment discovery.

The Tox21 10K compound library activity profiling data is a commonly used resource for studying the interactions between small molecules and genes, cells, and biochemical pathways. The Tox21 10K library, approximately a third of which are approved or investigational drugs, has been screened against 80 in vitro cell-based assays [3]. Large volumes of systematically preprocessed data with high-quality control for toxicology applications makes the Tox21 data an ideal resource for drug repurposing. Our previous studies successfully used the Tox21 bioassay screening data to cluster drugs with similar biological activities and identified enriched genes for each cluster [4]. The results of that study laid out the foundation of supporting novel gene target prediction.

Different artificial intelligence (AI) and ML tools have been widely used for drug safety evaluation offering beneficial information for drug repurposing [5, 6]. Application of modern machine learning approaches such as k-nearest neighbors (KNN) [7], Support Vector Machine (SVM) [8], and extreme gradient boosting (XGBoost) [9], in the drug repurposing exhibits high performance of predictivity.[6, 10] In this study, we developed four different ML models including SVM; KNN; Random Forest; XGBoost incorporating four multi-label classification algorithms including Binary Relevance (BR) [11]; Label Powerset (LP) [12]; Classifier Chain (CC) [13], and Multioutput Classifier (MC) [14] for multiple gene target prediction.

II. Methodology

A. Tox21 Data Set

In this study, we explored Tox21 bioassay screening data to predict novel gene targets for the chemical compounds. The Tox21 library contains 8,971 unique compounds, which have been screened against ~80 bioassays. We applied compound activity scores generated for each bioassay, which are determined by several features of the primary concentration-response curve including potency, efficacy, and quality, as features for our predictive modeling. Compound activity scores from Tox21 data are reported with the curve rank metric, which ranges between −9 to 9 [15]. While positive numbers indicate activation and negative numbers suggest inhibition, the absolute value of each number represents the strength of activation or inhibition. The larger the absolute value of the activity score, the stronger the effects of the compound.

B. Target Gene Selection

In our previous study,[4] gene enrichment analysis was performed on 129 generated clusters, and it results in 737 enriched genes among those clusters. In this study, we selected the top fifteen gene targets associated with the greatest number of compounds, as multi-label predictive targets, which is shown in Table 1.

TABLE 1.

Fifteen selected gene targets

Gene	Count
AR	223
CYP2D6	158
KCNH2	127
HTR2A	117
SI	113
DRD2	101
HRH1	100
CHRM1	98
SLC6A4	97
CYP1A2	97
CYP3A4	96
SIGMAR1	92
HTR2C	88
CHRM3	82
ADRA1A	82
HTR1A	81
ADRA2A	81
HTR2B	80
CYP2C19	79
DRD3	78
ADRA2B	78
ADRA2C	75
CHRM2	74
CHRM4	73
ADRA1D	72
CA1	71
SLC6A2	69
PTGS2	67
PTGS1	66
CA2	66
ACHE	63
ADRA1B	60
CHRM5	58
NR3C1	58
SLC6A3	56
DRD1	56
DRD4	54
CA12	53
CA9	52
ADRB2	51
ADRB1	51
HTR6	50
HTR7	49
ESR1	47
TRPA1	45
HTR1B	44
CYP2C9	44
HTR1D	43
DRD5	38
CA4	37
GABRA1	35
MAOA	34
SCN1A	34
CA14	34
CA7	33
SCN2A	32
SHBG	31
SCN3A	31
GABRB3	30
CA6	30
ESR2	29
ABCB1	28
PPARG	27
GABRG2	26
MAOB	25
OPRM1	25
PGR	25
GABRB2	24
C5	24
HTR1E	24
ADORA3	23
NAPRT	22
TOP2A	22
GABRA3	22
ADRB3	21
KCNJ1	21
GABRA2	21
CYP19A1	21
GLRA1	21
NR1I2	20
LOX	20
GABRA5	20
KCNJ11	19
SLCO1B1	19
SERPINA6	18
CACNA1C	18
F2	18
ALOX5	18
EBP	18
BCHE	17
PPARA	17
CYP2B6	17
ACE	17
TBXAS1	17
KCNA7	17
HRH2	16
NR3C2	16
PDE5A	16
HTR1F	16
GABRB1	16
HTR5A	16
KCNJ6	15
EGFR	15
CHRNA2	15
SLC12A3	15
GABRG3	15
ABCC8	14
KCNK2	14
CHRNA1	14
OPRD1	14
SLC47A1	14
CACNA1H	13
PDE4B	13
CACNA1G	13
HSP90AA1	13
TRPV3	12
SLC22A6	12
AOX1	12
CALM1	12
NISCH	12
TNF	12
CACNA1I	12
AKR1C3	12
ADA	12
PDE4A	11
SLC12A1	11
KCNJ5	11
PDE4D	11
AKR1C1	11
HTR3B	11
ESRRG	11
TRPC5	11
KCNQ1	10
DHFR	10
TRPM3	10
TTR	10
ADORA1	10
PLA2G1B	10
TUBB	10
TYMS	10
OR51E2	10
GLRB	10
HMGCR	10

C. Predictive modeling

We performed multi-label predictive modeling in our study. In total, we performed sixteen modeling experiments, which were categorized based on four different multi-label prediction algorithms: 1) Binary Relevance; 2) Label Powerset; 3) Classifier Chain; and 4) Multioutput Classifier, which is shown in Table 2. The parameters used in these models were set to default if not specified. We performed hyperparameter tuning for all models to determine the optimal values for a given model with 5-fold cross-validation (CV). The summary of optimal parameters is shown in Table 2. With the fine-tuning results, we calculated f1_score to evaluate the performance of all models. Macro f1_score is calculated as follows to determine the performance of modeling algorithms.

TABLE 2.

Models and the associated Parameters

Prediction algorithms	Estimators	Parameters
Binary	KNN	Leaf size=1; n neighbors=33; p=3
Relevance	XGBoot	Colsample_bytree=0.8;max_depth=3; reg_alpha =10; min_child_weight=7; subsample=0.7
	Random Forest	Max_depth=12; min_samples leaf=2; min_samples_split=2; n_estimators=17;
	SVM	C=100; gamma=0.001
Label Powerset	KNN	Leaf_size=1; n_neighbors=31; p=3
	XGBoot	Colsample_bytree=0.7; max_depth=5; reg_alpha =10; min_child_weight=7; subsample=0.6
	Random Forest	Max_depth=9; min_samples_leaf=2; min_samples_split=2; n_estimators=23;
	SVM	C=15; gamma=0.01
Classifier Chain	KNN	Leaf_size=1; n_neighbors=33; p=3
	XGBoot	Colsample_bytree=0.6; max_depth=7; reg_alpha =1; min_child_weight=1; subsample=0.8
	Random Forest	Max_depth=12; min_samples leaf=3; min_samples_split=2; n_estimators=20;
	SVM	C=10; gamma=0.1
Multioutput Classifier	KNN	Leaf_size=1; n_neighbors=25; p=2
	XGBoot	Colsample_bytree=0.8; max_depth=3; reg_alpha =10; min_child_weight=7; subsample=0.7
	Random Forest	Max_depth=12; min_samples_leaf=2; min_samples_split=2; n_estimators=17;
	SVM	C=10; gamma=0.01

$$f1\_\mathit{\text{score}}=\frac{2*\mathit{\text{precision}}*\mathit{\text{recall}}}{\mathit{\text{precision}}+\mathit{\text{recall}}}$$

$$\mathit{\text{Macro}}f1\_\mathit{\text{score}}=\frac{\sum _i^{n}f1\_\mathit{\text{scor}}{e}_i}{n}$$

III. Results

A. Data preparation

Out of 8,971 Tox21 compounds, we included 7,170 compounds with curve rank data across all assays in this study. From 737 unique enriched genes from our previous study, we selected 15 genes with the greatest number of known targeting compounds as endpoints in our models.

B. Predictive results

We tested 16 models in our study. We found that the best model is the binary relevance algorithm with XGBoost (f1_score = 0.757). We also observed that XGBoost and Random Forest have generally better performances (f1_score > 0.7) across the 4 multi-label modeling algorithms. The non-linear SVMs with radial basis function (RBF) kernel with higher computational complexity did not outperform simple KNN models. Finally, comparing the 4 different multi-label predictive modeling strategies, BR and MC have better predictive accuracy (Fig. 1). Since these two experiments were built based on the assumption that all ‘labels’ in the models are independent of each other, we concluded that there might not be interactions among the 15 selected genes.

Fig. 1.

Performance of Models (Figure shows the calculated macro f1_socre of 16 different models. Red bars represent models with macro f1_score lower than 0.7; blue bars represent better performance with macro f1_score greater than 0.7.)

IV. Discussion

While studies were published by using Tox21 data on toxicity prediction [5, 6, 7, 16], we explored Tox21 data for drug target prediction towards drug repurposing based on sixteen developed ML models. In this study, most of our developed models performed well (f1_score > 0.7), and we found that XGBoost is the best estimator compared with the other three estimators. Meanwhile, BR and MC always showed better performance across all estimators.

Due to the nature of data, models for some targets failed to predict, which negatively impacted the overall performance of our models. To further improve the performance of our models, we will exclude those targets with poor performance. In addition, more powerful tools such as deep neural networks will be performed to improve the accuracy of the predictions.

With more accurate drug target prediction models built from compound features, the underlying mechanisms between drugs and genetic rare diseases will be uncovered more easily. And more comprehensive investigation in multi-drug target prediction would be critical to drug repurposing for rare disease treatments.