TYPE Original Research

PUBLISHED 24 October 2022

DOI 10.3389/fimmu.2022.1015409

Modeling and insights into

OPEN ACCESS the structural characteristics
EDITED BY
Betty Diamond,
Feinstein Institute for Medical
of drug-induced
Research, United States

REVIEWED BY
autoimmune diseases
Huiyong Sun,
China Pharmaceutical University,
China Huizhu Guo 1, Peitao Zhang 1, Ruiqiu Zhang 1, Yuqing Hua 1,
Jing Xing,
Michigan State University,
Pei Zhang 1, Xueyan Cui 1, Xin Huang 1 and Xiao Li 1,2*
United States 1
Department of Clinical Pharmacy, The First Affiliated Hospital of Shandong First Medical University
*CORRESPONDENCE & Shandong Provincial Qianfoshan Hospital, Shandong Engineering and Technology Research
Xiao Li Center for Pediatric Drug Development, Shandong Medicine and Health Key Laboratory of Clinical
lixiao1688@163.com; Pharmacy, Jinan, China, 2 Department of Clinical Pharmacy, Shandong Provincial Qianfoshan
x.li@sdu.edu.cn Hospital, Shandong University, Jinan, China

SPECIALTY SECTION
This article was submitted to
Autoimmune and Autoinflammatory
Disorders : Autoimmune Disorders, The incidence and complexity of drug-induced autoimmune diseases (DIAD)
a section of the journal
have been on the rise in recent years, which may lead to serious or fatal
Frontiers in Immunology
consequences. Besides, many environmental and industrial chemicals can also
RECEIVED 09 August 2022
ACCEPTED 11 October 2022
cause DIAD. However, there are few effective approaches to estimate the DIAD
PUBLISHED 24 October 2022 potential of drugs and other chemicals currently, and the structural
CITATION
characteristics and mechanism of action of DIAD compounds have not been
Guo H, Zhang P, Zhang R, Hua Y, clarified. In this study, we developed the in silico models for chemical DIAD
Zhang P, Cui X, Huang X and Li X
prediction and investigated the structural characteristics of DIAD chemicals
(2022) Modeling and insights into the
structural characteristics of drug- based on the reliable drug data on human autoimmune diseases. We collected
induced autoimmune diseases. 148 medications which were reported can cause DIAD clinically and 450
Front. Immunol. 13:1015409.
doi: 10.3389/fimmu.2022.1015409
medications that clearly do not cause DIAD. Several different machine
learning algorithms and molecular fingerprints were combined to develop
COPYRIGHT
© 2022 Guo, Zhang, Zhang, Hua, the in silico models. The best performed model provided the good overall
Zhang, Cui, Huang and Li. This is an accuracy on validation set with 76.26%. The model was made freely available
open-access article distributed under
the terms of the Creative Commons
on the website http://diad.sapredictor.cn/. To further investigate the
Attribution License (CC BY). The use, differences in structural characteristics between DIAD chemicals and non-
distribution or reproduction in other DIAD chemicals, several key physicochemical properties were analyzed. The
forums is permitted, provided the
original author(s) and the copyright results showed that AlogP, molecular polar surface area (MPSA), and the
owner(s) are credited and that the number of hydrogen bond donors (nHDon) were significantly different
original publication in this journal is
cited, in accordance with accepted
between the DIAD and non-DIAD structures. They may be related to the
academic practice. No use, DIAD toxicity of chemicals. In addition, 14 structural alerts (SA) for DIAD
distribution or reproduction is toxicity were detected from predefined substructures. The SAs may be
permitted which does not comply with
these terms. helpful to explain the mechanism of action of drug induced autoimmune
disease, and can used to identify the chemicals with potential DIAD toxicity.

Frontiers in Immunology 01 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

The structural alerts have been integrated in a structural alert-based web server
SApredictor (http://www.sapredictor.cn). We hope the results could provide
useful information for the recognition of DIAD chemicals and the insights of
structural characteristics for chemical DIAD toxicity.

KEYWORDS

drug-induced autoimmune diseases, computational toxicology, machine learning,

molecular fingerprinting, structural alert

Introduction The mechanisms of DIAD have not been fully clarified. Most
DIAD chemicals are small molecules and have no
Autoimmune diseases refer to problems where the immune immunogenicity by themselves, but these small molecule
system attacks healthy cells in the body by mistake (1). It was substances or their metabolites are able bind to carrier
reported that there are about 20 million autoimmune disease proteins and become immunogens (8). These hemicantized
patients in the United States, accounting for about 7% to 9% of drugs become target antigens and induce an immune response
the total population (2). Meanwhile, the incidence of to themselves, leading to the production of autoantibodies.
autoimmune diseases in industrialized countries is also However, recent studies have found that most DIAD drugs do
increasing in recent years with the continuous deterioration of not induce specific T cell production, but induce autoimmune
the environment (3). Up to now, more than 100 types of response, so there may be other mechanisms for this process.
autoimmune diseases have been discovered, but the treatment The neutrophil extracellular traps (NETs) (9, 10) induced by
of autoimmune diseases can only repair the damage already drugs may cause or promote the occurrence of some
caused as the main direction. Most patients need long-term or autoimmune diseases. Therefore, the formation of NETs may
even lifelong medication, resulting in huge medical economic also be an important cause of drug-induced autoimmune disease
burden. It cost more than 100 billion US dollars every year in the (11). More recently, Li, et al. reported that ferroptosis is a major
USA healthcare system (4). More importantly, many patients’ factor in neutropenia and systemic autoimmune disease (12).
conditions are dangerous, seriously affecting the quality of life, Hence, drug-induced ferroptosis may be another possible
and even fatal. pathway for DIAD. Many of the DIAD drugs are used for
The development of autoimmune diseases requires both autoimmune diseases treatment themselves. These drugs may
genetic predisposition and environmental factors to jointly be immunostimulatory, and can act in an immunomodulatory
trigger immune pathways, which gradually develop and manner under different genetic and environmental
eventually lead to tissue destruction (5). As one of the conditions (8).
environmental factors, medications and industrial chemicals Until now, there is few effective approach to estimate the
also have been reported can interfere with human immune DIAD potential of drugs and other chemicals (4).
system and induce autoimmune diseases. For instance, drug- Computational toxicology is a structure-based, application-
induced lupus accounts for about 10% of all systemic lupus cases related management and analysis of experimental data from
in the USA (6). Besides, about 12-17% of autoimmune hepatitis toxicological tests that can provide viable mechanistic
cases were believed to be induced by clinical drugs (4). The explanations for the toxicity of compounds (13). This tool is
incidence and complexity of drug-induced autoimmune diseases particularly important in designing safe drugs and assessing
have been on the rise in recent years. Since sulfadiazine was first environmental risks (14–24). Using computational toxicology
reported to cause lupus-like symptoms in 1945, more than 100 tools to develop in silico models for chemical DIAD toxicity and
drugs have been found to cause drug-induced autoimmune analyze the structural characteristics of DIAD drugs not only
diseases (DIAD). As a special type B drug reaction, DIAD is helps to estimate the potential DIAD toxicity of compounds, but
unpredictable, with an incubation period of months or even also helps to explore the structural basis of chemical
years, sometimes leading to serious or fatal consequences. DIAD toxicity.
Compared with primary autoimmune diseases, DIAD has In the present study, we aim to develop the machine learning
more complex clinical manifestations, with significant models for chemical DIAD toxicity and investigate the structural
differences in epidemiology and pathology (7). characteristics of DIAD chemicals.

Frontiers in Immunology 02 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

Materials and methods The molecular description was implemented with several
different molecular fingerprints packages, which have been
Collection and preparation of DIAD and widely used in toxicity prediction of drugs and environmental
non-DIAD drugs chemicals. In this study, we used seven common fingerprints
packages, including the Estate fingerprint (Estate, 79 bits), CDK
Only approved small molecule drugs data related to DIAD fingerprint (FP, 1024 bits), CDK extended fingerprint
toxicity were included in this study. The DIAD drugs were (ExtendFP, 1024 bits), Klekota-Roth fingerprint (KRFP, 4860
extracted from two different sources: (1) the drugs with bits), MACCS keys (MACCS, 166 bits), PubChem fingerprint
autoimmune diseases associated side effects extracted from (PubChem, 881 bits), and Substructure fingerprint (SubFP, 307
the Side Effect Resource (SIDER) database (25); (2) positive bits). All the fingerprints were calculated with PaDEL
drugs for DIAD reported in the literature (4). SIDER is a Descriptor (35).
resource of adverse drug reaction (ADR), which contains the
information on marketed medicines and their recorded ADRs
on human. We retrieved the entire SIDER database, and Assessment of model performance
extracted the ADRs related to autoimmune diseases with
frequency ≥ 0.1%. The corresponding structures were The models were both validated with 5-fold cross validation
obtained from the PubChem compound database (26). The and a validation set. Several statistical parameters were
negative drugs for DIAD were all extracted from Wu’s work calculated for the assessment of model performance, including
(4). The structures were prepared by: (1) only keeping the main prediction accuracy (ACC), sensitivity (SE), specificity (SP) and
ingredients in mixtures; (2) excluding the inorganic and the Matthew’s correlation coefficient (MCC) (36), as shown with
organometallic compounds; (3) converting the salts into the Eqs (1-4).
parent forms; (4) removing the duplicate substances. The data TP + TN
standardization was performed on Online Chemical Database ACC = (1)
TP + FN + TN + FP
and Modeling Environment (OCHEM) platform (27), which is
a user friendly web-based platform for data exploring and TP
modeling. The details for structure preparation can be seen in SE = (2)
TP + FN
supporting information.
TN
SP = (3)
TN + FP
Model building for chemical TP*TN − FP*FN
DIAD toxicity MCC = pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (4)
ðTP + FPÞðTP + FN ÞðTN + FPÞðTN + FN Þ

As a specific artificial intelligence method, machine learning Where TP represented true positives, FP represented false
was always used for the model building which can access data positives, TN represented true negatives, and TN represented
and use data for automated learning (28). In this study, five false negatives.
commonly used machine learning methods were used for the In addition, the receiver operating characteristic (ROC)
model development, including Support Vector Machines (SVM) curve was plotted, and the values of the area under the ROC
(29), Naive Bayes (NB) (30), K-nearest Neighbor (kNN) (31), curves (AUC) were also computed.
Decision Tree (DT) (32) and Random Forest (RF) (33). These
methods have been extensively used in computational toxicity
studies due to the high effective and robust. The detailed Analysis of molecular properties for the
descriptions for the algorithms can be learned in the DIAD and non-DIAD drugs
corresponding literature. Herein, the SVM algorithm was
performed with the LIBSVM (LIBSVM 3.16 package) (34), and The molecular properties of compounds can play key roles
the parameters for Gaussian radial basis function (RBF) kernel in biological and toxicological activities. Eight important
were optimized with a grid search method based on 5-fold cross- molecular properties were calculated with PaDEL-Descriptor
validation. The other algorisms were implemented in the data package. These properties were molecular weight (MW),
mining tool Orange (version 2.7, freely available at https:// molecular polar surface area (MPSA), AlogP, molecular
orange.biolab.si/orange2/). For kNN, the parameter k was also solubility (LogS), the number of hydrogen bond acceptors
optimized based on 5-fold cross-validation. The parameters for (nHAcc) and donors (nHDon), the number of rotatable bonds
C4.5 DT, RF and NB algorithms were optimized with the default (nRotB) and the number of aromatic rings (nAR). The MW and
setting in the Orange toolbox. MPSA values can reflect the size and complexity of molecules to

Frontiers in Immunology 03 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

a certain extent. AlogP and LogS are usually used to represent chemical space. PCA can transform data into lower dimensions
the chemical lipophilicity and solubility in water. The nHAcc from high-dimensional data, and meanwhile the trends and
and nHDon values represent the hydrogen bonding ability of patterns can be retained as possible (49). Herein, the first two
compounds, which also can play an important role for principal components (PC) with cumulative proportion
chemical activities. 79.08% were kept to represent the chemical space. As shown
Because of disobeying the normal distribution, the data were in Figure 1, the results suggested that the chemical spaces of
expressed by the median and interquartile spacing and the training and validation sets were similar.
comparison between groups adopted Wilcoxon rank sum test
in this study. The p value < 0.05 was considered with
statistical significance.
Machine learning models

There were 35 different classification models developed

Analysis of structural alerts for chemical using the different machine learning algorithms combined
DIAD toxicity with fingerprint packages. The optimized parameters for SVM
and kNN models can be seen in Table S2. Considering the large
Structural alert (SA) was the toxophore (usually a specific
difference in the number of DIAD and non-DIAD drugs in this
substructure or fragment) which can lead to a particular toxicity
study, the ACC and MCC values were paid special attention
endpoint. It has been widely used for toxicity research for many
when evaluating the performance of the models, since MCC can
different toxicity endpoints (37–44). In this study, the SAs for
be influenced much less by imbalanced data. Most of the models
DIAD toxicity were detected by calculating the f-score (45) and
showed good performance on the 5-fold cross validation, as
frequency ratio of each fragment from KRFP fingerprint. The f-
shown in Table 2. The ACC values ranged from 60.42% to
score is a simple feature selection technique, which can measure
77.50%, and the MCC values ranged from 0.21 to 0.55. The
the discrimination of two sets. The larger f-score always
model developed with SVM method and MACCS keys provided
suggested the feature is more discriminative (45, 46). The
the best performance with the total accuracy 77.50%, SE value
positive rate (PR) of a substructure was defined as below:
76.52%, SP value 78.40%, AUC value 0.86 and MCC value 0.55.
Nfragment _ positive Besides, five models (ExtendFP_SVM, KRFP_SVM,
PR = (5) SubFP_SVM, MACCS_kNN, and ExtendFP_kNN) also
Nfragment
showed good predictive results on 5-fold cross validation with
Where Nfragment_positive was the number of DIAD compounds ACC > 75.00% and MCC > 0.50.
containing the substructure, and Nfragment was the total number Furthermore, the validation set was used to assess the
of all the compounds containing the substructure. generalization and robustness of the six models with top
performance on internal cross validation. Since the validation
set was completely independent from the training set, it can be
Results and discussion used to validate the predictive ability of models objectively. The
performances of models on validation set were shown in Table 3
Data set analysis and the ROC curves were shown in Figure 2. Most of the models,
except MACCS_kNN, performed well with the ACC values
After filtering and preparation, there were 598 organic higher than 75% and the MCC values over 0.25. The
compounds, including 148 DIAD drugs and 450 non-DIAD MACCS_SVM model also achieved good prediction accuracy
drugs, extracted from the SIDER database and literature. The of 75.98% on validation set and best AUC value of 0.84, the value
DIAD compounds were randomly divided into a training set of MCC was 0.33, and the values of SE and SP were 75.76% and
and a validation set with 80%:20%. Since the non-DIAD drugs 76.00%, respectively. In Wu’s work, the machine learning model
are much more than the DIAD drugs in number, and the based on structural alerts and daily dose as input features
imbalance may cause bias in model development, the non- showed a balanced accuracy of 69%, MCC of 0.47, and AUC
DIAD were randomly divided into training and validation set
with 25%:75%. As shown in Table 1, the training set contained
240 structures (115 DIAD drugs and 125 non-DIAD drugs) TABLE 1 Detailed statistical number of drugs in the data set.
and the validation set contained 358 structures (33 DIAD drugs
DIAD structures Non-DIAD structures Total
and 325 non-DIAD drugs). The structure information of the
drugs can be seen in Table S1. The structural diversity of Training set 115 125 240
chemical compounds is important for the global models (47). Validation set 33 325 358
The principal component analysis (PCA) (48) was performed Total 148 450 598
based on the eight physical-chemical properties to generate the

Frontiers in Immunology 04 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

FIGURE 1
Chemical space defined by the first two principal components of physical-chemical descriptors. Red squares stand for the training set, blue
circles stand for the validation set.

of 70% on the test set. Our model covered more compounds and surface area, and there may be no significant correlation between
showed better predictive ability. the chemical DIAD toxicity and structure size.
The analysis for chemical hydrogen bonding ability (nHAcc
and nHDon) suggested that nHDon may be obviously associated
with DIAD toxicity, but the nHAcc was not. The median nHDon
The comparisons of molecular properties of DIAD group was 2 (1, 3), and that of non-DIAD group was 1
between DIAD and non-DIAD chemicals (1, 2), with p < 0.01. The difference is not significant (p=0.09) in
nHAcc between the groups.
In this study, we compared the distributions of several The DIAD toxicity was also not obviously associated with
important molecular properties between DIAD and non-DIAD nRotB and nAR, since the differences between DIAD and non-
structures, as shown in Figure 3 and Table 4. The results DIAD structures were not significant (p = 0.53 for nRotB, p =
indicated that several properties were significantly different 0.10 for nAR).
between DIAD and non-DIAD groups, including AlogP, In fact, the individual chemical descriptors are not sufficient
nHDon, and MPSA. The lipophilicity of compounds is always to fully explain the mechanism of DIAD toxicity, since DIAD
represented with AlogP property. The median AlogP of DIAD toxicity is a very complex endpoint. But we think the results of
group was 1.92 (-0.25, 3.75), which was significantly lower than the study could provide useful information for a further
that of non-DIAD group with 2.60 (0.89, 3.99), with p = 0.03. It understanding of DIAD toxicity.
indicated that lipophilicity should be a molecular property
associated with chemical DIAD toxicity. Meanwhile, there may
be no association between the molecular solubility in water with Structural alerts responsible for
the DIAD toxicity, since there was no significant difference in DIAD toxicity
logS between DIAD and non-DIAD groups (p = 0.44).
The median MPSA of DIAD group (95.08 (58.29, 120.70)) In this study, only the fragments existed in ≥ 6 structures
was larger than that of non-DIAD group (78.29 (49.77, 110.51)), were kept for the structural alert detection. We identified the
with p = 0.02, while the median MW was not significantly privileged substructures which presented much more frequently
different between DIAD and non-DIAD groups (p = 0.71). The in DIAD structures than in non-DIAD structures, with f-score ≥
results suggested that DIAD chemicals may have larger polar 0.018 and positive rate (PR) ≥ 0.75. Finally, we obtained 14

Frontiers in Immunology 05 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

TABLE 2 Performances of models on 5-fold cross-validation.

Model ACC (%) SE (%) SP (%) MCC AUC

Estate_kNN 70.83 77.39 64.80 0.42 0.77

Estate_SVM 72.08 67.83 76.00 0.44 0.79
Estate_RF 67.08 52.17 80.80 0.35 0.76
Estate_NB 65.42 66.09 64.80 0.31 0.65
Estate_CT 67.92 69.57 66.40 0.36 0.74
ExtendFP_kNN 75.42 72.17 78.40 0.51 0.82
ExtendFP_SVM 75.83 70.43 80.80 0.52 0.82
ExtendFP_RF 69.17 58.26 79.20 0.38 0.76
ExtendFP_NB 66.67 62.61 70.40 0.33 0.70
ExtendFP_CT 60.42 65.22 56.00 0.21 0.61
FP_kNN 72.92 71.3 74.40 0.46 0.81
FP_SVM 72.08 70.43 73.60 0.44 0.77
FP_RF 66.25 59.13 72.80 0.32 0.75
FP_NB 64.58 62.61 66.40 0.29 0.68
FP_CT 60.42 62.61 58.40 0.21 0.62
MACCS_kNN 77.50 80.87 74.40 0.55 0.78
MACCS_SVM 77.50 76.52 78.40 0.55 0.86
MACCS_RF 70.83 61.74 79.20 0.42 0.78
MACCS_NB 63.33 66.96 60.00 0.27 0.71
MACCS_CT 66.25 71.3 61.60 0.33 0.65
Pubchem_kNN 70.83 75.65 66.40 0.42 0.75
Pubchem_SVM 74.58 73.91 75.20 0.49 0.80
Pubchem_RF 65.42 59.13 71.20 0.31 0.75
Pubchem_NB 64.58 65.22 64.00 0.29 0.69
Pubchem_CT 67.08 66.09 68.00 0.34 0.68
SubFP_kNN 69.58 77.39 62.40 0.40 0.77
SubFP_SVM 75.42 70.43 80.00 0.51 0.78
SubFP_RF 62.08 50.43 72.80 0.24 0.71
SubFP_NB 63.75 67.83 60.00 0.28 0.69
SubFP_CT 68.75 71.3 66.40 0.38 0.69
KRFP_kNN 73.75 71.3 76.00 0.47 0.85
KRFP_SVM 77.50 73.04 81.60 0.55 0.85
KRFP_RF 62.50 25.22 96.80 0.32 0.77
KRFP_NB 67.08 65.22 68.80 0.34 0.70
KRFP_CT 70.83 73.91 68.00 0.42 0.71

TABLE 3 Performances of best performed models on validation set.

Model ACC (%) SE (%) SP (%) MCC AUC

MACCS_SVM 76.26 75.76 76.31 0.33 0.84

ExtendFP_SVM 79.05 57.58 81.23 0.27 0.78
KRFP_SVM 78.77 72.73 79.38 0.35 0.80
SubFP_SVM 76.54 66.67 77.54 0.29 0.81
MACCS_kNN 68.16 84.85 66.46 0.31 0.76
ExtendFP_ kNN 76.26 60.61 77.85 0.25 0.78

Frontiers in Immunology 06 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

FIGURE 2
ROC curve of models on validation set. Each color line represents a model.

representative fragments for DIAD toxicity. More structural The chemical structures used in this study only contained
alerts responsible for DIAD were proposed in this study than clinical drugs already approved on the market, and the limitation
Wu’s work. Six substructures appeared in DIAD chemicals only, of chemical space may hinder the generalization ability of the
which covered 32 DIAD drugs. All the privileged substructures structural alerts. Nevertheless, these privileged substructures
were listed in Table 5. found in this study can provide the alert help, to a certain
Oxidative stress is common in many autoimmune diseases extent, for the early assessment and mechanism of action of
and is accompanied by overproduction of reactive oxygen species DIAD toxicity.
(ROS) and reactive nitrogen (RNS). The role of oxidative stress in
autoimmune diseases is complex and unclear. Smallwood, et al.
provided insights on the pathophysiological events of oxidative Availability of machine learning models
stress in autoimmune rheumatic diseases (50). The role of ROS and structural alerts
and RNS in the occurrence, detection and treatment of
autoimmune diseases was summarized. In the present study, We made the best performed model developed with SVM
most of the structural alerts (No.1, No.2, No.4, No.5, No.6, method and MACCS keys available at a webserver named
No.7, No.9, No.10, No.11, No.12, and No.14) have the potential DIADpredictor, which can be freely accessed via http://diad.
to produce ROS or RNS (51–56), which we infer may be related to sapredictor.cn/. As shown in Figure 4, users can upload a.smi file
their role in inducing autoimmune diseases. Among them, the or print the SMILES formula to predict whether the chemicals
No.1 fragments has also been identified as an alert for have DIAD toxicity freely.
nephrotoxicity in our previous study. Interestingly, Hultqvist, The structural alerts responsible for DIAD toxicity have been
et al. reported the protective role of ROS in autoimmune disease integrated into SApredictor (http://www.sapredictor.cn/) (42),
(57), just as many of the DIAD drugs we collected are themselves which is an expert system for screening chemicals against
used to treat autoimmune diseases. Phenothiazine was structural alerts. Users can evaluate the DIAD potential for
demonstrated can induce the increase in thyroid autoantigens query structures, and find the specific structural alerts which
and costimulatory molecules on thyroid cells, which may be a leading to DIAD toxicity intuitively. It should be noted
pathophysiological mechanism for drug-induced autoimmunity that the web servers are not suitable for inorganics and
(58). In our study, we found phenothiazine (No.3) only presented organometallic compounds, since they were excluded from the
in DIAD positive structures (8 drugs). modeling dataset.

Frontiers in Immunology 07 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

FIGURE 3
Distributions of the molecular properties for DIAD and non-DIAD chemicals.

Conclusions The model developed with SVM method and MACCS keys
performed best on validations. We made it available freely at
In summary, we developed the machine learning models for http://diad.sapredictor.cn. The analysis of molecular properties for
DIAD toxicity based on the DIAD data of clinical medications. DIAD and non-DIAD compounds indicated that AlogP,

Frontiers in Immunology 08 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

TABLE 4 Distributions of the molecular properties for DIAD and non-DIAD chemicals.

Molecular properties M (P25, P75) p values

DIAD chemicals non-DIAD chemicals

ALogP 1.92 (-0.25, 3.75) 2.60 (0.89, 3.99) 0.03

logS -4.15 (-5.37, -2.86) -4.36 (-5.98, -2.74) 0.44
MW 354.92 (259.60, 429.03) 353.37 (272.14, 441.53) 0.71
MPSA 95.08 (58.29, 120.70) 78.29 (49.77, 110.51) 0.02
nHAcc 5 (3, 7) 4 (3, 6) 0.09
nHDon 2 (1, 3) 1 (1, 2) <0.01
nRotB 5 (2, 8) 5 (3, 8) 0.53
nAR 1 (1, 2) 1 (0, 2) 0.10

TABLE 5 Structural alerts responsible for chemical DIAD toxicity detected from KRFP fragments.

No. General structure Num _P Num _N f-score PR

1 14 0 0.065 1.00

2 8 0 0.036 1.00

3 8 0 0.036 1.00

4 6 0 0.026 1.00

5 12 1 0.051 0.92

6 17 2 0.072 0.89

7 6 1 0.022 0.86

8 20 5 0.075 0.80

(Continued)

Frontiers in Immunology 09 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

TABLE 5 Continued

No. General structure Num _P Num _N f-score PR

9 15 4 0.053 0.79

10 7 2 0.023 0.78

11 7 2 0.023 0.78

12 9 3 0.029 0.75

13 6 2 0.019 0.75

14 6 2 0.019 0.75

FIGURE 4
Main page of DIAD predictor web server. From this page, users can submit the query structures.

Frontiers in Immunology 10 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

molecular polar surface area (MPSA), and the number of project of Clinical Toxicology of Chinese Society of
hydrogen bond donors (nHDon) may be obviously associated Toxicology (CST2020CT104).
with chemical DIAD toxicity. In addition, the structural alerts
responsible for chemical DIAD toxicity were detected from
defined fragments, and made available on SApredictor (www. Acknowledgments
sapredictor.cn). The computational models and the structural
features could provide useful information and understanding for We would like to thank the staff at the Center for Big Data
DIAD toxicity in drug and chemical hazard assessment. Research in Health and Medicine, The First Affiliated Hospital
of Shandong First Medical University & Shandong Provincial
Qianfoshan Hospital, for their valuable contribution. The
Data availability statement authors also gratefully acknowledge the encouragement and
support from Miss Chaoyue Yang and Miss Liying Zhao.
The original contributions presented in the study are
included in the article/Supplementary Material. Further
inquiries can be directed to the corresponding author. Conflict of interest
The authors declare that the research was conducted in the
Author contributions absence of any commercial or financial relationships that could
be construed as a potential conflict of interest.
HG: Data curation, Methodology, Software, Investigation,
Writing - original draft. PTZ: Data curation, Methodology,
Software, Investigation. RZ: Data curation, Methodology, Software, Publisher’s note
Investigation, Writing - original draft. YH: Data curation,
Methodology, Validation, Visualization, Writing - review and All claims expressed in this article are solely those of the
editing. PZ: Methodology, Validation, Writing - review and editing. authors and do not necessarily represent those of their affiliated
XC: Methodology, Validation, Writing - review and editing. XH: organizations, or those of the publisher, the editors and the
Methodology, Validation, Writing - review and editing. XL: reviewers. Any product that may be evaluated in this article, or
Conceptualization, Project administration, Funding acquisition, claim that may be made by its manufacturer, is not guaranteed
Writing - review and editing. All authors contributed to the article or endorsed by the publisher.
and approved the submitted version.

Supplementary material
Funding
The Supplementary Material for this article can be found
This work was supported by the National Natural Science online at: https://www.frontiersin.org/articles/10.3389/
Foundation of China (grant 81803433) and the Special Research fimmu.2022.1015409/full#supplementary-material

References
1. Davidson A, Diamond B. Autoimmune diseases. N Engl J Med (2001) 345 6. Vedove CD, Del Giglio M, Schena D, Girolomoni G. Drug-induced lupus
(5):340–50. doi: 10.1056/nejm200108023450506 erythematosus. Arch Dermatol Res (2009) 301(1):99–105. doi: 10.1007/s00403-008-
2. Cooper GS, Bynum MLK, Somers EC. Recent insights in the epidemiology of 0895-5
autoimmune diseases: Improved prevalence estimates and understanding of 7. Pé rez-De-Lis M, Retamozo S, Flores-Chá vez A, Kostov B, Perez-Alvarez R,
clustering of diseases. J Autoimmun (2009) 33(3):197–207. doi: 10.1016/ Brito-Zeró n P, et al. Autoimmune diseases induced by biological agents. A review
j.jaut.2009.09.008 of 12,731 cases (BIOGEAS registry). Expert Opin Drug Saf (2017) 16(11):1255–71.
3. Rose NR. Prediction and prevention of autoimmune disease in the 21st doi: 10.1080/14740338.2017.1372421
century: A review and preview. Am J Epidemiol (2016) 183(5):403–6. doi: 10.1093/ 8. Chang C, Gershwin ME. Drugs and autoimmunity – a contemporary review
aje/kwv292 and mechanistic approach. J Autoimmun (2010) 34(3):J266–75. doi: 10.1016/
4. Wu Y, Zhu J, Fu P, Tong W, Hong H, Chen M. Machine learning for j.jaut.2009.11.012
predicting risk of drug-induced autoimmune diseases by structural alerts and daily 9. Brinkmann V, Reichard U, Goosmann C, Fauler B, Uhlemann Y, Weiss DS,
dose. Int J Environ Res Public Health (2021) 18(13):7139. doi: 10.3390/ et al. Neutrophil extracellular traps kill bacteria. Science (2004) 303(5663):1532–5.
ijerph18137139 doi: 10.1126/science.1092385
5. Wang L, Wang F-S, Gershwin ME. Human autoimmune diseases: A 10. Thiam HR, Wong SL, Qiu R, Kittisopikul M, Vahabikashi A, Goldman AE,
comprehensive update. J Internal Med (2015) 278(4):369–95. doi: 10.1111/joim.12395 et al. NETosis proceeds by cytoskeleton and endomembrane disassembly and

Frontiers in Immunology 11 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

PAD4-mediated chromatin decondensation and nuclear envelope rupture. Proc 33. Svetnik V, Liaw A, Tong C, Culberson JC, Sheridan RP, Feuston BP.
Natl Acad Sci (2020) 117(13):7326–37. doi: 10.1073/pnas.1909546117 Random forest: A classification and regression tool for compound classification
11. Irizarry-Caro JA, Carmona-Rivera C, Schwartz DM, Khaznadar SS, Kaplan and QSAR modeling. J Chem Inf Comput Sci (2003) 43(6):1947–58. doi: 10.1021/
MJ, Grayson PC. Brief report: Drugs implicated in systemic autoimmunity ci034160g
modulate neutrophil extracellular trap formation. Arthritis Rheumatol (2018) 70 34. Chang C-C, Lin C-J. LIBSVM: A library for support vector machines. ACM
(3):468–74. doi: 10.1002/art.40372 Trans Intell Syst Technol (2011) 2(3):27. doi: 10.1145/1961189.1961199
12. Li P, Jiang M, Li K, Li H, Zhou Y, Xiao X, et al. Glutathione peroxidase 4– 35. Yap CW. PaDEL-descriptor: An open source software to calculate molecular
regulated neutrophil ferroptosis induces systemic autoimmunity. Nat Immunol descriptors and fingerprints. J Comput Chem (2011) 32(7):1466–74. doi: 10.1002/
(2021) 22(9):1107–17. doi: 10.1038/s41590-021-00993-3 jcc.21707
13. Kleinstreuer NC, Tong W, Tetko IV. Computational toxicology. Chem Res 36. Chicco D, Jurman G. The advantages of the matthews correlation coefficient
Toxicol (2020) 33(3):687–8. doi: 10.1021/acs.chemrestox.0c00070 (MCC) over F1 score and accuracy in binary classification evaluation. BMC
14. Li X, Zhang Y, Chen H, Li H, Zhao Y. Insights into the molecular basis of the Genomics (2020) 21(1):6. doi: 10.1186/s12864-019-6413-7
acute contact toxicity of diverse organic chemicals in the honey bee. J Chem Inf 37. Sushko I, Salmina E, Potemkin VA, Poda G, Tetko IV. ToxAlerts: A web server
Modeling (2017) 57(12):2948–57. doi: 10.1021/acs.jcim.7b00476 of structural alerts for toxic chemicals and compounds with potential adverse
15. Cui X, Liu J, Zhang J, Wu Q, Li X. ). in silico prediction of drug-induced reactions. J Chem Inf Modeling (2012) 52(8):2310–6. doi: 10.1021/ci300245q
rhabdomyolysis with machine-learning models and structural alerts. J Appl Toxicol 38. Yang H, Sun L, Li W, Liu G, Tang Y. In silico prediction of chemical toxicity
(2019) 39(8):1224–32. doi: 10.1002/jat.3808 for drug design using machine learning methods and structural alerts. Front Chem
16. Cordero JA, He K, Janya K, Echigo S, Itoh S. Predicting formation of (2018) 6:30. doi: 10.3389/fchem.2018.00030
haloacetic acids by chlorination of organic compounds using machine-learning- 39. Baderna D, Gadaleta D, Lostaglio E, Selvestrel G, Raitano G, Golbamaki A,
assisted quantitative structure-activity relationships. J Hazardous Materials (2021) et al. New in silico models to predict in vitro micronucleus induction as marker of
408:124466. doi: 10.1016/j.jhazmat.2020.124466 genotoxicity. J Hazardous Materials (2020) 385:121638. doi: 10.1016/
17. Cui X, Yang R, Li S, Liu J, Wu Q, Li X. Modeling and insights into molecular j.jhazmat.2019.121638
basis of low molecular weight respiratory sensitizers. Mol Diversity (2021) 25 40. Yang H, Lou C, Li W, Liu G, Tang Y. Computational approaches to identify
(2):847–59. doi: 10.1007/s11030-020-10069-3 structural alerts and their applications in environmental toxicology and drug
18. Hua Y, Shi Y, Cui X, Li X. In silico prediction of chemical-induced discovery. Chem Res Toxicol (2020) 33(6):1312–22. doi: 10.1021/
hematotoxicity with machine learning and deep learning methods. Mol Diversity acs.chemrestox.0c00006
(2021) 25(3):1585–96. doi: 10.1007/s11030-021-10255-x 41. Zhang H, Mao J, Qi H-Z, Xie H-Z, Shen C, Liu C-T, et al. Developing novel
19. Huang X, Tang F, Hua Y, Li X. In silico prediction of drug-induced computational prediction models for assessing chemical-induced neurotoxicity
ototoxicity using machine learning and deep learning methods. Chem Biol Drug using naïve bayes classifier technique. Food Chem Toxicol (2020) 143:111513.
Design (2021) 98(2):248–57. doi: 10.1111/cbdd.13894 doi: 10.1016/j.fct.2020.111513

20. Roostaei J, Colley S, Mulhern R, May AA, Gibson JM. Predicting the risk of 42. Hua Y, Cui X, Liu B, Shi Y, Guo H, Zhang R, et al. SApredictor: An expert
GenX contamination in private well water using a machine-learned Bayesian system for screening chemicals against structural alerts. Front Chem (2022)
network model. J Hazardous Materials (2021) 411:125075. doi: 10.1016/ 10:916614. doi: 10.3389/fchem.2022.916614
j.jhazmat.2021.125075 43. Jia X, Wen X, Russo DP, Aleksunes LM, Zhu H. Mechanism-driven
21. Bernardes RC, Botina LL, da Silva FP, Fernandes KM, Lima MAP, Martins modeling of chemical hepatotoxicity using structural alerts and an in vitro
GF. Toxicological assessment of agrochemicals on bees using machine learning screening assay. J Hazardous Materials (2022) 436:129193. doi: 10.1016/
tools. J Hazardous Materials (2022) 424:127344. doi: 10.1016/j.jhazmat. j.jhazmat.2022.129193
2021.127344 44. Wang C-C, Liang Y-C, Wang S-S, Lin P, Tung C-W. A machine learning-
22. Shi Y, Hua Y, Wang B, Zhang R, Li X. Prediction and insights into the driven approach for prioritizing food contact chemicals of carcinogenic concern
structural basis of drug induced nephrotoxicity. Front Pharmacol (2022) based on complementary in silico methods. Food Chem Toxicol (2022) 160:112802.
12:793332. doi: 10.3389/fphar.2021.793332 doi: 10.1016/j.fct.2021.112802

23. Tang W, Liu W, Wang Z, Hong H, Chen J. Machine learning models on 45. Song Q, Jiang H, Liu J. Feature selection based on FDA and f-score for
chemical inhibitors of mitochondrial electron transport chain. J Hazardous multi-class classification. Expert Syst Appl (2017) 81:22–7. doi: 10.1016/
Materials (2022) 426:128067. doi: 10.1016/j.jhazmat.2021.128067 j.eswa.2017.02.049

24. Wu X, Zhou Q, Mu L, Hu X. Machine learning in the identification, 46. Chen Y-W, Lin C-J. Combining SVMs with various feature selection
prediction and exploration of environmental toxicology: Challenges and strategies. In: Guyon I, Nikravesh M, Gunn S, Zadeh LA, editors. Feature
perspectives. J Hazardous Materials (2022) 438:129487. doi: 10.1016/ extraction: Foundations and applications, vol. 315-324 . Berlin, Heidelberg:
j.jhazmat.2022.129487 Springer Berlin Heidelberg (2006). doi: 10.1007/978-3-540-35488-8_13

25. Kuhn M, Letunic I, Jensen LJ, Bork P. The SIDER database of drugs and side 47. Katritzky AR, Maran U, Lobanov VS, Karelson M. Structurally diverse
effects. Nucleic Acids Res (2015) 44(D1):D1075–9. doi: 10.1093/nar/gkv1075 quantitative structure-property relationship correlations of technologically relevant
physical properties. J Chem Inf Comput Sci (2000) 40(1):1–18. doi: 10.1021/
26. Kim S, Thiessen PA, Bolton EE, Chen J, Fu G, Gindulyte A, et al. PubChem ci9903206
substance and compound databases. Nucleic Acids Res (2015) 44(D1):D1202–13.
doi: 10.1093/nar/gkv951 48. Abdi H, Williams LJ. Principal component analysis. WIREs Comput Stat
(2010) 2(4):433–59. doi: 10.1002/wics.101
27. Sushko I, Novotarskyi S, Körner R, Pandey AK, Rupp M, Teetz W, et al.
Online chemical modeling environment (OCHEM): Web platform for data 49. Jolliffe IT, Cadima J. Principal component analysis: A review and recent
storage, model development and publishing of chemical information. J developments. Philos Trans R Soc A: Mathematical Phys Eng Sci (2016) 374
Computer-Aided Mol Design (2011) 25(6):533–54. doi: 10.1007/s10822-011- (2065):20150202. doi: 10.1098/rsta.2015.0202
9440-2 50. Smallwood MJ, Nissim A, Knight AR, Whiteman M, Haigh R, Winyard PG.
28. Choi RY, Coyner AS, Kalpathy-Cramer J, Chiang MF, Campbell JP. Oxidative stress in autoimmune rheumatic diseases. Free Radical Biol Med (2018)
Introduction to machine learning, neural networks, and deep learning. Trans 125:3–14. doi: 10.1016/j.freeradbiomed.2018.05.086
Vision Sci Technol (2020) 9(2):14–4. doi: 10.1167/tvst.9.2.14 51. Dedon PC, Tannenbaum SR. Reactive nitrogen species in the chemical
29. Hearst MA, Dumais ST, Osuna E, Platt J, Scholkopf B. Support vector biology of inflammation. Arch Biochem Biophysics (2004) 423(1):12–22.
machines. IEEE Intelligent Syst their Appl (1998) 13(4):18–28. doi: 10.1109/ doi: 10.1016/j.abb.2003.12.017
5254.708428 52. Magder S. Reactive oxygen species: toxic molecules or spark of life? Crit
30. Webb GI, Keogh E, Miikkulainen R. Naïve bayes. In: Encyclopedia of Care (2006) 10(1):208. doi: 10.1186/cc3992
machine learning, Springer, Boston, MA vol. 15. (2010). p. 713–4. doi: 10.1007/ 53. Hartung J. Organic radical reactions associated with nitrogen monoxide.
978-1-4899-7502-7_581-1 Chem Rev (2009) 109(9):4500–17. doi: 10.1021/cr900085j
31. Peterson LE. K-Nearest neighbor. Scholarpedia (2009) 4(2):1883. 54. d’Ischia M, Napolitano A, Manini P, Panzella L. Secondary targets of nitrite-
doi: 10.4249/scholarpedia.1883 derived reactive nitrogen species: Nitrosation/Nitration pathways, antioxidant
32. Myles AJ, Feudale RN, Liu Y, Woody NA, Brown SD. An introduction to defense mechanisms and toxicological implications. Chem Res Toxicol (2011) 24
decision tree modeling. J Chemometrics (2004) 18(6):275–85. doi: 10.1002/cem.873 (12):2071–92. doi: 10.1021/tx2003118

Frontiers in Immunology 12 frontiersin.org

Guo et al. 10.3389/fimmu.2022.1015409

55. Krumova K, Cosa G. Chapter 1 overview of reactive oxygen species. In: 57. Hultqvist M, Olsson LM, Gelderman KA, Holmdahl R. The protective role of ROS
Singlet oxygen: Applications in biosciences and nanosciences, volume 1. London, UK: in autoimmune disease. Trends Immunol (2009) 30(5):201–8. doi: 10.1016/j.it.2009.03.004
The Royal Society of Chemistry (2016). p. 1–21. doi: 10.1039/9781782622208- 58. Takorabet L, Ropars A, Stasiuk L, Raby C, Charreire J. Phenothiazine-
00001 induced increase in thyroid autoantigens and costimulatory molecules on thyroid
56. Yang B, Chen Y, Shi J. Reactive oxygen species (ROS)-based nanomedicine. cells: A pathophysiological mechanism for drug-induced autoimmunity? Clin Exp
Chem Rev (2019) 119(8):4881–985. doi: 10.1021/acs.chemrev.8b00626 Immunol (2001) 111(2):415–21. doi: 10.1046/j.1365-2249.1998.00507.x

Frontiers in Immunology 13 frontiersin.org