Prediction of Drug-Induced Liver Injury: From Molecular Physicochemical Properties and Scaffold Architectures to Machine Learning Approaches

doi:10.21203/rs.3.rs-4268191/v1

Download PDF

Research Article

Prediction of Drug-Induced Liver Injury: From Molecular Physicochemical Properties and Scaffold Architectures to Machine Learning Approaches

https://doi.org/10.21203/rs.3.rs-4268191/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The process of developing new drugs is widely acknowledged as being time-intensive and requiring substantial financial investment. Despite ongoing efforts to reduce time and expenses in drug development, ensuring medication safety remains an urgent problem. One of the major problems involved in drug development is hepatotoxicity, specifically known as drug-induced liver injury (DILI). The popularity of new drugs often poses a significant barrier during development and frequently leads to their recall after launch. In silico methods have many advantages compared with traditional in vivo and in vitro assays. To establish a more precise and reliable prediction model, it is necessary to utilize an extensive and high-quality database consisting of information on drug molecule properties and structural patterns. In addition, we should also carefully select appropriate molecular descriptors that can be used to accurately depict compound characteristics. The aim of this study was to conduct a comprehensive investigation into the prediction of DILI. First, we conducted a comparative analysis of the physicochemical properties of extensively well-prepared DILI-positive and DILI-negative compounds. Then, we used classic substructure dissection methods to identify structural pattern differences between these two different types of chemical molecules. These findings indicate that it is not feasible to establish property or substructure-based rules for distinguishing between DILI-positive and DILI-negative compounds. Finally, we developed quantitative classification models for predicting DILI using the naïve Bayes classifier (NBC) and recursive partitioning (RP) machine learning techniques. The optimal DILI prediction model was obtained using NBC, which combines 21 physicochemical properties, the VolSurf descriptors, and the LCFP_10 fingerprint set. This model achieved a global accuracy (GA) of 0.855 and an area under the curve (AUC) of 0.704 for the training set, while the corresponding values were 0.619 and 0.674 for the test set, respectively. Moreover, indicative substructural fragments favorable or unfavorable for DILI were identified from the best naïve Bayesian classification model. These findings may help prioritize lead compounds in the early stage of drug development pipelines.

drug-induced liver injury (DILI)

physicochemical property

scaffold feature

naïve Bayesian classifier (NBC)

recursive partitioning (RP)

machine learning

The challenging nature of drug development is an unequivocal reality, demanding substantial time and financial resources. Remarkably, statistical data reveal that the average expenditure to bring a single drug to fruition reaches an astounding $2.6 billion[1, 2]. Within this intricate landscape, drug-induced liver injury (DILI) has emerged as a formidable adverse concern, representing a primary factor for the eventual fall of drug development and necessitating market recalls[3]. Recognizing the criticality of this issue is vital and it is imperative to prioritize comprehensive safety evaluations throughout the entire drug development pipeline. Therefore, identifying a practical and trustworthy method for accurately assessing DILI prior to finalizing the development of a drug is crucial. Currently, the evaluation of DILI relies primarily on in vivo and in vitro assays; however, these techniques face challenges such as the long time needed, high cost, and remarkable variations between different species, thereby impeding their effectiveness in accurately predicting DILI in real-world scenarios[4]. In contrast, in silico models such as the quantitative structure‒activity relationship (QSAR) model have advanced in surmounting various challenges posed by in vivo and in vitro assays, offering the ability to predict DILI at an earlier stage, even before the synthesis of physical drug molecules[5]. This capacity is well-suited to the demands and objectives of both the academic and industrial realms[6].

The AI, which refers specifically to the concept of weak AI, has been widely applied for predicting DILI. In recent years, significant advancements have been made by various researchers in the application of artificial intelligence (AI) methodologies for predicting DILI. These algorithms encompass conventional methods such as decision tree, naïve Bayes, random forest, and support vector machine (SVM), as well as emerging artificial neural network (ANN) algorithms such as deep neural network (DNN), convolutional neural network (CNN), and graph neural network (GNN). Among all the related research dedicated to predicting DILI using machine learning algorithms[7–13], one prominent model is the ensemble approach proposed by Mora et al[14]. The authors replace the traditional 0-3D indices with the more informative 0-2.5D algebraic indices obtained from QuBiLS-MAS software. The training set comprised 1075 compounds, while two separate test sets were utilized for model evaluation: one consisting of 120 molecules and the other consisting of 47. Additionally, they employed five external test sets for further performance evaluation. Subsequently, a series of ensemble learning models were built, using k-NN, MLP, RF, NB, SVM, logistic regression, a classification tree, Fisher's linear discriminant analysis (FLDA), and Bayes network algorithms. Among these models, the one with the optimal performance achieved an accuracy of 84% in 10-fold cross-validation for prediction ability. Wang and Chen[15] proposed five consensus models using SVM and hybrid quantum particle swarm optimization (HQPSO) algorithms. These models were constructed based on PaDEL descriptors, and substructures, as well as Klekota-Roth, Estate, and PubChem fingerprints. The dataset that was employed for model development included 2608 chemical compounds. The best-performing consensus model developed by researchers achieved an accuracy (ACC) of 80% in predicting the hepatotoxicity of 528 tested chemical compounds. The model exhibited a sensitivity (SE) of 83.9% and a specificity (SP) of 73.3%, accurately identifying hepatotoxic compounds with high accuracy. Furthermore, deep learning algorithms have been applied in several studies to improve the accuracy of DILI prediction. Li et al.[16] utilized a deep learning algorithm that makes great improvements compared with other methods, as it combines model-level representations generated by conventional machine learning algorithms with Mold2 descriptors instead of raw data representing molecules. The dataset used consisted of 1002 drug molecules, achieving an ACC of 68.7% and a MCC of 0.331. The performance surpassed that of five traditional machine learning models and two advanced ensemble learning models. Nguyen-Vo et al.[17] developed a DILI prediction model using CNN algorithms and molecular fingerprint-embedded features. They utilized a development set comprising 1597 compounds and a test set containing 322 compounds. Their study demonstrated exceptional performance with an ACC of 89% and a ROC-AUC of 96%. Ma et al.[10] employed a GNN-based method that incorporated attribute increments from four toxicity datasets, which resulted in an expanded dataset comprising 15675 compounds for DILI prediction. The utilization of this methodology notably enhanced both the precision (from 78.8–81.4%) and the AUC (from 86.6–88.2%) for tasks associated with DILI prediction. Hwang et al.[18] developed GLIT, a graph neural network-based model that adopts ECFP molecular fingerprints for drug structure representation while considering the chirality of drug molecules. A total of 12348 samples were curated from both the Cmap dataset and the LTKB dataset. Their model achieved an ACC of 77.3% and a MCC of 0.553, outperforming three previous benchmark models, namely, the MLP, GCNN, and GraphSAGE models[19]. Recently, Shang et al.[20] proposed ResNet18DNN, a novel predictive model for DILI, inspired by convolutional neural networks commonly used in computer vision research. They evaluated the model on a dataset consisting of 1446 compounds and achieved an accuracy of 95%. Although these models generally show satisfactory performance, they occasionally encounter challenges such as imbalanced sensitivity versus specificity, an ambiguous standard for hepatotoxicity definition, or limited availability of DILI-positive/negative compound data. These challenges considerably restrict the practical applicability of these models in real-world scenarios[9, 21, 22].

In this study, we conducted a deep examination of factors closely associated with DILI based on the physicochemical properties and structural characteristics of a meticulously curated dataset comprising 2145 compounds. This dataset was collected and compiled through an extensive literature review. Of the 2145 compounds, 1141 were classified as DILI-positive compounds, while the remaining 1004 were categorized as DILI-negative compounds according to experimental toxicity. To the best of our knowledge, the number of molecules investigated in this work surpasses that of previously published studies. We are confident that our comprehensive analysis of DILI data will ensure reliable findings and robust conclusions. First, we carefully analyzed factors closely linked to DILI from the perspectives of both the physicochemical properties and structural features of the compounds. Subsequently, two distinct machine learning algorithms, the naïve Bayes classifier and recursive partitioning, were utilized to develop prediction models for DILI based on different combinations of molecular physicochemical properties, VolSurf descriptors and molecular fingerprints. The favorable or unfavorable structural features for DILI, identified from the optimal naïve Bayesian classification model, were briefly analyzed and discussed. We anticipate that our study will provide valuable insights for prioritizing lead compounds and preventing DILI during the drug discovery process.

2.1. Data collection. Our dataset was obtained primarily from the work of Kotsampasakou et al.,[21] who utilized various search tools such as PubMed, Google, and Scopus to extract 12 distinct datasets related to DILI in humans[23–34]. After preprocessing, which included deduplication, curation of class labels, and standardization, we generated a dataset comprising 2145 compounds (1141 DILI-positive and 1004 DILI-negative compounds). The supplementary information provides detailed data on DILI-positive and DILI-negative compounds. The molecules in both databases were subjected to molecular mechanics (MM) minimization using the MMFF94 force field[35] of Discovery Studio software[36]. The detailed protocol for dataset preparation can be found in our previous studies[37–42].

2.2. Calculation of molecular physicochemical property descriptors. Thirty representative molecular property descriptors widely used in ADME/T predictions were chosen and analyzed in the present study[10, 37, 38, 41–44]. These descriptors encompass a range of properties, including the octanol–water partition coefficient based on the Ghose and Crippen methods (AlogP)[45], the apparent distribution coefficient at pH = 7.4 based on the Csizmadia method (logD)[46], the molecular solubility based on multiple linear regression developed by Tekto (logS)[47], the molecular weight (MW), the molecular polar surface area (MPSA), 22 descriptors (6 ~ 27 in Table 1) encompassing atoms and bonds, and ring counts with varying complexities, as well as the capacity for hydrogen bond formation (28 ~ 29 in Table 1), which were considered alongside a classical drug-likeness criterion known as the QED[48]. All the molecular descriptors were calculated using Discovery Studio software in our study[36].

2.3. Calculation of molecular fingerprints. We utilized two distinct types of molecular fingerprints in our study, namely SciTegic extended-connectivity fingerprints and Daylight-style path-based fingerprints[50–52]. The SciTegic extended-connectivity fingerprints (ECFP, ECFC, FCFP, and LCFP) are molecular fingerprint algorithms developed by Accelrys to characterize molecular structures. They enhance conventional connectivity-based molecular fingerprints by incorporating additional improvements to capture both local and global structural information of molecules. These algorithms iteratively analyze the connectivity patterns between atoms within a single compound, transforming the original molecular structure into a binary bit string representation. Starting from each atom constituent of a compound, algorithms investigate the structural features of its neighboring atoms and determine the presence or absence of specific topological substructures. This process continues recursively until the specified fingerprint radius is reached. These methods not only consider the occurrence of local fragments within a compound but also incorporate their spatial relationships within the molecule. Compared to standard connectivity-based fingerprints, SciTegic extended-connectivity fingerprints offer further progress in encoding molecular features and show greater discrimination power.

Table 1

Descriptions of the 30 molecular descriptors used for comparing physicochemical properties
No.	Molecular Parameter	Parameter Meaning
1	AlogP	Octanol-water partition coefficient
2	logD_7.4	Apparent distribution coefficient at pH 7.4
3	logS	Molecular water solubility
4	MW	Molecular weight
5	MPSA	Molecular polar surface area
6	N_Rot	Number of rotatable bonds
7	N_Rings	Number of rings
8	N_AR	Number of aromatic rings
9	N_C	Number of carbon atoms
10	N_N	Number of nitrogen atoms
12	N_Atoms	Number of atoms
13	N_Bonds	Number of bonds
14	N_DoubleBonds	Number of double bonds
15	N_BHA	Number of bridgehead atoms
16	N_RingBonds	Number of ring bonds
17	N_{RingFusionBonds}	Number of fused ring bonds
18	N_{AromaticBonds}	Number of aromatic bonds
19	N_BridgeBonds	Number of bridge bonds
20	N_RA	Number of ring assemblies
21	N_R3	Number of 3-membered rings
22	N_R4	Number of 4-membered rings
23	N_R5	Number of 5-membered rings
24	N_R6	Number of 6-membered rings
25	N_R7	Number of 7-membered rings
26	N_R8	Number of 8-membered rings
27	N_R9+	Number of rings with 9 or more members
28	N_HBA	Number of hydrogen bond acceptors
29	N_HBD	Number of hydrogen bond donors
30	QED[49]	Quantitative estimation of drug-likeness index

Then, we employed Daylight-style path-based fingerprints as another type of molecular fingerprint. This approach can capture crucial structural information by analyzing the paths that atoms follow within a molecule. By considering all possible paths linking pairs of atoms, path-based fingerprints offer an exhaustive depiction of a molecule's connectivity; these fingerprints can take various forms and involve multiple types of chemical bonds. The fingerprint contains embedded indicators that can denote the presence or absence of specific substructures along these paths, which are determined through an iterative process until the predefined path length is achieved. This results in a binary fingerprint that encodes the existence or absence of these substructures. Compared with traditional connectivity-based fingerprints, Daylight-style path-based fingerprints provide a more comprehensive description of both local and global connectivity patterns and have been extensively applied in various molecular property prediction tasks such as biological activity modeling, virtual screening, and clustering analysis[52].

In general, molecular fingerprint symbols typically consist of four uppercase letters, each with distinct meanings. The first letter is usually F, E, or L, where F represents a functional role code that encompasses aromatic elements, positively and negatively charged atoms, halogens, hydrogen bond donors and acceptors, among others. E represents an atom type code that contains the total count of specific atoms meeting certain criteria as well as their elemental type, charge and atomic weight. L represents the atom type code used for calculating the partition coefficient AlogP and comprises the 120 atom types required for AlogP calculations. The second uppercase letter in the molecular fingerprint symbol, either C or P, denotes distinct categories of fingerprints. Specifically, C represents connectivity-based fingerprints while P denotes path-based fingerprints. The fourth letter of the fingerprint symbol represents the maximum interatomic distance considered. For example, a functionalized path-based fingerprint with a maximum distance of 4 can be denoted as FPFP_4. However, in this study, we did not employ a distance parameter of 2 for fingerprinting because it is important to maintain an appropriate range between molecular fragments that are neither excessively large nor too small. Various types of fingerprints of different lengths were utilized in this research including FCFC, FCFP, FPFC, FPFP, ECFC, ECFP, EPFC, EPFP, LCFC, LCFP, LPFC and LPFP; these molecular fingerprints were generated using the Discovery Studio molecular simulation package (version 3.1)[36].

Furthermore, we used VolSurf fingerprints[53], a distinctive fingerprinting technique that captures essential information based on the three-dimensional shape and electrostatic potentials of molecules, thereby providing a comprehensive characterization of their physicochemical properties. Therefore, the present study incorporated a comprehensive set of 76 VolSurf descriptors, which have demonstrated efficacy in predicting pharmacokinetic properties and were calculated using the MOE simulation package[54].

2.4. Generation of scaffold features. In this study, we applied two primary methods, namely, the Murcko framework and Scaffold Tree methods, to extract molecular substructure features from chemical compounds (Fig. 1). The Murcko framework method was introduced by Bemis et al.,[55] while the Scaffold Tree method was proposed by Schuffenhauer et al[56]. First, we utilized the Scaffold Tree method to hierarchically partition the small molecule compounds. The process can be summarized into several steps. Initially, the side chains of the molecules were removed from the main scaffold architectures, corresponding to a single cut, resulting in the formation of the Murcko framework structure. Subsequently, based on the predefined iterative cutting rules set in the Scaffold Tree method, we incrementally divided the Murcko framework structure, layer by layer, until only a basic ring structure remained in the compound. Ultimately, we organized these various cyclic substructures that were generated from the above procedures. These substructures were designated Level 0 to Level n fragments, encompassing the entire cutting process of the Scaffold Tree method. In subsequent investigations, determining the appropriate level for distinguishing and classifying DILI is a pivotal concern. The diverse complexity of the studied molecular structures may result in certain structures being relatively simple and not reaching Level 3 or higher after being processed via the Scaffold Tree method. Moreover, Level 0 substructures are excessively simplistic and homogeneous, rendering them unsuitable for further research. Several researchers have suggested that fragments generated at Level 1 through the Scaffold Tree can effectively facilitate chemical structure analysis[57]. For this study, we employed the Murcko frameworks together with the Levels 1 and 2 frameworks of Scaffold Tree for further investigation. Additionally, we developed a straightforward protocol to identify scaffolds by examining the presence of cyclic substructures, including simple rings, ring assemblies, and bridge assemblies (Fig. 1). The side chains of molecules also possess significant value for a wide range of applications, such as enhancing synthetic accessibility, improving solubility, and addressing concerns related to metabolism and toxicity[58]. In our analysis, we also considered the side chains linked to Murcko frameworks (Fig. 1). The Generate Fragments module in Pipeline Pilot 7.5 software was used to generate structures with varying levels of complexity.

2.5 Analysis of scaffold diversity. The datasets of DILI-positive/DILI-negative compounds were subjected to scaffold diversity analysis, wherein duplicated side chains and ring systems were removed to obtain unique ones for comparison with the scaffold diversity in the analyzed datasets. The distribution of molecules across distinct scaffolds within each dataset and the structural diversity of these scaffolds were utilized as metrics to define their respective scaffold diversities. In this study, the distribution of molecules across distinct scaffolds was quantified using scaffold counts and cumulative scaffold frequencies (CSFs)[57], while Tree Maps were utilized to characterize both the distribution and structural diversity of the scaffolds.

2.6 Tree Maps generation. The Tree Maps method, proposed by Schneiderman[59], differs significantly from traditional tree-like distribution representation methods because it employs a 2D packing approach to visually depict the structures and properties of molecules. Currently, this technique is extensively employed in molecular substructure analysis. The molecular substructures are visually represented in Tree Maps, with variations in sizes, quantities, and colors to depict different structures. Prior to constructing the Tree Maps, cluster analysis of the given substructures was required based on the similarity of molecular fingerprints, such as ECFP_6. Typically, a restriction of 50 or fewer substructure features is defined within each category. This clustering process uses the Cluster Ligands module in Discovery Studio software and can be repeated multiple times to mitigate the negative impact of random variations when clustering molecules. Based on our clustering results, Tree Maps were generated to visually represent the distinctive characteristics of substructures within the DILI-positive/negative categories.

2.7. Machine learning approaches. We employed two machine learning algorithms, namely the naïve Bayes algorithm and the recursive partitioning algorithm, to establish and evaluate predictive models for DILI classification.

2.7.1 Naïve Bayes. The naïve Bayes method was developed based on the principles of probability and statistics utilizing the Bayesian theorem. The Bayesian theorem is defined as follows:

$$P\left(c|F\right)=\frac{P\left(F|c\right)P\left(c\right)}{P\left(F\right)}=\frac{P\left(c\right)}{P\left(F\right)}\prod _{i=1}^{n}P\left({f}_{i}|c\right)$$

It is important to note that we make an idealized treatment in the naïve Bayes algorithm, which states that all the molecular descriptor attributes are independent. Additionally, for attributes with continuous values, we assume that their probability distribution follows a normal distribution to calculate their probabilities.

To investigate the performance of the naïve Bayes algorithm, we employed the Create Bayesian Model protocol in Discovery Studio software. Our training set for model building consisted of 1717 compounds, with a diversified selection achieved by extracting 80% of the compounds based on their physicochemical property distribution. The remaining 20% served as our test set for model evaluation. We built various naïve Bayes machine learning models using different combinations of molecular property descriptors and fingerprints to achieve high prediction accuracy.

During the process of implementing the naïve Bayes algorithm in machine learning, particular attention was given to two specific issues: missing values and zero statistics. Missing values refer to a situation where the molecular descriptors in the test set used for assessing the model's predictive performance are not present in the training set. Additionally, zero statistics refer to a circumstance where the training set incompletely covers all molecular descriptors. For the latter scenario, Laplacian correction is commonly employed as a solution. However, for the former scenario, ignoring missing values is typically adopted to enhance model reliability. Compared to other machine learning algorithms, the naïve Bayes algorithm exhibits outstanding advantages such as requiring less training data, generally achieving high prediction accuracy, and having faster training speed[38, 41, 60, 61].

2.7.2 Recursive partitioning. The recursive partitioning (RP) algorithm is widely utilized in machine learning to categorize initially mixed data into different categories based on specific splitting criteria, which are defined in the formation of trees. At each node of the tree, the RP algorithm selects a variable from the structural descriptors that describe compound molecule features and calculates a splitting threshold based on predefined rules for classifying the original data. This iterative process continues until either the maximum depth of the decision tree is reached or a node lacks sufficient data for further division. In this study, using Discovery Studio software, we established a minimum sample size requirement of 5 samples per node and limited the maximum depth of the tree to 10 during model construction. In comparison to the naïve Bayes classifier, the RP classifier exhibits enhanced intuitiveness and comprehensibility; nevertheless, it also contains inherent risks such as overfitting and potentially diminished prediction accuracy.

2.8. Classification model evaluation. We evaluated the performance of machine learning models by assessing the numbers of true positive (TP), true negative (TN), false positive (FP), and false negative (FN) samples. Our primary focus was on the following parameters: sensitivity (SE), specificity (SP), positive predictive value (Q+), negative predictive value (Q-), and overall accuracy (GA).

SE = TP/(TP + FN) (1)

SP = TN/(TN + FP) (2)

Q + = TP/(TP + FP) (3)

Q- = TN/(TN + FN) (4)

GA = (TP + TN)/(TP + TN + FP + FN) (5)

In addition to our analysis, we further evaluated the predictive performance of the classification models by assessing the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. The AUC values typically range from 0 to 1, where a value of 1 indicates excellent predictive performance and a value below 0.5 suggests inadequate discriminatory ability.

3.1. DILI prediction based on molecular physicochemical properties. The differences in 36 molecular properties between compounds with and without DILI were analyzed and are summarized in Table 2 to elucidate the relationships between the essential physicochemical properties of the molecules and DILI. Additionally, we examined the distribution patterns of eight physicochemical properties commonly employed in ADME/T predictions within datasets categorized as DILI-positive or DILI-negative (Fig. 2). These properties include the octanol–water partition coefficient (AlogP), apparent distribution coefficient (logD), solubility (logS), molecular weight (MW), molecular polar surface area (MPSA), number of rotatable bonds (N_Rot), number of hydrogen bond donors (N_HBD), and number of hydrogen bond acceptors (N_HBA). The distribution differences between these two compound types were assessed using Student’s t-test based on various molecular descriptors.

Table 2

The 36 representative physicochemical properties of the 1141 DILI-positive and 1004 DILI-negative compounds
No	Descriptors	DILI-positive			DILI-negative
No	Descriptors	Max	Min	Average (± StdDev)	Max	Min	Average (± StdDev)
1	AlogP	13.41	-18.27	2.24 ± 2.70^a	14.69	-26.25	1.67 ± 3.60
2	logD_7.4	13.41	-23.45	1.64 ± 2.92	10.25	-32.81	0.99 ± 3.89
3	logS	2.47	-14.72	-4.28 ± 2.45	4.58	-31.28	-4.12 ± 3.32
4	MW	2639.13	42.08	361.59 ± 188.12	4541.07	44.05	379.18 ± 347.31
5	N_Rot	40	0	5.29 ± 4.60	148	0	6.63 ± 11.29
6	N_Rings	13	0	2.75 ± 1.57	16	0	2.56 ± 1.84
7	N_AR	6	0	1.44 ± 1.11	8	0	1.20 ± 1.17
8	N_HBA	28	0	3.09 ± 2.68	47	0	3.03 ± 4.07
9	N_HBD	39	0	1.59 ± 2.36	55	0	2.16 ± 4.67
10	MPSA	1310.93	0	94.94 ± 81.34	1861.54	0	104.89 ± 152.65
11	N_C	102	1	17.96 ± 9.73	207	0	18.62 ± 15.83
12	N_N	36	0	2.33 ± 2.37	56	0	2.38 ± 4.56
13	N_O	38	0	3.74 ± 3.57	58	0	4.22 ± 5.26
14	N_Atoms	177	2	25.02 ± 13.34	322	3	25.98 ± 24.08
15	N_Bonds	180	1	26.77 ± 14.45	333	3	27.53 ± 25.11
16	N_DoubleBonds	31	0	6.38 ± 3.51	70	0	5.77 ± 5.61
17	N_BHA	13	0	0.15 ± 0.90	16	0	0.15 ± 0.90
18	N_RingBonds	93	0	15.19 ± 9.11	78	0	13.89 ± 9.75
19	N_{RingFusionBonds}	14	0	0.86 ± 1.22	12	0	0.86 ± 1.30
20	N_{AromaticBonds}	34	0	8.26 ± 6.26	50	0	6.88 ± 6.59
21	N_BridgeBonds	87	0	0.82 ± 4.95	56	0	0.73 ± 3.95
22	N_RA	8	0	1.83 ± 1.10	11	0	1.62 ± 1.15
23	N_R3	3	0	0.04 ± 0.21	3	0	0.03 ± 0.18
24	N_R4	2	0	0.05 ± 0.23	2	0	0.04 ± 0.19
25	N_R5	4	0	0.59 ± 0.78	8	0	0.59 ± 0.95
26	N_R6	10	0	1.98 ± 1.27	10	0	1.84 ± 1.41
27	N_R7	2	0	0.06 ± 0.24	1	0	0.03 ± 0.18
28	N_R8	1	0	0 ± 0.07	1	0	0.01 ± 0.08
29	N_R9+	3	0	0.04 ± 0.22	2	0	0.03 ± 0.17
30	QED	0.94	0.01	0.55 ± 0.23	0.92	0.01	0.53 ± 0.23
^arepresents standard deviation (SD)

According to our calculations of the eight molecular physicochemical properties, we first assessed the hydrophobicity-related properties which included the lipid-water distribution coefficient (AlogP), apparent distribution coefficient (logD), and solubility (logS). The AlogP values for the DILI-positive compounds ranged from − 18.27 to 13.41, with an average value of 2.24, while those for the DILI-negative compounds ranged from − 26.25 to 14.69, with an average value of 1.67. For logD, DILI-positive compounds had a range between − 23.45 and 13.41, with an average value of 1.64, whereas DILI-negative compounds had a range between − 32.81 and 10.25, with an average value of 0.99. The logS values for the DILI-positive compounds ranged between − 14.72 and 2.47, with an average value of -4.28, whereas for the DILI-negative compounds, the values ranged between − 31.28 and 4.58, with an average value of -4.12. Based on these results, it can be concluded that, on average, compounds demonstrating hepatotoxicity generally exhibit greater lipophilicity than those without hepatotoxicity.

The maximum molecular weight for the DILI-positive compounds was 2639.13, while it reached 4541.07 for the DILI-negative compounds. However, the average values between these two groups were comparable, suggesting that the molecular weight is not a highly discriminative property for distinguishing DILI. Conversely, in terms of the number of rotatable bonds (N_Rot), DILI-negative compounds exhibited higher counts than DILI-positive compounds, indicating that greater flexibility is associated with nonhepatotoxicity.

The remaining three molecular descriptors included the polar surface area, hydrogen bond acceptor count, and hydrogen bond donor count. These descriptors are usually used for characterizing the hydrogen bonding and electrostatic properties of molecules. The p-values for these descriptors were also calculated across different compound categories. The obtained p-values for the polar surface area, hydrogen bond acceptor count, and hydrogen bond donor count were 0.06, 0.7, and 0.000303 respectively, suggesting that these descriptors do not possess discriminatory capabilities in predicting DILI. For constructing a high-performing and reliable predictive model, it is preferable for the p-values to be below the 10^− 20 threshold level. As a result, none of these eight molecular descriptors can be used to effectively discriminate compounds for their hepatotoxicity properties. Furthermore, regarding the other 28 descriptors encompassing various physicochemical parameters evaluated in our study, the statistical results indicate that accurate discrimination between DILI-positive and DILI-negative compounds cannot be achieved solely based on these simplistic physicochemical property-based rules/filters.

3.2. DILI prediction based on scaffold features.

3.2.1. Cumulative scaffold frequencies (CSFs) of the Murcko frameworks and the Levels 1 and 2 scaffolds. We utilized the Scaffold Tree tool to fragment the compound molecules, generating a series of substructures including Murcko frameworks, rings, ring assemblies, and bridge assemblies. These substructures collectively represent the cyclic components present in the molecules. Based on our analysis results (Table 3), a majority of the compounds used in this study exhibited cyclic structures, while only a small proportion lacked such structures. For example, the proportion of DILI-positive compounds containing Level 1 substructures was 79.05%, while for DILI-negative compounds, it was 68.13%. For the Level 2 scaffolds, the percentages of molecules exhibiting them in the DILI-positive and DILI-negative categories were 56.35% and 47.91%, respectively. However, the proportion of molecules possessing scaffolds at Level 3 or higher was less than 30%. Therefore, employing Level 1 or Level 2 scaffolds to characterize the structural features of both DILI-positive and DILI-negative compounds, as well as comparing scaffold diversity between them, is appropriate and justified.

Table 3

The number of scaffolds at different levels of the Scaffold Tree for 1141 DILI-positive and 1004 DILI-negative compounds
Level	No. of Scaffolds		No. of non-duplicated Scaffolds
Level	DILI-positive	DILI-negative	DILI-positive	DILI-negative
Level 0	1056(92.55%)	880(87.65%)	211(18.49%)	169(16.83%)
Level 1	902(79.05%)	684(68.13%)	421(36.90%)	331(32.97%)
Level 2	643(56.35%)	481(47.91%)	415(36.37%)	303(30.18%)
Level 3	339(29.71%)	283(28.19%)	229(20.07%)	175(17.43%)
Level 4	101(8.85%)	102(10.16%)	85(7.45%)	72(7.17%)
Level 5	42(3.68%)	56(5.58%)	35(3.07%)	43(4.28%)
Level 6	22(1.93%)	26(2.59%)	19(1.67%)	21(2.09%)
Level 7	6(0.53%)	16(1.59%)	6(0.53%)	14(1.39%)
Level 8	4(0.35%)	7(0.70%)	4(0.35%)	7(0.70%)
Level 9	2(0.18%)	2(0.20%)	2(0.18%)	2(0.20%)
Level 10	1(0.09%)	1(0.10%)	1(0.09%)	1(0.10%)
Level 11	1(0.09%)	1(0.10%)	1(0.09%)	1(0.10%)
Level 12	1(0.09%)	0	1(0.09%)	0

As illustrated in Table 4, the presence of Murcko frameworks was observed in 92.55% and 87.65% of the DILI-positive and DILI-negative compounds, respectively, echoing the trends detailed in Table 3 and further substantiating the prevalence of cyclic structures among the analyzed compounds. The Murcko framework refers to the core structural skeleton of molecules remaining after the removal of their side chains. With respect to our datasets comprising DILI-positive and DILI-negative compounds, we identified 680 and 558 unique Murcko frameworks, respectively, indicating a certain level of structural diversity within these two types of molecules. A comparison of other cyclic systems, which include rings, ring assemblies, and bridge assemblies, revealed that the counts associated with DILI-positive compounds tended to be greater than those associated with DILI-negative compounds. A possible reason for this could be the relatively larger number of DILI-positive compounds in our dataset. Notably, chain assemblies exhibit an intriguing pattern; specifically, 668 unique chain assemblies were identified within the dataset of DILI-positive compounds, whereas 693 were identified within the dataset of DILI-negative compounds. This suggests that DILI-negative compounds exhibit greater variation in terms of side chains and linkers connecting major functional groups than the DILI-positive counterparts. Based on the above findings, it can be inferred that the Murcko framework, Level 1 and Level 2 scaffolds serve as representative substructures for effectively capturing the comprehensive structural features of the investigated molecules.

Table 4

The number of fragments with Murcko frameworks, ring assemblies, rings, bridge assemblies and chain assemblies present in 1141 DILI-positive and 1004 DILI-negative compounds
Scaffold architecture	No. of Scaffolds		No. of non-duplicated Scaffolds
Scaffold architecture	DILI-positive	DILI-negative	DILI-positive	DILI-negative
Murcko framework	1056(92.55%)	880(87.65%)	680(59.60%)	558(55.58%)
Rings	3137(2.75)	2569(2.56)	206(18.05%)	168(16.73%)
Ring assemblies	2086(1.83)	1626(1.62)	353(30.94%)	301(29.98%)
Bridge assemblies	52(4.56%)	51(5.08%)	35(3.07%)	32(3.19%)
Chain assemblies	8406(7.37)	7173(7.14)	668(58.55%)	693(69.02%)

Here, we conducted a statistical analysis of the cumulative scaffold frequencies (CSFs) of Murcko frameworks, Level 1, and Level 2 substructures of DILI-positive/negative compounds. The results were obtained and are presented in Figs. 3 and 4 by sorting the frequencies of each representative scaffold. As depicted in Fig. 3a, with an increasing number of distinct Murcko frameworks, the growth rate of DILI-negative compounds slightly surpassed that of DILI-positive compounds. This finding can be explained by the fact that there were fewer DILI-negative compounds in our dataset than DILI-positive compounds, which resulted in a reduced variety of Murcko frameworks. The top ten most frequently occurring Murcko frameworks in both the DILI-positive compounds and DILI-negative compounds datasets are displayed in Fig. 3b. The analysis of representative Murcko frameworks from both datasets revealed that certain frameworks were more prevalent in DILI-positive compounds, but they also appeared in DILI-negative compounds, such as benzene (81) and diphenylmethane (10). Furthermore, some basic fragments like the pyridine ring (11) and quinoline ring (9) are common across a wide range of compound classes, suggesting their limited ability to distinguish between DILI-positive and DILI-negative compounds. However, there are also specific fragments with high frequencies observed only in positive DILI compounds, including N-(5-thio-1-diazabicyclo [4.2.0]oct-2-en-7-yl)-2-(thiazol-4-yl)acetamide (10), 4-phenyl-3,4-dihydropyridine (8), and benzophenone (7), indicating their potential as structural warning fragments.

The CSF plots of the Level 1 substructure fragments were also calculated and are depicted in Fig. 4a. These plots illustrate the growth rate of DILI-positive compounds initially trailing behind that of DILI-negative compounds; however, this phenomenon was later overtaken by the former, suggesting a relatively uniform distribution of diverse Level 1 substructures within the dataset containing DILI-positive compounds. As depicted in Fig. 4b, we have identified several fundamental segments, alongside others that are present in both DILI-negative and DILI-positive compounds. Moreover, several distinctive frameworks have captured our attention. The presence of 1,2,3,4,4a,5,6,8a-octahydronaphthalene (12) and 3-methylene-4-phenyl-3,4-dihydropyridine (12) as characteristic fragments strongly suggests their potential hepatotoxicity.

The CSF curves of the Level 2 scaffold revealed consistent growth trends for both DILI-positive and DILI-negative compounds as the number of Level 2 substructures increased, indicating a uniform dispersion of these substructures in both compound types at the Level 2 scaffold. (Fig. 4c). However, notable disparities exist in the occurrence of certain substructures (Fig. 4d). The scaffolds, namely N-(8-oxo-5-thio-1-diazabicyclo[4.2.0]oct-2-en-7-yl)-2-(thiazol-4-yl)acetamide (15) and 7-(piperazin-1-yl)-1,4-dihydroquinolin-4-one (8) are frequently observed in DILI-positive compounds, but rarely found in compounds without DILI risk, suggesting their potential as structural warning fragments. Other fragments have limited representation due to their low occurrence in DILI-positive compounds or multiple appearances in DILI-negative compounds.

3.2.2. Tree Maps. Based on the aforementioned analysis, we obtained a comprehensive understanding of the distribution pattern of molecular substructures within the DILI-positive/negative compound datasets. However, the distribution of substructure fragment similarities among different compound categories has not been fully elucidated. To bridge this gap in knowledge, we employed Tree Maps to investigate and analyze the similarity of distribution patterns exhibited by substructure fragments, which belong to two distinct compound categories. Specifically, we employed a clustering analysis based on ECFP_6 molecular fingerprint similarity, which incorporated representative structures, including Murcko frameworks and Level 1 scaffolds from DILI-positive/negative categories. Finally, our clustering results were visually represented through the Tree Maps strategy (Figs. 5 and 6).

In the Tree Maps, each distinct structural category is depicted by a gray circle, where the circle’s size reflects the quantity of substructure fragments it represents. The Tree Maps of each compound category highlight the six largest gray circles, signifying the predominant Murcko frameworks or Level 1 scaffold of the Scaffold Tree. Additionally, the number of gray circles in the Tree Maps for the DILI-positive compounds was slightly greater than that for the DILI-negative compounds, indicating a slightly greater diversity of structures in the DILI-positive compounds. Finally, we observed variations in the most frequently occurring scaffolds within clusters for DILI-positive and DILI-negative compounds. As depicted in Fig. 5, specific substructures of the Murcko framework, such as 6-phenyl-2,3,5,6-tetrahydroimidazo[2,1-b]thiazole and phenyl 2-phenoxyacetate, exhibit distinct characteristics compared to DILI-negative fragments. It is important to note, nevertheless, that certain DILI-negative fragments may be considered as modifications or additions to comparable DILI-positive fragments. For example, the presence of 1-cyclopropylethan-1-one moiety on the methylene group relative to 1-benzylisoquinoline in the DILI-positive fragment 1-cyclopropyl-2-(6,7-dihydrothieno[3,2-c]pyridin-5(4H)-yl)-2-phenylethan-1-one could offer novel insights for mitigating DILI risk. Additionally, as a common thiophene ring was observed in both DILI-positive and negative compounds, its association with DILI remains inconclusive. In Fig. 6, the three Level 1 scaffold categories of DILI-positive compounds are represented by an octahydro-1H-indole isocyclic ring, a 2H-indazole ring, and a 1,3,3a,7a-tetrahydrobenzo[c][1, 2, 5]oxadiazole ring. Similarly, the three Level 1 scaffold categories of DILI-positive compounds also include bicyclic rings with similar structural patterns such as an indole ring, a tetrahydroisoquinoline ring, and an octahydronaphthalene ring. The only notable differences lie in slight modifications to their unilateral rings. Consequently, these Level 1 scaffold architectures of DILI-negative compounds have the potential to function as bioisosteres for reducing DILI risk. Additionally, it was observed that introducing oxygen atoms into 5-cyclopentylpent-4-en-1-ylbenzene attenuated its potential for causing DILI.

3.3 DILI prediction models based on the naïve Bayes classifier. Previous sections have shown that neither individual physicochemical descriptors nor molecular substructures can serve as effective classification criteria for predicting DILI. Hence, we employed the naïve Bayes classifier (NBC) machine learning approach to develop more robust DILI classification models. We partitioned the entire dataset of 2145 compound molecules (1141 DILI-positive and 1004 DILI-negative) into two distinct subsets: a training set and a test set. The training set comprised 80% of the original dataset, totaling 1717 compound molecules (913 DILI-positive and 804 DILI-negative), while the remaining 20% of the data consisted of 428 compound molecules (228 DILI-positive and 200 DILI-negative) that formed the test set. Each compound in the dataset was meticulously labeled to verify its association with DILI. To identify the most suitable molecular features for accurate prediction of DILI, we constructed naïve Bayes models using 21 physicochemical properties, 76 VolSurf descriptors, and different molecular fingerprints (48 combinations in total). These molecular fingerprint descriptors included the ECFC, ECFP, EPFC, EPFP, FCFC, FCFP, FPFC, FPFP, LCFC, LCFP, LPFC and LPFP with varying bond lengths ranging from 4 to 10. The statistical results obtained from the leave-one-out (LOO) cross-validation procedure of NBC are summarized in Table 5 for the training set and in Table 6 for the test set.

Table 5

Performance of the naïve Bayesian classifiers for the training set based on different combinations of molecular descriptors
Descriptors	TP	FN	FP	TN	SE	SP	Q+	Q-	GA	AUC
MP^a	499	414	237	567	0.547	0.705	0.678	0.578	0.621	0.621
V^b	527	386	241	563	0.577	0.700	0.686	0.593	0.635	0.629
MP + V	460	453	165	639	0.504	0.795	0.736	0.585	0.640	0.638
MP + V + ECFC_4	699	214	161	643	0.766	0.800	0.813	0.750	0.782	0.714
MP + V + ECFC_6	722	191	98	706	0.791	0.878	0.880	0.787	0.832	0.716
MP + V + ECFC_8	744	169	91	713	0.815	0.887	0.891	0.808	0.849	0.713
MP + V + ECFC_10	735	178	79	725	0.805	0.902	0.903	0.803	0.850	0.710
MP + V + ECFP_4	619	294	87	717	0.678	0.892	0.877	0.709	0.778	0.717
MP + V + ECFP_6	734	179	106	698	0.804	0.868	0.874	0.796	0.834	0.717
MP + V + ECFP_8	712	201	69	735	0.780	0.914	0.912	0.785	0.843	0.712
MP + V + ECFP_10	719	194	68	736	0.788	0.915	0.914	0.791	0.847	0.709
MP + V + EPFC_4	639	274	257	547	0.700	0.680	0.713	0.666	0.691	0.663
MP + V + EPFC_6	648	265	175	629	0.710	0.782	0.787	0.704	0.744	0.671
MP + V + EPFC_8	661	252	131	673	0.724	0.837	0.835	0.728	0.777	0.672
MP + V + EPFC_10	684	229	122	682	0.749	0.848	0.849	0.749	0.796	0.669
MP + V + EPFP_4	623	290	219	585	0.682	0.728	0.740	0.669	0.704	0.684
MP + V + EPFP_6	577	336	109	695	0.632	0.864	0.841	0.674	0.741	0.680
MP + V + EPFP_8	598	315	92	712	0.655	0.886	0.867	0.693	0.763	0.673
MP + V + EPFP_10	615	298	85	719	0.674	0.894	0.879	0.707	0.777	0.670
MP + V + FCFC_4	576	337	153	651	0.631	0.810	0.790	0.659	0.715	0.697
MP + V + FCFC_6	738	175	164	640	0.808	0.796	0.818	0.785	0.803	0.712
MP + V + FCFC_8	641	272	69	735	0.702	0.914	0.903	0.730	0.801	0.712
MP + V + FCFC_10	653	260	58	746	0.715	0.928	0.918	0.742	0.815	0.711
MP + V + FCFP_4	602	311	147	657	0.659	0.817	0.804	0.679	0.733	0.712
MP + V + FCFP_6	636	277	80	724	0.697	0.900	0.888	0.723	0.792	0.722
MP + V + FCFP_8	662	251	64	740	0.725	0.920	0.912	0.747	0.817	0.719
MP + V + FCFP_10	675	238	62	742	0.739	0.923	0.916	0.757	0.825	0.716
MP + V + FPFC_4	525	388	164	640	0.575	0.796	0.762	0.623	0.679	0.661
MP + V + FPFC_6	619	294	166	638	0.678	0.794	0.789	0.685	0.732	0.673
MP + V + FPFC_8	636	277	130	674	0.697	0.838	0.830	0.709	0.763	0.679
MP + V + FPFC_10	650	263	120	684	0.712	0.851	0.844	0.722	0.777	0.679
MP + V + FPFP_4	626	287	204	600	0.686	0.746	0.754	0.676	0.714	0.688
MP + V + FPFP_6	568	345	97	707	0.622	0.879	0.854	0.672	0.743	0.691
MP + V + FPFP_8	580	333	87	717	0.635	0.892	0.870	0.683	0.755	0.688
MP + V + FPFP_10	599	314	74	730	0.656	0.908	0.890	0.699	0.774	0.686
MP + V + LCFC_4	779	134	211	593	0.853	0.738	0.787	0.816	0.799	0.711
MP + V + LCFC_6	720	193	96	708	0.789	0.881	0.882	0.786	0.832	0.710
MP + V + LCFC_8	738	175	80	724	0.808	0.900	0.902	0.805	0.851	0.707
MP + V + LCFC_10	750	163	81	723	0.821	0.899	0.903	0.816	0.858	0.705
MP + V + LCFP_4	647	266	99	705	0.709	0.877	0.867	0.726	0.787	0.715
MP + V + LCFP_6	725	188	93	711	0.794	0.884	0.886	0.791	0.836	0.712
MP + V + LCFP_8	721	192	71	733	0.790	0.912	0.910	0.792	0.847	0.707
MP + V + LCFP_10	731	182	67	737	0.801	0.917	0.916	0.802	0.855	0.704
MP + V + LPFC_4	687	226	92	712	0.752	0.886	0.882	0.759	0.815	0.701
MP + V + LPFC_6	739	174	90	714	0.809	0.888	0.891	0.804	0.846	0.694
MP + V + LPFC_8	765	148	88	716	0.838	0.891	0.897	0.829	0.863	0.682
MP + V + LPFC_10	776	137	101	703	0.850	0.874	0.885	0.837	0.861	0.672
MP + V + LPFP_4	656	257	83	721	0.719	0.897	0.888	0.737	0.802	0.701
MP + V + LPFP_6	712	201	83	721	0.780	0.897	0.896	0.782	0.835	0.696
MP + V + LPFP_8	746	167	91	713	0.817	0.887	0.891	0.810	0.850	0.686
MP + V + LPFP_10	764	149	94	710	0.837	0.883	0.890	0.827	0.858	0.677
^aMP represents 21 molecular physicochemical properties;

^bV represents 76 VolSurf descriptors.

Table 6

Performance of the naïve Bayesian classifiers for the test set based on different combinations of molecular descriptors
Descriptors	TP	FN	FP	TN	SE	SP	Q+	Q-	GA	AUC
MP^a	98	130	73	127	0.430	0.635	0.573	0.494	0.526	0.548
V^b	126	102	89	111	0.553	0.555	0.586	0.521	0.554	0.562
MP + V	122	106	86	114	0.535	0.570	0.587	0.518	0.551	0.563
MP + V + ECFC_4	135	93	78	122	0.592	0.610	0.634	0.567	0.600	0.657
MP + V + ECFC_6	135	93	74	126	0.592	0.630	0.646	0.575	0.610	0.666
MP + V + ECFC_8	136	92	74	126	0.596	0.630	0.648	0.578	0.612	0.671
MP + V + ECFC_10	138	90	74	126	0.605	0.630	0.651	0.583	0.617	0.673
MP + V + ECFP_4	144	84	84	116	0.632	0.580	0.632	0.580	0.607	0.653
MP + V + ECFP_6	137	91	77	123	0.601	0.615	0.640	0.575	0.607	0.663
MP + V + ECFP_8	141	87	75	125	0.618	0.625	0.653	0.590	0.621	0.668
MP + V + ECFP_10	141	87	75	125	0.618	0.625	0.653	0.590	0.621	0.670
MP + V + EPFC_4	89	139	50	150	0.390	0.750	0.640	0.519	0.558	0.613
MP + V + EPFC_6	95	133	46	154	0.417	0.770	0.674	0.537	0.582	0.625
MP + V + EPFC_8	95	133	40	160	0.417	0.800	0.704	0.546	0.596	0.635
MP + V + EPFC_10	100	128	41	159	0.439	0.795	0.709	0.554	0.605	0.641
MP + V + EPFP_4	103	125	59	141	0.452	0.705	0.636	0.530	0.570	0.610
MP + V + EPFP_6	90	138	45	155	0.395	0.775	0.667	0.529	0.572	0.624
MP + V + EPFP_8	84	144	38	162	0.368	0.810	0.689	0.529	0.575	0.633
MP + V + EPFP_10	88	140	39	161	0.386	0.805	0.693	0.535	0.582	0.638
MP + V + FCFC_4	113	115	64	136	0.496	0.680	0.638	0.542	0.582	0.631
MP + V + FCFC_6	115	113	60	140	0.504	0.700	0.657	0.553	0.596	0.649
MP + V + FCFC_8	120	108	59	141	0.526	0.705	0.670	0.566	0.610	0.658
MP + V + FCFC_10	118	110	57	143	0.518	0.715	0.674	0.565	0.610	0.663
MP + V + FCFP_4	125	103	71	129	0.548	0.645	0.638	0.556	0.593	0.637
MP + V + FCFP_6	135	93	69	131	0.592	0.655	0.662	0.585	0.621	0.655
MP + V + FCFP_8	111	117	57	143	0.487	0.715	0.661	0.550	0.593	0.664
MP + V + FCFP_10	113	115	55	145	0.496	0.725	0.673	0.558	0.603	0.668
MP + V + FPFC_4	85	143	48	152	0.373	0.760	0.639	0.515	0.554	0.606
MP + V + FPFC_6	90	138	47	153	0.395	0.765	0.657	0.526	0.568	0.615
MP + V + FPFC_8	94	134	50	150	0.412	0.750	0.653	0.528	0.570	0.624
MP + V + FPFC_10	102	126	52	148	0.447	0.740	0.662	0.540	0.584	0.630
MP + V + FPFP_4	101	127	59	141	0.443	0.705	0.631	0.526	0.565	0.615
MP + V + FPFP_6	86	142	50	150	0.377	0.750	0.632	0.514	0.551	0.620
MP + V + FPFP_8	87	141	49	151	0.382	0.755	0.640	0.517	0.556	0.624
MP + V + FPFP_10	87	141	49	151	0.382	0.755	0.640	0.517	0.556	0.627
MP + V + LCFC_4	118	110	56	144	0.518	0.720	0.678	0.567	0.612	0.659
MP + V + LCFC_6	123	105	61	139	0.539	0.695	0.668	0.570	0.612	0.665
MP + V + LCFC_8	122	106	55	145	0.535	0.725	0.689	0.578	0.624	0.669
MP + V + LCFC_10	127	101	60	140	0.557	0.700	0.679	0.581	0.624	0.671
MP + V + LCFP_4	125	103	63	137	0.548	0.685	0.665	0.571	0.612	0.658
MP + V + LCFP_6	133	95	74	126	0.583	0.630	0.643	0.570	0.605	0.666
MP + V + LCFP_8	137	91	76	124	0.601	0.620	0.643	0.577	0.610	0.671
MP + V + LCFP_10	133	95	68	132	0.583	0.660	0.662	0.581	0.619	0.674
MP + V + LPFC_4	121	107	51	149	0.531	0.745	0.703	0.582	0.631	0.668
MP + V + LPFC_6	116	112	48	152	0.509	0.760	0.707	0.576	0.626	0.671
MP + V + LPFC_8	121	107	51	149	0.531	0.745	0.703	0.582	0.631	0.672
MP + V + LPFC_10	125	103	53	147	0.548	0.735	0.702	0.588	0.636	0.672
MP + V + LPFP_4	109	119	46	154	0.478	0.770	0.703	0.564	0.614	0.663
MP + V + LPFP_6	115	113	52	148	0.504	0.740	0.689	0.567	0.614	0.668
MP + V + LPFP_8	126	102	58	142	0.553	0.710	0.685	0.582	0.626	0.671
MP + V + LPFP_10	122	106	58	142	0.535	0.710	0.678	0.573	0.617	0.671
^aMP represents 21 molecular physicochemical properties;

^bV represents 76 VolSurf descriptors.

According to the statistical results, we found that the Bayesian classifiers failed to achieve satisfactory discrimination capacities, as indicated by the relatively low GA and AUC values based solely on the 21 molecular properties, VolSurf descriptors, or a combination of both. The corresponding global accuracies were only 0.621, 0.635, and 0.640 respectively, while the AUC values were 0.621, 0.629, and 0.638, respectively. The prediction accuracy of NBC models for DILI can be improved significantly by adding any type of molecular fingerprint. The significance of molecular fingerprint descriptors in the construction of the naïve Bayesian classification model is highlighted.

The performance of the Bayesian classifier was subsequently assessed on an independent test set. Among the various combinations of molecular descriptors, the model that integrated the LCFP_10 molecular fingerprint with 21 physicochemical properties and VolSurf descriptors demonstrated the most optimal performance. It achieved a sensitivity of 0.583, a specificity of 0.660, a global accuracy of 0.619, and an AUC value of 0.674. Notably, the model's performance on the test set was noticeably inferior to its performance on the internal training set. Furthermore, our model exhibited a slightly superior predictive performance compared to that of Kotsampasakou et al.’s[21] model, which attained a maximum prediction accuracy of only 0.60 and an AUC value of 0.64 on the external testing set. Additionally, when combining molecular physicochemical properties and VolSurf descriptors with different molecular fingerprints such as ECFC_8, ECFC_10, ECFP_10, LCFC_10, LCFP_8, LPFC_6, LPFC_8, LPFC_10, LPFP_8, and LPFP_10, the resulting models yielded similar AUC values to that obtained by our best-performing model. This observation implies that utilizing larger step sizes of molecular fingerprints when constructing NBC models for DILI prediction, could offer certain advantages.

In addition to its application in constructing classification models for the DILI prediction, the Bayesian classifier can also employ Bayesian scoring to identify crucial molecular substructures that are highly relevant or irrelevant to DILI. For pharmaceutical chemists, these identified key substructures can serve as valuable references during drug design/development, aiding in the prevention and mitigation of hepatotoxicity. Figure 7 shows the top 20 molecular substructures most strongly associated with DILI (Fig. 7a) and the top 20 molecular substructures least significantly related with DILI (Fig. 7b), ranked based on their Bayesian scores.

The data presented above demonstrate the absence of identical substructures between Figs. 7a and 7b. Based on the results, it is evident that compounds with multi-substituted pyridine-like features are frequently associated with DILI-positive warnings, while DILI-negative compounds do not exhibit such structural characteristics. Although the pyridine ring is generally considered a universe functional group in medicinal chemistry, its multiple substitutions may increase the risk of DILI. Additionally, this study has identified novel fragments related to DILI, such as methyl 2-methyl-2-phenoxypropanoate and their corresponding substructures which strongly correlate with the occurrence of DILI. The presence of a 4-chlorophenyl ethers fragment should be avoided in drug design due to its association with hepatotoxicity. Furthermore, comparing Figs. 7a and 7b reveals that symmetrical alkane fragments containing nitrogen or oxygen atoms are frequently observed in DILI-positive compounds; however, removing these heteroatoms reduces their potential for causing liver injury. The occurrence of nonhepatotoxicity-related fragments (Figs. 7b) also demonstrates that long-chain alkenes or long-chain acids containing alkenes, certain benzoates as well as common heterocycles such as imidazole, pyrrole and indole can be considered structurally safe for drug design purposes. Nevertheless, it is worth noting that the presence of methyl-substituted on five-membered heterocycles may increase the risk of DILI. The findings may have significant implications for the identification of more promising drug candidates with reduced risk of DILI in the drug development.

3.4. DILI prediction models based on recursive partitioning. Compared to the naïve Bayesian technique, RP can provide simpler and more comprehensible classification models characterized by a series of hierarchical rules. Consequently, RP was employed to construct decision trees for discriminating DILI-positive compounds from DILI-negative compounds, aiming to compare the performance of Bayesian and RP classifiers based on identical molecular descriptors and datasets. The complexity of a decision tree in the RP analysis was judiciously regulated by calibrating a crucial parameter, the tree depth, based on predictions observed on the test set. The range of tree depths varied from 5 to 15, and the corresponding models were assessed for their classification accuracies. Subsequently, a series of RP models were generated using Discovery Studio software, and the model with the best predictive performance was chosen (Fig. 8). The optimal classification model, trained with a tree depth of 12, exhibited a cross-validation AUC of 0.619 in the training set and an AUC of 0.569 in the test set. It is evident that the NBC significantly outperforms the RP classifier when using identical molecular descriptors and datasets.

In our study, we conducted a comprehensive analysis of the factors influencing DILI using a well-prepared dataset. Initially, our research concentrated on utilizing the basic physicochemical properties and substructural features to develop simple rule-based filters for distinguishing DILI-positive from DILI-negative compounds. Among these properties, the apparent distribution coefficient (logD) exhibited significant variation across different types of molecular compounds. However, relying solely on individual physicochemical property descriptors has proven insufficient for accurate DILI prediction. Next, we applied the Scaffold Tree tool to hierarchically partition and analyze substructures of two compound datasets, aiming to discern the inherent structural disparities between DILI-positive and DILI-negative compounds. The obtained results enabled us to pinpoint various structurally indicative fragments that are closely associated with DILI. Ultimately, using a variety of physicochemical properties, VolSurf descriptors and molecular fingerprints, we employed two well-established machine learning approaches, namely the NBC and RP classifiers, for the quantitative prediction of DILI. The NBC classification model, which incorporates physicochemical properties, VolSurf descriptors, and LCFP_10, exhibited superior performance, with a GA of 0.619 and an AUC of 0.674 in the external test set, representing a modest improvement over the results of previous investigations. Additionally, Bayesian analysis successfully identified the most crucial molecular fragments associated with either favorable or unfavorable DILI properties, thereby offering valuable insights for designing high-quality lead compounds in the initial phases of drug design/discovery.

Significant progress has been made in utilizing machine learning algorithms for the prediction of DILI. However, there are ample opportunities for further advancements in this field. A primary challenge lies in the limited sample size incorporated in the training dataset. Currently, the most advanced AI algorithms utilized in computer science, such as computer vision and natural language processing, predominantly rely on an array of neural networks constructed using deep learning methodologies. Nevertheless, it continues to be difficult to discern a substantial disparity in predictive efficacy between DILI classification models generated through deep learning and conventional machine learning algorithms. This issue arises due to the scarcity of available samples that fail to meet the requirements for training deep neural networks with millions or even billions of parameters. In future research, the integration of our constructed models into an ensemble learning framework could be explored. Moreover, comprehensive characterization of drug molecules through molecular graphs, coupled with the use of deep learning algorithms such as GNN on broader and more robust datasets, may significantly boost the predictive performance of DILI classification models, and thus promote existing research.

Conflict of interests:

The authors declare no conflict of interest.

Funding

This study was supported by the Priority Academic Program Development of the Jiangsu Higher Education Institutes (PAPD) and the Suzhou Science and Technology Development Plan Project (Grant No. SKY2023127).

Author Contribution

YLZ. and ZDZ. contributed equally to this work; Conceptualization: ST., XTK., and HQL.; Methodology: YLZ., ZDZ., and KW.; Analysis: YLZ., ZDZ., KW., JJ., YXW., and ST.; Writing: YLZ. and ST.; Supervision: ST., XTK., and HQL. All authors have read and agreed to the published version of the manuscript.

Acknowledgement

We thank Prof. Lei Xu at the Institute of Bioinformatics and Medical Engineering at Jiangsu University of Technology for providing the MOE, Discovery Studio, and Pipeline Pilot softwares used to construct and validate the DILI prediction model.

Data Availability

DILI prediction data (1141 DILI-positive compounds and 1004 DILI-negative compounds) that support the findings of this study are available in the supplementary information of this paper.

Ashburn TT, Thor KB (2004) Drug repositioning: identifying and developing new uses for existing drugs. Nat Rev Drug Discovery 3(8):673–683. https://doi.org/10.1038/nrd1468
Paul SM, Mytelka DS, Dunwiddie CT, Persinger CC, Munos BH, Lindborg SR, Schacht AL (2010) How to improve R&D productivity: the pharmaceutical industry's grand challenge. Nat Rev Drug Discovery 9(3):203–214. https://doi.org/10.1038/nrd3078
Wilke RA, Lin DW, Roden DM, Watkins PB, Flockhart D, Zineh I, Giacomini KM, Krauss RM (2007) Identifying genetic risk factors for serious adverse drug reactions: current progress and challenges. Nat Rev Drug Discovery 6(11):904–916. https://doi.org/10.1038/nrd2423
Ozer J, Ratner M, Shaw M, Bailey W, Schomaker S (2008) The current state of serum biomarkers of hepatotoxicity. Toxicology 245(3):194–205. https://doi.org/10.1016/j.tox.2007.11.021
DiMasi JA, Grabowski HG, Hansen RW (2016) Innovation in the pharmaceutical industry: New estimates of R&D costs. J Health Econ 47:20–33. https://doi.org/10.1016/j.jhealeco.2016.01.012
Regev A (2014) Drug-induced liver injury and drug development: industry perspective. Semin Liver Dis 34(2):227–239. https://doi.org/10.1055/s-0034-1375962
Chen MJ, Bisgin H, Tong L, Hong HX, Fang H, Borlak J, Tong WD (2014) Toward predictive models for drug-induced liver injury in humans: are we there yet? Biomark Med 8(2):201–213. https://doi.org/10.2217/bmm.13.146
Ekins S (2014) Progress in computational toxicology. J Pharmacol Toxicol Methods 69(2):115–140. https://doi.org/10.1016/j.vascn.2013.12.003
Minerali E, Foil DH, Zorn KM, Lane TR, Ekins S (2020) Comparing Machine Learning Algorithms for Predicting Drug-Induced Liver Injury (DILI). Mol Pharm 17(7):2628–2637. https://doi.org/10.1021/acs.molpharmaceut.0c00326
Ma H, An W, Wang Y, Sun H, Huang R, Huang J (2021) Deep Graph Learning with Property Augmentation for Predicting Drug-Induced Liver Injury. Chem Res Toxicol 34(2):495–506. https://doi.org/10.1021/acs.chemrestox.0c00322
Williams DP, Lazic SE, Foster AJ, Semenova E, Morgan P (2020) Predicting Drug-Induced Liver Injury with Bayesian Machine Learning. Chem Res Toxicol 33(1):239–248. https://doi.org/10.1021/acs.chemrestox.9b00264
Liu L, Fu L, Zhang JW, Wei H, Ye WL, Deng ZK, Zhang L, Cheng Y, Ouyang DF, Cao Q, Cao DS (2019) Three-Level Hepatotoxicity Prediction System Based on Adverse Hepatic Effects. Mol Pharm 16(1):393–408. https://doi.org/10.1021/acs.molpharmaceut.8b01048
Li X, Chen YJ, Song XR, Zhang Y, Li HH, Zhao Y (2018) The development and application of in silico models for drug induced liver injury. RSC Adv 8(15):8101–8111. https://doi.org/10.1039/c7ra12957b
Mora JR, Marrero-Ponce Y, García-Jacas CR, Suarez Causado A (2020) Ensemble Models Based on QuBiLS-MAS Features and Shallow Learning for the Prediction of Drug-Induced Liver Toxicity: Improving Deep Learning and Traditional Approaches. Chem Res Toxicol 33(7):1855–1873. https://doi.org/10.1021/acs.chemrestox.0c00030
Wang Y, Chen X (2021) Joint Decision-Making Model Based on Consensus Modeling Technology for the Prediction of Drug-Induced Liver Injury. J Chem 2021:2293871. https://doi.org/10.1155/2021/2293871
Li T, Tong W, Roberts R, Liu Z, Thakkar S (2021) Chem Res Toxicol 34(2):550–565. https://doi.org/10.1021/acs.chemrestox.0c00374. DeepDILI: Deep Learning-Powered Drug-Induced Liver Injury Prediction Using Model-Level Representation
Nguyen-Vo T-H, Nguyen L, Do N, Le PH, Nguyen T-N, Nguyen BP, Le L (2020) Predicting Drug-Induced Liver Injury Using Convolutional Neural Network and Molecular Fingerprint-Embedded Features. ACS Omega 5(39):25432–25439. https://doi.org/10.1021/acsomega.0c03866
Hwang D, Jeon M, Kang J (2020) A Drug-Induced Liver Injury Prediction Model using Transcriptional Response Data with Graph Neural Network, IEEE International Conference on Big Data and Smart Computing (BigComp), 19–22 Feb. 2020; 2020; pp 323–329
Hamilton WL, Leskovec RYJ (2017) Inductive Representation Learning on Large Graphs. arXiv ; Vol. 1706.02216.
Chen Z, Jiang Y, Zhang X, Zheng R, Qiu R, Sun Y, Zhao C, Shang H (2021) ResNet18DNN: prediction approach of drug-induced liver injury by deep neural network with ResNet18. Brief Bioinform 23(1). https://doi.org/10.1093/bib/bbab503
Kotsampasakou E, Montanari F, Ecker GF (2017) Predicting drug-induced liver injury: The importance of data curation. Toxicology 389:139–145. https://doi.org/10.1016/j.tox.2017.06.003
Rao M, Nassiri V, Alhambra C, Snoeys J, Van Goethem F, Irrechukwu O, Aleo MD, Geys H, Mitra K, Will Y (2023) AI/ML Models to Predict the Severity of Drug-Induced Liver Injury for Small Molecules. Chem Res Toxicol 36(7):1129–1139. https://doi.org/10.1021/acs.chemrestox.3c00098
O'Brien PJ, Irwin W, Diaz D, Howard-Cofield E, Krejsa CM, Slaughter MR, Gao B, Kaludercic N, Angeline A, Bernardi P, Brain P, Hougham C (2006) High concordance of drug-induced human hepatotoxicity with in vitro cytotoxicity measured in a novel cell-based model using high content screening. Arch Toxicol 80(9):580–604. https://doi.org/10.1007/s00204-006-0091-3
Rodgers AD, Zhu H, Fourches D, Rusyn I, Tropsha A (2010) Modeling Liver-Related Adverse Effects of Drugs Using kNearest Neighbor Quantitative Structure Activity Relationship Method. Chem Res Toxicol 23(4):724–732. https://doi.org/10.1021/tx900451r
Fourches D, Barnes JC, Day NC, Bradley P, Reed JZ, Tropsha A (2010) Cheminformatics Analysis of Assertions Mined from Literature That Describe Drug-Induced Liver Injury in Different Species. Chem Res Toxicol 23(1):171–183. https://doi.org/10.1021/tx900326k
Greene N, Fisk L, Naven RT, Note RR, Patel ML, Pelletier DJ (2010) Developing Structure-Activity Relationships for the Prediction of Hepatotoxicity. Chem Res Toxicol 23(7):1215–1222. https://doi.org/10.1021/tx1000865
Ekins S, Williams AJ, Xu JHJ (2010) A Predictive Ligand-Based Bayesian Model for Human Drug-Induced Liver Injury. Drug Metab Dispos 38(12):2302–2308. https://doi.org/10.1124/dmd.110.035113
Chen MJ, Vijay V, Shi Q, Liu ZC, Fang H, Tong WD (2011) FDA-approved drug labeling for the study of drug-induced liver injury. Drug Discovery Today 16(15–16):697–703. https://doi.org/10.1016/j.drudis.2011.05.007
Liu ZC, Shi Q, Ding D, Kelly R, Fang H, Tong WD (2011) Translating Clinical Findings into Knowledge in Drug Safety Evaluation - Drug Induced Liver Injury Prediction System (DILIps). PLoS Comput Biol 7(12):1002310. https://doi.org/10.1371/journal.pcbi.1002310
Zhu X, Kruhlak NL (2014) Construction and analysis of a human hepatotoxicity database suitable for QSAR modeling using post-market safety data. Toxicology 321:62–72. https://doi.org/10.1016/j.tox.2014.03.009
Liu RF, Yu XP, Wallqvist A (2015) Data-driven identification of structural alerts for mitigating the risk of drug-induced human liver injuries. J Cheminform 7:4. https://doi.org/10.1186/s13321-015-0053-y
Liew CY, Lim YC, Yap CW (2011) Mixed learning algorithms and features ensemble in hepatotoxicity prediction. J Comput Aided Mol Des 25(9):855–871. https://doi.org/10.1007/s10822-011-9468-3
Mulliner D, Schmidt F, Stolte M, Spirkl HP, Czich A, Amberg A (2016) Computational Models for Human and Animal Hepatotoxicity with a Global Application Scope. Chem Res Toxicol 29(5):757–767. https://doi.org/10.1021/acs.chemrestox.5b00465
Chen MJ, Suzuki A, Thakkar S, Yu K, Hu CC, Tong WD (2016) DILIrank: the largest reference drug list ranked by the risk for developing drug-induced liver injury in humans. Drug Discovery Today 21(4):648–653. https://doi.org/10.1016/j.drudis.2016.02.015
Bush BL, Bayly CI, Halgren TA (1998) Consensus bond-charge increments fitted to ab-initio electrostatic potentials of the MMFF94 training set. Abstracts Papers Am Chem Soc 216:U702–U702
Discovery Studio 3.1 Guide (2012) Accelrys Inc., San Diego, http://www.accelrys.com
Shen MY, Tian S, Li YY, Li Q, Xu XJ, Wang JM, Hou TJ (2012) Drug-likeness analysis of traditional Chinese medicines: 1. property distributions of drug-like compounds, non-drug-like compounds and natural compounds from traditional Chinese medicines. J Cheminform 4:31. https://doi.org/10.1186/1758-2946-4-31
Shi HL, Tian S, Li YY, Li D, Yu HD, Zhen XC, Hou TJ (2015) Absorption, Distribution, Metabolism, Excretion, and Toxicity Evaluation in Drug Discovery. 14. Prediction of Human Pregnane X Receptor Activators by Using Naive Bayesian Classification Technique. Chem Res Toxicol 28(1):116–125. https://doi.org/10.1021/tx500389q
Tian S, Li YY, Wang JM, Xu XJ, Xu L, Wang XH, Chen L, Hou TJ (2013) Drug-likeness analysis of traditional Chinese medicines: 2. Characterization of scaffold architectures for drug-like compounds, non-drug-like compounds, and natural compounds from traditional Chinese medicines. J Cheminform 5:5. https://doi.org/10.1186/1758-2946-5-5
Tian S, Wang J, Li Y, Li D, Xu L, Hou T (2015) The application of in silico drug-likeness predictions in pharmaceutical research. Adv Drug Deliv Rev 86:2–10. https://doi.org/10.1016/j.addr.2015.01.009
Tian S, Wang J, Li Y, Xu X, Hou T (2012) Drug-likeness Analysis of Traditional Chinese Medicines: Prediction of Drug-likeness Using Machine Learning Approaches. Mol Pharm 9(10):2875–2886. https://doi.org/10.1021/mp300198d
Zhan ZX, Li LL, Tian S, Zhen XC, Li YY (2017) Prediction of chemical biodegradability using computational methods. Mol Simul 43(13–16):1277–1290. https://doi.org/10.1080/08927022.2017.1328556
Lei TL, Sun HY, Kang Y, Zhu F, Liu H, Zhou WF, Wang Z, Li D, Li YY, Hou TJ (2017) ADMET Evaluation in Drug Discovery. 18. Reliable Prediction of Chemical-Induced Urinary Tract Toxicity by Boosting Machine Learning-Approaches. Mol Pharm 14(11):3935–3953. https://doi.org/10.1021/acs.molpharmaceut.7b00631
Leeson PD (2018) Impact of Physicochemical Properties on Dose and Hepatotoxicity of Oral Drugs. Chem Res Toxicol 31(6):494–505. https://doi.org/10.1021/acs.chemrestox.8b00044
Ghose AK, Viswanadhan VN, Wendoloski JJ (1998) Prediction of hydrophobic (lipophilic) properties of small organic molecules using fragmental methods: An analysis of ALOGP and CLOGP methods. J Phys Chem A 102(21):3762–3772. https://doi.org/10.1021/jp980230o
Csizmadia F, TsantiliKakoulidou A, Panderi I, Darvas F (1997) Prediction of distribution coefficient from structure.1. Estimation method. J Pharm Sci 86(7):865–871. https://doi.org/10.1021/js960177k
Tetko IV, Tanchuk VY, Kasheva TN, Villa AEP (2001) Estimation of aqueous solubility of chemical compounds using E-state indices. J Chem Inf Comput Sci 41(6):1488–1493. https://doi.org/10.1021/ci000392t
Ritchie TJ, Macdonald SJF (2014) How drug-like are 'ugly' drugs: do drug-likeness metrics predict ADME behaviour in humans? Drug Discovery Today 19(4):489–495. https://doi.org/10.1016/j.drudis.2014.01.007
Bickerton GR, Paolini GV, Besnard J, Muresan S, Hopkins AL (2012) Quantifying the chemical beauty of drugs. Nat Chem 4(2):90–98. https://doi.org/10.1038/nchem.1243
Rogers D, Hahn M (2010) Extended-Connectivity Fingerprints. J Chem Inf Model 50(5):742–754. https://doi.org/10.1021/ci100050t
Rogers D, Brown RD, Hahn M (2005) Using extended-connectivity fingerprints with Laplacian-modified Bayesian analysis in high-throughput screening follow-up. J BioMol Screen 10(7):682–686. https://doi.org/10.1177/1087057105281365
Holliday JD, Hu CY, Willett P (2002) Grouping of coefficients for the calculation of inter-molecular similarity and dissimilarity using 2D fragment bit-strings. Comb Chem High Throughput Screen 5(2):155–166. https://doi.org/10.2174/1386207024607338
Cruciani G, Pastor M, Guba W (2000) VolSurf: a new tool for the pharmacokinetic optimization of lead compounds. Eur J Pharm Sci 11. https://doi.org/https://doi.org/10.1016/S0928-0987(00)00162-7. :S29-S39
MOE molecular simulation package, Chemical Computing Group Inc (2010) Montreal, Candada, http://www.chemcomp.com
Bemis GW, Murcko MA (1996) The properties of known drugs.1. Molecular frameworks. J Med Chem 39(15):2887–2893. https://doi.org/10.1021/jm9602928
Schuffenhauer A, Ertl P, Roggo S, Wetzel S, Koch MA, Waldmann H (2007) The scaffold tree - Visualization of the scaffold universe by hierarchical scaffold classification. J Chem Inf Model 47(1):47–58. https://doi.org/10.1021/ci600338x
Langdon SR, Brown N, Blagg J (2011) Scaffold Diversity of Exemplified Medicinal Chemistry Space. J Chem Inf Model 51(9):2174–2185. https://doi.org/10.1021/ci2001428
Bemis GW, Murcko MA (1999) Properties of known drugs. 2. Side chains. J Med Chem 42(25):5095–5099. https://doi.org/10.1021/jm9903996
Shneiderman B (1992) TREE VISUALIZATION WITH TREE-MAPS – 2-D SPACE-FILLING APPROACH. Acm Trans Graphics 11(1):92–99. https://doi.org/10.1145/102377.115768
Lei TL, Chen F, Liu H, Sun HY, Kang Y, Li D, Li YY, Hou TJ (2017) ADMET Evaluation in Drug Discovery. Part 17: Development of Quantitative and Qualitative Prediction Models for Chemical-Induced Respiratory Toxicity. Mol Pharm 14(7):2407–2421. https://doi.org/10.1021/acs.molpharmaceut.7b00317
Wang SQ, Sun HY, Liu H, Li D, Li YY, Hou TJ (2016) ADMET Evaluation in Drug Discovery. 16. Predicting hERG Blockers by Combining Multiple Pharmacophores and Machine Learning Approaches. Mol Pharm 13(8):2855–2866. https://doi.org/10.1021/acs.molpharmaceut.6b00471

No competing interests reported.

DILIpositivecompoundsTotal1141.xlsx
DILInegativecompoundsTotal1004.xlsx
image1.jpg
For Table of Contents Use Only Systematic computational prediction of drug-induced liver injury (DILI) can offer valuable insights for prioritizing lead compounds during the early stages of drug development.

Download PDF

Version 1

posted

You are reading this latest preprint version

Prediction of Drug-Induced Liver Injury: From Molecular Physicochemical Properties and Scaffold Architectures to Machine Learning Approaches

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and methods

3. Results and discussion

3.2. DILI prediction based on scaffold features.

4. Conclusions and future perspectives

Declarations

Funding

Author Contribution

Acknowledgement

Data Availability

References

Additional Declarations

Supplementary Files

Status:

Version 1