S2DV: converting SMILES to a drug vector for predicting the activity of anti-HBV small molecules

Abstract

In the past few decades, chronic hepatitis B caused by hepatitis B virus (HBV) has been one of the most serious diseases to human health. The development of innovative systems is essential for preventing the complex pathogenesis of hepatitis B and reducing side effects caused by drugs. HBV inhibitory drugs have been developed through various compounds, and they are often limited by routine experimental screening and delay drug development. More recently, virtual screening of compounds has gradually been used in drug research with strong computational capability and is further applied in anti-HBV drug screening, thus facilitating a reliable drug screening process. However, the lack of structural information in traditional compound analysis is an important hurdle for unsatisfactory efficiency in drug screening. Here, a natural language processing technique was adopted to analyze compound simplified molecular input line entry system strings. By using the targeted optimized word2vec model for pretraining, we can accurately represent the relationship between the compound and its substructure. The machine learning model based on training results can effectively predict the inhibitory effect of compounds on HBV and liver toxicity. The reliability of the model is verified by the results of wet-lab experiments. In addition, a tool has been published to predict potential compounds. Hence, this article provides a new perspective on the prediction of compound properties for anti-HBV drugs that can help improve hepatitis B diagnosis and further develop human health in the future.

Introduction

The design and synthesis of compounds that establish high activity for specific targets is one of the most significant processes in pharmaceutical chemistry. Drug design is the inventive process of finding new therapeutic entities based on high-throughput virtual screening. Despite advances in pharmaceutical chemistry, drug design is still a slow and difficult process for evaluating the pharmacological activity of drug molecules against specific targets. In this study, we use a simplified molecular input line entry system (SMILES) as input to process chemical tasks. Deep learning is a typical method for representation learning that has shown very high performance across many different natural language processing tasks. Supervised classification models for SMILES encoded spatial vectors are used to predict the activity of small molecules to specific targets.

Hepatitis B is a chronic infectious disease caused by hepatitis B virus (HBV). It can lead to serious health problems, including cirrhosis, liver cancer, liver failure and even death. According to statistics, ~350 million people worldwide are chronically infected with HBV, and >600 000 patients die of chronic infection and complications of HBV every year [1].

The virus particles are internalized into the body by endocytosis, and the nucleocapsid is subsequently released into the cytoplasm. The HBV genomic material in the nucleocapsid (i.e. relaxed circular DNA, rcDNA) is transported to the nucleus, and part of the double-stranded virus rcDNA is repaired to form covalently closed circular DNA (cccDNA) using the host DNA repair mechanism. Subsequently, the virus is transcribed and translated into a new nucleocapsid, in which the second generation of viral DNA is synthesized by reverse transcription. Subsequently, mature newborn HBV nucleocapsids containing rcDNA are secreted from host cells [2]. Currently, the main anti-HBV drugs in clinical application are HBV DNA polymerase/reverse transcriptase inhibitor nucleoside analogs [3–5] and the immune modulator interferon [6–9]. The resistance of hepatitis B to drugs is complex, and current antiviral drugs can have unpleasant side effects. In recent years, researchers have synthesized many nonnucleoside active compounds with novel structures, which depict an important class of current anti-HBV therapies. Withdrawal causes and rebound indications usually occur after the discontinuation and/or reduction of these drugs.

Antiviral compounds (AVCs) inhibit the development of viruses in the host cell and are relatively harmless to the host. However, designing safe and effective antiviral drugs is a difficult task due to the high genetic diversity and consequent drug resistance in viruses [10]. Initially, antivirals were discovered using traditional trial-and-error methods [11]. However, the discovery of effective antivirals was a very lengthy process. Later, research on virology helped to identify many target pathways to block viral multiplication [12, 13]. Scientists are now using rational drug design strategies for developing antivirals that target viruses at different stages of their life cycles [14]. To save time and money for discovering a new drug, researchers have widely used various computational methods to screen virtual libraries of compounds before the synthesis and animal testing of chemicals. Among the different approaches, quantitative structure–activity relationships (QSARs) are mostly used [15–17]. In this approach, relationships connecting molecular descriptors and activity are used to predict the properties of other molecules [18]. Molecular descriptors transform the chemical information (structure and linking of groups) of a molecule into simple numbers. QSAR-based virtual screening is an effective computational technique leading toward the identification and design of novel antiviral agents [19]. AVCpred was developed as a web server for the prediction and design of AVCs. In this method, scientists used previously known AVCs against HIV, hepatitis C virus, HBV, human herpesvirus and 26 other viruses with experimentally validated percentage inhibition from ChEMBL—a large-scale bioactivity database for drug discovery [20].

Limited by the conventional drug development process, large-scale drug molecule design, synthesis and anti-HBV activity experiments consume many human and material resources, including time and cost, and decrease the speed of drug development. Artificial Intelligence aided (AI-aided) drug screening prediction models play a substantial role in drug research and development with their advantages of fast speed, low cost and high efficiency in predicting active drug molecules [21–24].

SMILES

The SMILES is a ‘chemical language’ [25] that represents a chemical structure in compact text notation. The application of molecular graph theory for SMILES enables structure specification with the help of a very small and natural grammar. SMILES is also suitable for computer recognition and processing. It is easy to read and can be used by researchers and computers, so it is suitable for designing and producing unique numerical symbols, continuous database retrieval, smooth substructure retrieval and performance prediction models [26].

Extended connectivity fingerprint

Molecular fingerprints are representations of chemical structures originally designed to aid in chemical database substructure searches but later used for analytical tasks such as similarity search, clustering and classification. Molecular fingerprints are commonly used in several areas of drug discovery, including MACCS [27], PubChem [28], Tree Fingerprint [29], MolPrint2D [30, 31], extended connectivity fingerprint (ECFP; [32]), UNITY 2D [28] and MP-MFP [33]. ECFP is a fingerprint methodology specifically designed to capture molecular characteristics associated with molecular activity. ECFP is applied to predict drug activity, but they are not designed for substructure search. A variety of methods, such as similarity search, clustering and virtual filtering, have been used by ECFP to capture molecular features. Since 2000, ECFP has been applied to a wide range of science-related issues [32–34].

Word2vec

Word2vec methodology such as continuous bag-of-words (CBOW) and skip-gram has seen increasing interest in recent years because of its ability to understand word representations using a pair of architectures. CBOW is similar to skip-gram in general. The main difference between the two structures is that CBOW uses context to predict words, whereas skip-gram uses target words to predict context.

There are many gradient schemes for neural networks [35], such as the stochastic gradient descent scheme (SGD), momentum [36] and Adagrad [37]. Current word embedding models use a simple stochastic gradient optimization method, i.e. for different words in the word bag, the gradient equally contributes to each word vector. The most commonly used scheme in word2vec is SGD.

Database

ChEMBL [38] is a free online database developed by EBI (European Bioinformatics Institute) that contains a large amount of binding, function and ADMET (absorption, distribution, metabolism, excretion and toxicity) information on medicinal compounds. These data were regularly extracted manually from major published literature and then further collated and standardized. The database currently contains 2 105 464 compounds, 14 554 targets and 1 383 553 drug activity assays. Through this database, the reported compounds and their activity information of a target can be quickly queried. The data in this database are all from various reported studies, which are reliable and can be traced to the source of the data. Through this database, accurate chemical compounds and their biological data can be obtained quickly, which can further accelerate the speed of drug design and drug development. In this study, potential anti-HBV drugs were screened, and liver toxicity was considered. Therefore, compounds targeted at HBV and HepG2 cells were screened as training data.

Methods

Tokenization

Usually, word segmentation is required before the text is pretrained for the next step. As shown in Figure 1A, for chemical structure coding SMILES, split character (SC) and molecular fingerprint (ECFP) can be used.

Figure 1.

The pipeline overview of S2DV method. Part A: Two tokenization method: ‘SMILES to SC’ and ‘SMILES to ECFP’. Part B: The process of embedding generation.

SMILES to SC

• (i)According to the order of SMILES, the sliding window displacement is cut with a specific window size and the displacement is one character each time.
• (ii)Delete the ``(`` and '')'' used to represent the branch chain at the branch chain, then use a single letter ``R'' to retain the branch chain information.
• (iii)Double characters such as ‘Cl’ and ‘Fe’ indicate a single atom. Replace it with single characters such as ‘L’ and ‘E’ that are not confused with other characters.
• (iv)For each SMILES, a linear word segmentation list is formed by splicing back and forth according to the sequence of the main chain and branch chain.

As shown in Figure 1A, SMILES expressed as ‘Nc1cc(OCCOCP(=O)(O)O) NC (N) N1’ compounds were processed by SMILES to SC to form 17 tokens as follows:

Nc1c/c1cc/1ccR/ccRn/cRnc/RncR/ncRn/cRn1/OCCO/CCOC/COCP/OCPR/CPRR/PRRO/N/=O/O.

SMILES to ECFP

Morgan FP is generated by producing and assigning a unique identifier (Morgan identifier) to all substructures around all heavy atoms in the molecule within a defined radius. These identifiers usually hash vectors of fixed length [39].

For each compound encoded by SMILES, SMILES to ECFP generates identifiers for all atoms of a fixed radius and then arrange the identifiers for each atom into a ‘molecular sentence’ in the order in which the atoms are arranged by SMILES. If an atom has multiple identifiers at the same time, the atomic identifiers are arranged according to their recognition radius from short to long.

Embedding generation

SMILES-encoded texts are made up of different atoms and atomic structures, and various chemical structures, such as atoms, chemical bonds, branched chains, ring structures and ionic bonds, are described through codes. These chemical structures are related to each other and can be distinguished from each other to form the meaning of similar entities. The concept of an entity refers to the object or concept that exists in the objective world and can be distinguished from each other. An Ising model is proposed to explain a phase change of the ferromagnetic material. The magnetic and nonmagnetic shifts between two phases can be the same in text data to construct an entity of sparse matrix. Between entities, there is no change to reflect. After verification, it can also describe the chemical structure that exists, and there is no change in the entity. Different components of the Ising model system are associated with each other over a long range, which is exactly what is needed in the construction of the global correlation matrix through local interaction [40, 41].

The module shown in Figure 1B is a new word2vec model combined with the Ising model design (Ising-word2vec), which is used to flexibly capture global and local connections in the process of generating word embedding. Tokens extracted from SMILES were input to the model, and the corresponding global gradient modified word embedding model was obtained through the model.

At this point, the chemical structure entity vector will no longer evenly distribute the weight to update the embedding but redistribute the combination with to form a new chemical structure entity embedding. Compound chemical structure embedding is repeatedly trained by a negative sampling function, and a stable representation of compound chemical structure embedding is finally obtained as follows:

Step 1: The token data of the compound structure extracted from SMILES were used to construct the sparse matrix according to the Ising model data structure Ising sparse matrix ().

Step 2: The whole correlation matrix () was obtained by large-scale sparse data processing of sparse learning with efficient projections.

In this study, logistic regression is adopted to solve the regularization logic problem of the Ising spares matrix.

(1)

(2)

Formula (1) shows the process of solving , where is the weight of the tth entity in all chemical structure entities, is the tth column of the we input and is the tth column of the kth row extracted from the sparse matrix, which is used to solve the correlation between and other entities. is the solution of the correlation corresponding to , λ is the regularization parameter of the norm and ρ is the regularization parameter of the square 2 norm.

Step 3: To solve the problem of memory crash caused by a large amount of data, the input text data entities are divided into N batches to process data, and the ith batch is taken as Batch_i. The entities in Batch_i are

(3)

Step 4: Input the chemical structure entities of Batch_i   to obtain the initialization of chemical structure entity embedding

(4)

In formula (4), j in is the jth entity in Batch_i, and correspondingly, is the corresponding embedding of the jth entity. is the weight matrix during vector generation.

Step 5: Batch_i was paired to generate a skip-gram relation, and and of each chemical structure were obtained. According to the study, the predicted probability is calculated. ，.

Among them:

(5)

According to formula (5) above, , , σ is the sigmoid activation function and indicates that the working parameters corresponding to are also to be trained.

The loss function formula (6) is calculated according to the prediction probability between word pairs in Batch_i   count context and negative sampling word space

(6)

Step 6: According to the loss function obtained in Equation (6), multiple optimization schemes can be adopted to obtain the gradient to optimize the parameters, and the gradient descent method is adopted for detailed analysis to facilitate understanding.

Might as well set:

(7)

The gradients of (8) and (9) are calculated according to Equation (7):

(8)

(9)

Step 7: Entities in Batch_i are used to construct skip-gram pairing relationships and negative sampling pairing relationships between entities in Step 5. Based on the pairing of entities, the subcorrelation matrix () corresponding to the entity pairing relationship and the negative correlation matrix () corresponding to the negative sampling pairing relationship were obtained from .

Step 8: and are combined with the gradient, and the weights are redistributed instead of distributed evenly. Then, backpropagation updates word embedding , and the corresponding auxiliary vector for word .

(10)

(11)

Evaluation of anti-HBV activity and cytotoxicity

All the synthesized compounds were evaluated for in vitro anti-HBV activity and cytotoxicity in HepG2 2.2.15 cells using real-time quantitative PCR and MTT methods, respectively. The concentration of compound required for 50% inhibition of DNA replication was defined as the IC_50, and the concentration of compound that induced the death of the HepG2 2.2.15 cell cultures by 50% was defined as the CC₅₀ [42–44].

Detailed wet-lab experimental operations are described in the supplementary file.

Results

Embedding generation

In this study, the performance difference of the word2vec model before and after the Ising model was compared.

Dataset collection

To reduce the computational dimension required by training and generate a unique vector space with the chemical structure distribution of specific target compounds to accurately represent the spatial distribution relationship of the chemical structure, all the compounds targeted at HBV and HepG2 were screened to generate compound structure word embedding models for HBV and HepG2. In Table 1, a total of 6705 experimental activity values were obtained for screening HBV as a target, of which 4411 were semi-inhibitory concentration values. A total of 182 550 results were obtained targeting HepG2 cells, of which 2270 were semitoxic concentrations.

Table 1.

ChEMBL database filtering results

Target name	Standard type	Target organism	Assays number
Hepatitis B virus	ALL	Hepatitis B virus	6705
Hepatitis B virus	EC50/IC50	Hepatitis B virus	4411
HepG2	ALL	Homo sapiens	182 550
HepG2	CC50	H. sapiens	2270

Tokenization

Considering the SMILES to SC method and the influence of different lengths on embedding performance, we use different lengths for embedding generation. When the split char length is 1 and 2, <100 tokens are generated, and word2vec cannot be used. However, when the split char length is 5, a large number of SMILES cannot generate a sufficient number of tokens. Therefore, we chose 3 and 4 for comparative experiments. As shown in Table 2, better performance can be obtained when the tokenization character length is 4 than when the token extract character length is 3. Therefore, in the subsequent SMILES to SC method, the extracted length defaults to 4.

Table 2.

Performance comparison of different split char lengths in SMILES with SC

Embedding model	Splitted char length	Score	LR	LDA	KNN	CART	NB	SVM	XG-Boost	RD-Forest
Ising-word2vec	4	Accuracy	0.790	0.788	0.794	0.736	0.706	0.854	0.849	0.834
		Precision	0.698	0.701	0.655	0.597	0.544	0.832	0.827	0.883
		F1	0.668	0.661	0.718	0.606	0.599	0.760	0.750	0.696
		Recall	0.641	0.627	0.796	0.616	0.666	0.699	0.687	0.575
	3	Accuracy	0.776	0.771	0.802	0.728	0.686	0.844	0.845	0.838
		Precision	0.691	0.687	0.679	0.595	0.531	0.826	0.819	0.882
		F1	0.654	0.641	0.732	0.610	0.584	0.750	0.753	0.716
		Recall	0.621	0.602	0.795	0.626	0.650	0.687	0.697	0.603
word2vec	4	Accuracy	0.786	0.784	0.798	0.721	0.709	0.851	0.848	0.835
		Precision	0.692	0.693	0.665	0.573	0.548	0.831	0.826	0.887
		F1	0.661	0.655	0.720	0.587	0.603	0.754	0.749	0.696
		Recall	0.634	0.621	0.786	0.603	0.671	0.690	0.685	0.573
	3	Accuracy	0.777	0.772	0.799	0.736	0.693	0.844	0.847	0.839
		Precision	0.690	0.689	0.676	0.610	0.541	0.822	0.826	0.881
		F1	0.657	0.644	0.726	0.617	0.586	0.750	0.753	0.720
		Recall	0.627	0.604	0.785	0.624	0.639	0.690	0.698	0.609

For tokenization of the ECFP model, we took HBV as an example to compare the model classification performance under different ECFP sampling radii. As shown in Table 3, in the word2vec and Ising-word2vec models, the sampling radius of 1 is not significantly inferior to the larger sampling radius (Radius = 2 or 3). Considering that fewer tokens are generated when the sampling radius is lower (when Radius is 1, 2, 3, n-token is 1474, 7187, 17 668, respectively), which avoids the generation of an overly sparse matrix and greatly improves the calculation requirements. The default sampling radius is 1 in subsequent experiments.

Table 3.

Comparison of classification effects of ECFP models with different sampling radii

Embedding method	Radius	N-Token	Model	Accuracy	Precision	Recall	F1
Ising-word2vec	1	1474	LR	0.825	0.827	0.830	0.828
			LDA	0.830	0.835	0.829	0.832
			KNN	0.823	0.824	0.830	0.826
			CART	0.810	0.825	0.795	0.810
			NB	0.699	0.653	0.873	0.747
			SVM	0.832	0.830	0.842	0.836
			XGBoost	0.835	0.825	0.858	0.841
			RDForest	0.837	0.820	0.871	0.845
	2	7187	LR	0.829	0.826	0.841	0.833
			LDA	0.834	0.837	0.837	0.837
			KNN	0.829	0.820	0.851	0.834
			CART	0.813	0.830	0.794	0.812
			NB	0.690	0.646	0.866	0.740
			SVM	0.836	0.837	0.841	0.839
			XGBoost	0.841	0.833	0.859	0.846
			RDForest	0.834	0.819	0.865	0.841
	3	17 668	LR	0.816	0.818	0.822	0.819
			LDA	0.825	0.833	0.821	0.826
			KNN	0.830	0.830	0.839	0.834
			CART	0.805	0.823	0.786	0.804
			NB	0.695	0.653	0.857	0.741
			SVM	0.841	0.843	0.845	0.844
			XGBoost	0.833	0.823	0.857	0.839
			RDForest	0.839	0.824	0.868	0.846
word2vec	1	1474	LR	0.823	0.824	0.829	0.826
			LDA	0.823	0.831	0.818	0.824
			KNN	0.828	0.825	0.841	0.833
			CART	0.815	0.831	0.799	0.815
			NB	0.696	0.649	0.878	0.746
			SVM	0.830	0.829	0.841	0.834
			XGBoost	0.836	0.826	0.860	0.842
			RDForest	0.837	0.821	0.868	0.844
	2	7187	LR	0.816	0.816	0.824	0.819
			LDA	0.820	0.825	0.822	0.823
			KNN	0.828	0.826	0.839	0.832
			CART	0.807	0.828	0.783	0.805
			NB	0.701	0.654	0.875	0.748
			SVM	0.839	0.841	0.846	0.843
			XGBoost	0.841	0.831	0.864	0.847
			RDForest	0.836	0.819	0.870	0.844
	3	17 668	LR	0.816	0.816	0.824	0.820
			LDA	0.817	0.819	0.821	0.820
			KNN	0.825	0.826	0.832	0.828
			CART	0.807	0.824	0.790	0.806
			NB	0.702	0.662	0.849	0.744
			SVM	0.838	0.837	0.847	0.842
			XGBoost	0.832	0.821	0.856	0.838
			RDForest	0.838	0.821	0.871	0.845

ChEMBL dataset validation

To verify whether anti-HBV compounds have potential research prospects, it is necessary to verify their anti-HBV ability (IC₅₀) and hepatocellular toxicity (CC₅₀).

The embedding model was verified in 4411 HBV inhibition tests and 2270 HepG2 toxicity tests. Eight machine learning models were used to construct compound classification models. In the initial biological data prediction of compounds targeting HBV and HepG2 cells, the threshold value of the semieffective inhibitory concentration of HBV (IC₅₀/EC₅₀) was set at 1 μM. The threshold of the HepG2 semitoxic concentration (CC₅₀) was set at 100 μM to convert the activity data of the compound against the target into binary data for prediction. The experimental results with different units in the dataset were all converted in μM. The model evaluation parameters were obtained using tenfold cross-validation.

ChEMBL (https://chembl.gitbook.io/chembl-interface-documentation/downloads) was used for the latest version of the drug activity database. Filter fixed targets based on fields such as target names and target organizations. Considering that the inhibition rate of HBV (IC₅₀ and EC₅₀) has the same effect, we reserved the EC50 and IC₅₀ data for training and testing of the classification model to reserve as many samples as possible.

HBV inhibition rate (IC₅₀)

Tables 4 and 5 show the comparison of the classification performance of the word2vec model and Ising-word2vec model in the two tokenization methods. Table 4 uses the split SMILES to generate tokens, and Table 5 uses the tokens generated by Morgan’s algorithm. Conclusion word2vec combined with the Ising model has a better classification effect than the original word2vec in predicting molecular activity, and RandomForest is the best classifier. Comparing the two different tokenization methods, SMILES to SC (Accuracy: 0.841, Precision: 0.843, Recall: 0.881, F1: 0.853 and AUC:0.92) has better classification results than ECFP generated by Morgan algorithm (Accuracy: 0.837, Precision: 0.835, Recall: 0.878, F1: 0.845 and AUC: 0.90). This may be because the split SMILES method has more advantages than Morgan’s algorithm in representation of the chemical structures of rings and branched chains.

Table 4.

Prediction result of inhibition (IC₅₀) using SMILES to ECFP

Embedding method	Classification	Accuracy	Precision	Recall	F1
word2vec	LR	0.823	0.824	0.829	0.826
	LDA	0.823	0.831	0.818	0.824
	KNN	0.828	0.825	0.841	0.833
	CART	0.815	0.831	0.799	0.815
	NB	0.696	0.649	0.878	0.746
	SVM	0.830	0.829	0.841	0.834
	XGBoost	0.836	0.826	0.860	0.842
	RDForest	0.837	0.821	0.868	0.844
Ising-word2vec	LR	0.825	0.827	0.830	0.828
	LDA	0.830	0.835	0.829	0.832
	KNN	0.823	0.824	0.830	0.826
	CART	0.810	0.825	0.795	0.810
	NB	0.699	0.653	0.873	0.747
	SVM	0.832	0.830	0.842	0.836
	XGBoost	0.835	0.825	0.858	0.841
	RDForest	0.837	0.820	0.871	0.845

Table 5.

Prediction result of inhibition (IC₅₀) using SMILES to SC

Embedding method	Classification	Accuracy	Precision	Recall	F1
word2vec	LR	0.824	0.834	0.830	0.832
	LDA	0.825	0.835	0.831	0.833
	KNN	0.825	0.828	0.841	0.834
	CART	0.812	0.834	0.800	0.817
	NB	0.688	0.665	0.815	0.732
	SVM	0.838	0.843	0.849	0.846
	XGBoost	0.840	0.833	0.870	0.851
	RDForest	0.838	0.824	0.881	0.851
Ising-word2vec	LR	0.818	0.825	0.829	0.826
	LDA	0.811	0.825	0.812	0.818
	KNN	0.831	0.836	0.845	0.840
	CART	0.809	0.835	0.794	0.813
	NB	0.681	0.660	0.807	0.726
	SVM	0.836	0.842	0.847	0.844
	XGBoost	0.837	0.831	0.867	0.848
	RDForest	0.841	0.828	0.881	0.853

Figure 2 shows that when using ECFP as a tokenization method, both word2vec and Ising-word2vec will yield the same AUC score (0.90). In Figure 3, the AUC score of Ising-word2vec (0.92) is better than word2vec (0.91). Therefore, in the calculation of HBV IC₅₀, word Ising-word2vec can lead to a better classification model.

Figure 2.

The ROC curve of prediction model with using ‘SMILES to ECFP’ for tokenization to predict inhibitory effect on HBV (IC50). (A) Embedding generated by word2vec. (B) Embedding generated by Ising-word2vec.

Figure 3.

The ROC curve of prediction model with using ‘SMILES to SC’ for tokenization to predict inhibitory effect on HBV (IC50). (A) Embedding generated by word2vec. (B) Embedding generated by Isingword2vec.

Toxicity of HepG2 (CC₅₀)

We used the same method to train the classification model and obtain conclusion statistics for the toxicity of HepG2 cells. In Tables 6 and 7, we can see that SMILES to SC and SMILES to ECFP were used for tokenization directly, and the Ising-word2vec model with the addition of SMILES to SC produced a slight improvement in performance compared with the direct use of word2vec (SMILES to SC: Accuracy = 0.902, Precision = 0.908, Recall = 0.972, F1 = 0.935, AUC = 0.91;SMILES to ECFP: Accuracy = 0.912, Precision = 0.974, Recall = 0.973, F1 = 0.943 and AUC = 0.94). Moreover, the AUC value is also excellent, which means that the expressions of SMILES can be accurately learned from the Ising model by adding the word2vec model. The Ising-word2vec model can be used to train and accurately predict the drug-like properties and other parameters of a variety of compounds, which will help pharmaceutical personnel avoid compounds with no medicinal potential in advance during drug screening, reduce drug screening input and improve drug screening efficiency. In Figure 4, the Ising-word2vec model yielded the highest AUC value (0.94). In Figure 5, the word2vec model yielded the highest AUC value (0.93). This seems to indicate that a higher classification effect can be achieved when different embedding methods are selected.

Table 6.

Prediction result of toxicity (CC₅₀) using SMILES to ECFP

Embedding method	Classification	Accuracy	Precision	Recall	F1
word2vec	LR	0.879	0.913	0.929	0.921
	LDA	0.874	0.916	0.918	0.917
	KNN	0.886	0.911	0.942	0.926
	CART	0.858	0.915	0.897	0.906
	NB	0.710	0.833	0.772	0.801
	SVM	0.893	0.893	0.977	0.933
	XGBoost	0.912	0.918	0.971	0.943
	RDForest	0.907	0.911	0.972	0.940
Ising-word2vec	LR	0.888	0.914	0.941	0.927
	LDA	0.883	0.921	0.925	0.923
	KNN	0.884	0.910	0.940	0.924
	CART	0.861	0.912	0.904	0.908
	NB	0.710	0.832	0.774	0.802
	SVM	0.889	0.890	0.975	0.930
	XGBoost	0.912	0.916	0.971	0.943
	RDForest	0.904	0.907	0.973	0.939

Table 7.

Prediction result of toxicity (CC₅₀) using SMILES to SC

Embedding method	Classification	Accuracy	Precision	Recall	F1
word2vec	LR	0.870	0.908	0.917	0.912
	LDA	0.863	0.904	0.911	0.907
	KNN	0.875	0.900	0.935	0.917
	CART	0.835	0.892	0.885	0.888
	NB	0.692	0.819	0.750	0.782
	SVM	0.892	0.892	0.972	0.930
	XGBoost	0.900	0.907	0.965	0.934
	RDForest	0.894	0.898	0.967	0.931
Ising-word2vec	LR	0.869	0.908	0.916	0.912
	LDA	0.859	0.900	0.911	0.905
	KNN	0.876	0.899	0.939	0.918
	CART	0.816	0.884	0.866	0.874
	NB	0.692	0.817	0.753	0.783
	SVM	0.892	0.892	0.972	0.930
	XGBoost	0.902	0.907	0.966	0.935
	RDForest	0.897	0.899	0.971	0.933

Figure 4.

The ROC curve of prediction model with using ‘SMILES to ECFP’ for tokenization to predict liver toxicity (CC50). (A) Embedding generated by word2vec. (B) Embedding generated by Ising-word2vec.

Figure 5.

The ROC curve of prediction model with using ‘SMILES to SC’ for tokenization to predict liver toxicity (CC50). (A) Embedding generated by word2vec. (B) Embedding generated by Ising-word2vec.

In vitro validation

To verify the reliability of the model, we used in vitro cell experiments to test the prediction model of the compound inhibition rate of HBV and the prediction model of the compound exclusivity of HepG2 cells. Throughout the trial, we tested data on the activity of 56 compounds, including HBV half-inhibitory concentration (IC₅₀ and EC₅₀) and HepG2 half-toxic concentration (CC₅₀).

After unifying the units of data, SMILES was input into the model corresponding to the compound, and the binary classification results of the compound for the specified threshold were obtained. A post prediction confusion matrix was established to calculate the accuracy, precision, recall and F1 value. The ROC curve was drawn to calculate the AUC value of the area under the curve.

In compound activity tests, the evaluation criteria for compound activity are not fixed values, and a real-time response threshold change interval should be formed according to the overall activity distribution of the current set of compounds to be tested. In the drug activity test, we tried the HBV IC₅₀ threshold from 1 to 20 μM and the HepG2 CC₅₀ threshold from 10 to 100 μM.

To establish a relative threshold ratio between compounds to IC₅₀ and CC₅₀, IC₅₀ and CC₅₀ can maintain a large concentration difference to maintain a reasonable active toxicity ratio, which is entirely dependent on the screening needs of drug testers. In this study, we focused more on the drug’s ability to inhibit HBV. Therefore, the IC₅₀ threshold was set on a low level. In contrast, the toxicity threshold for HepG2 cells was set loosely at 30 μM.

Tables 8 and 9 show the validation results of the prediction model for HBV drug IC₅₀ with a threshold value of 1 μM and HepG2 drug CC₅₀ with a threshold value of 30 μM, respectively. In the in vitro anti-HBV experiment, the model still obtained relatively high prediction accuracy under a specific threshold value. When using the XGBoost classification model, the highest accuracy of IC₅₀ is 0.732, and the highest AUC value is 0.962. When using the SVM classification model, the highest accuracy of CC₅₀ is 0.893, and the highest AUC is 0.943. The results for other thresholds of IC₅₀ and CC₅₀ are described in the supplementary file (Supplementary Table S1 and S2).

Table 8.

HBV Drug experimental compound model prediction validation

Threshold	Train set (Pos/Neg)	Test set (Pos/Neg)	Model	Accuracy	AUC
1	3582/829	37/19	LR	0.661	0.463
			LDA	0.643	0.736
			KNN	0.679	0.718
			CART	0.714	0.579
			NB	0.536	0.393
			SVM	0.661	0.748
			XGBoost	0.732	0.962
			RDForest	0.714	0.927

Table 9.

HepG2 drug experimental compound model prediction validation

Threshold	Train set (Pos/Neg)	Test set (Pos/Neg)	Model	Accuracy	AUC
30	1502/768	50/6	LR	0.893	0.427
			LDA	0.875	0.800
			KNN	0.750	0.517
			CART	0.804	0.722
			NB	0.321	0.558
			SVM	0.893	0.943
			XGBoost	0.821	0.763
			RDForest	0.875	0.360

Discussion and conclusion

In this work, we propose a word2vec model combined with Ising gradient correction, which shows better performance than traditional word2vec in different target datasets. In addition, in the downstream task after vector training, we have a definite advantage in predicting the compound’s HBV inhibition rate and hepatocyte toxicity. This method has a good ability to screen potential anti-HBV drugs. In addition, in the process of generating tokens, a variety of methods were compared to try to find token extraction parameters that could achieve the optimal performance of the model. We not only demonstrate the advantages of the model for compound representation learning in a public dataset but also demonstrate the predictive power of the classification model based on Ising-word2vec pretraining results in cell experiments. We demonstrate the potential of word2vec combined with the Ising gradient correction model for screening new potential drug-like compounds. This method can be widely applied to predict the drug-like properties of other compounds with different targets, thus simplifying the drug development process.

In Figure 6, tool S2DV is published for anti-HBV drug screening by predicting inhibitory effect on HBV (IC₅₀) and liver toxicity (CC₅₀) of potential compounds on the web (http://www.vectorspaceai.cn/S2DV/home).

Figure 6.

Screen capture of the web tool driven by S2DV.

Since our experiment demonstrated the superior performance of word2vec combined with Ising gradient modification for drug compound classification tasks, we also wanted to explore the universality of word2vec combined with Ising gradient modification in different fields. Not only can it effectively improve the performance of models, but its inherent global gradient correction mechanism can also provide many interesting insights.

The scheme still has shortcomings: from the perspective of the model algorithm itself, the model combines local and global, but the global matrix is not optimized with vector time; i.e. the global relational matrix is not completely correct in correcting the gradient. Second, from the perspective of model training and testing, the data amount is limited in the training and verification process of compound medicinal properties. Most of the drugs collected in the database are reported to have higher target activity, and a large number of negative samples are not needed to train a more accurate model, which means that more data of the same target test results are needed to comprehensively optimize the model in later work. In addition, the model only focuses on the compound activity association of a certain target drug and does not make in-depth use of the compound activity data of similar targets and related targets. Therefore, the application of the model needs to be extended to more target data for overall optimization and improvement.

Key Points

• By using the inherent global relationship between molecular SMILES strings, the Ising model can effectively improve the performance of the gradient correction mechanism.

• S2DV demonstrates the superior performance of Ising-word2vec for drug compound inhibition and toxicity classification. The generated molecular embeddings can be reused through GitHub.

• S2DV was developed as an online tool to predict the inhibitory effects of potential compounds on HBV (IC₅₀) and liver toxicity (CC₅₀).

Jinsong Shao is currently a graduated master student at the School of Public Health, Nantong University. His research focus on leveraging artificial intelligence for drug discovery.

Qineng Gong is currently a PhD student at Department of Medicinal Chemistry, School of Pharmacy, Fudan University. His research focus on antitumor drug molecule design.

Zeyu Yin is currently a master student at the School of Information Science and Technology, Nantong University, Nantong, China. His research focus on natural language processing.

Wenjie Pan is an undergraduate at the department of medical informatics, Nantong University. Her main research interests include text mining and drug discovery.

Sanjeevi Pandiyan is an associate professor of the research center for Intelligence Information Technology, Nantong University. His main research interest is medical artificial intelligence.

Li Wang is currently a professor of the research center for Intelligence Information Technology, Nantong University. His main research interests include medical nature language processing and data mining.

Funding

National Science Foundation of China (No.81873915).

Data availability statement

All datasets and codes used in this study are available at GitHub: https://github.com/NTU-MedAI/S2DV.