IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 13, No. 4, December 2024, pp. 4380 ∼ 4387 ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i4.pp4380-4387 ❒ 4380 Character N-gram model f or toxicity pr ediction Eman Shehab 1 , Hamada Nay el 2 , Mohamed T aha 2 1 Department of Computer Science, F aculty of Computers and Articial Intelligence, Uni v ersity of Sadat City , Sadat, Egypt 2 Department of Computer Science, F aculty of Computers and Articial Intelligence, Benha Uni v ersity , Benha, Egypt Article Inf o Article history: Recei v ed Jan 24, 2024 Re vised Mar 24, 2024 Accepted Apr 17, 2024 K eyw ords: Feature e xtraction Machine learning Molecular toxicity prediction N-gram Simplied molecular -input line-entry system ABSTRA CT Molecular toxicity prediction is a crucial step in the drug disco v ery process. It has a direct relationship with human heal th and medical destin y . Accurately assessing a molecule’ s toxicity can aid in the weeding out of lo w-quality com- pounds early in the drug disco v ery phase, a v oiding depletion later in the drug de v elopment process. Computational models ha v e been used automatically for molecular toxicity prediction. In this paper , a machine learning-based model has been proposed. TF/IDF representation scheme has been used for N-gram and inte grated with simplied molecular -input line-entry system (SMILES). Multi- ple machine learning classiers such as logistic re gression (LR), support v ector machine (SVM), random forest (RF), decision tree (DT), k-nearest neighbors (KNN), AdaBoost, multi-layer perceptron (MLP), and st ochastic gradient de- scent (SGD) classiers ha v e been implemented. A wide range of N-gram mod- els ha v e been implemented and trigram reported the best resul ts. RF and SVM achie v ed 85% and 84% accurac y respecti v ely . Comparable to state-of-the-art models, our results are acceptable as we used minimum a v ailable resources. This is an open access article under the CC BY -SA license . Corresponding A uthor: Eman Shehab Department of Computer Science, F aculty of Computers and Articial Intelligence, Uni v ersity of Sadat City Sadat, Egypt Email:eman.shehab@fcai.usc.edu.e g 1. INTR ODUCTION Numerous therapeutic candidates ha v e recently , f ailed in late-stage clinical trials [1]. Chemical respi- ratory toxicity i s a primary cause of clinical trial f ailures, and it has also resulted in the withdra w al of numerous medications from the mark et [2], [3]. Since there are no kno wn adv erse ef fects of medicinal products in human respiratory systems [4], [5], it is important to analyze potential for respiratory toxici ty from compounds as soon as possible during drug disco v ery . Rapid screening of drug candidates is made possible by accurately anticipating the properties of drug molecules, which helps to sa v e both time and mone y . Pharmacokinetic char - acteristics absorption, distrib ution, metabolism, e xcretion, and toxicity (ADMET) are major concerns during the screening stage of drug candidate [6], [7]. At an early stage of the drug de v elopment process, the ADMET property assessment approach can ef fecti v ely address the issue of species dif ferences, signicant ly increase the success rate and reduce the cost of drug disco v ery . The process of bringing F ood and Drug Administration (FD A) medicine to mark et requires more than ten years and $200 million [8], [9]. The main cause of such high costs is medication safety , which accounts for 96% of therapeutic f ailures [10]. In the nal stages of drug research, drug toxicity , and side ef fects are a crucial practiced challenge [11]–[14]. Therefore, to a v oid high-cost consumption, predicting molecular toxicity should be performed as soon as possible during the de v elopment stage of a drug. Machine learning techniques are becoming v ery popular in the pharmaceutical sector , which mak es it J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4381 possible to analyze a great deal of data a v ailable f ast er and easier [15]. Class ication, re gression, clustering, and pattern recognition are som e of the main tasks carried out by articial intelligence algorithms in a wide data set. Ne w molecular characteristics, interactions, biological acti vities, and side ef fects of medicines are predicted by a wide v ariety of machine learning methods in the pharmaceutical industry [16]–[19]. According to W ang et al. [20], for predicting the chemical respiratory toxicity , six machine learning methods with nine types of molecular ngerprints ha v e been used. According to Peng et al. [21], a no v el method of molecular representation and de v eloped the corresponding deep neural netw ork frame w ork, inte- grates designed data preprocessing techniques, a recurrent neural netw ork (RNN) based on the bidirectional g ated recurrent unit and fully connected neural netw orks for end-to-end molecular representation learning and chemical toxicity prediction. F or predicting chemical toxicity , a graph con v olution neural netw ork ha v e been de v eloped and trains by mean teacher algorithm, based on the success of semi-supervised learning (SSL) algorithm [22]. According to Jag anathan et al. [23], using machine learning algorithms, and systematic tool se- lection methods to select features of the molecular describer sets, authors set up quantitati v e structure-acti vity relationship models. According to Huang [24], dif ferent con v e n t ional machine learning and deep learning algorithms ha v e been applied to an online chemica l database and model en vironment, creating a series of com- putational models. According to Feng et al. [25], to predict the compound’ s reproducti v e toxicity , ensembles of learning models ha v e been de v eloped using 9 molecular ngerprints, random forest (RF), e xtreme gradient boosting methods and support v ector machine (SVM). Zhang et al. [26] studied a combination of deep neural netw orks with predictors based on the coherence forecasting frame w ork in order to generate high probability models with clear uncertainties. Hua et al. [27] focusing on using machine learning and deep learning tech- niques to predict chemically induced hematotoxicity in silico, using QNPR descriptors and the random forest re gression (RFR) and classication method. The rest of this paper is sho wn as follo ws. Section 2 presented materials and data preparat ion, section 3 introduced method, section 4 presented results and discussion. Finally , the conclusion is presented in section 5. 2. MA TERIALS AND D A T A PREP ARA TION 2.1. Dataset The dataset used in this study w as collected by g athering positi v e data from the three datasets [20]. The rst database is the Pneumotox database, which includes medicines that cause respiratory disorders [28]. Adv erse drug reaction classication system (ADReCS) database that contains a wide range of undesirable reactions [29]. W e’ v e been focusing on the ne g ati v e ef fects of medicines on the respiratory system. From the hazardous chemical information system, we selected substances that ha v e a detrimental ef fect on respiratory systems (published on M ay 9, 2018). In addition, from the rele v ant literature we ha v e obtained positi v ely and ne g ati v ely char ged chemicals [30]–[35]. In the ChemIDplus database, all chemicals were matched to simplied molecular -input line-entry system (SMILES) [36]. 2.2. Molecular r epr esentation In this section, we discussed the most common techniques that are used for molecule repres entation, string representation and molecular graphs. String representations, with a lar ge selection of sequence modeling techniques, e.g. RNN, attention mechanisms, and dilated con v olutions, ha v e been quickly adopted for gener - ati v e models that represent chemical structures as a string. SMILES is the most often used string encoding for generati v e machine learning model s [37]. The SMILES technique preserv es atom and edge tok ens while tra v ersing the spanning tree of a chemical graph in depth-rst order . Specic tok ens for branching and edges that are not part of a spanning tree shall also be used by SMILES. Because ther e are multiple spanning trees on a molecule, numerous SMILES strings coul d be representing the same molecule. Is it possible to create the SMILES string uniquely from a molecule, its ambiguity can also enrich and impro v e generati v e models [38]. Molecular graphs, graph representations ha v e been used in chemoinformatics for a long time to st ore and analyze biological data. Each node on the molecular graph represents the atom, and all edges represent a connection. In such a graph, the h ydrogens can be specied directly or implicitly . The number of h ydrogens can then be calculated based on their atomic v alue in this case. Char acter N-gr am model for toxicity pr ediction (Eman Shehab) Evaluation Warning : The document was created with Spire.PDF for Python.
4382 ❒ ISSN: 2252-8938 3. METHOD Figure 1 represents the general structure of the proposed model. It comprises of six phases: dataset g athering, data pre-proces sing, feature e xtraction, data splitting, learning models and e v aluation. Details of dataset g athering and pre-processing aforement ioned. The feature e xtraction phase has been implemented using the ngram term frequenc y-in v er se document frequenc y (TF-IDF) based model. The dataset must be ran- domly split into train and test sets with a ratio of 8:2 respecti v ely . Logistic re gression (LR), SVM, RF , decision tree (DT), k-nearest neighbors (KNN), AdaBoost, multi-layer perceptron (MLP), and stochastic gradient de- scent (SGD) classiers ha v e been used to train the proposed model. The last phase is the e v aluation of the performance of the proposed model. These procedures were put into practice in scikit-learn 0.24.1 [39]. Figure 1. Sho ws the proposed model for predicting the chemical respiratory toxicity 3.1. F eatur e extraction In machine learning algorithms, the TF-IDF with n-gram pro vides numerical weights to te xtual con- tent for mi ning. It measures ho w important a te rm is within a compound relati v e to a col lection of com- pounds. Choosing the right feature parameters can help to increase the classication accurac y of predecting toxicity [40]. 3.2. Logistic r egr ession LR is a multi v ariate analytic discrete choice model. This approach is most commonly used for stati sti- cal analysis of biostatistics, sociology , quantitati v e psychology , clinical practice, econometrics, and mark eting as well as to compare it with machine learning studies. It of fers numerous benets such as accurac y and strong po wer [41]. 3.3. Support v ector machine SVM is a v ery commonly used algorithm of machine learning. It is capable of both linear and non- linear classication [42]. W e set the k ernel parameter to Gaussian radial basis k ernel function (RBF). By using a grid search, the parameters C and g amma ha v e been determined. 3.4. Random f or est RF are an ensemble method of machine learning. RF generates a lar ge number of DT randomly from the training set and can infer v alues for v arious DT to forecast their o v erall se v erity decit [41]. The parameters used in the dataset are criterion (entrop y), class weight (balanced), and n estimators (n=200). 3.5. Decision tr ee This classier has a tree structure, with internal nodes that represent features, branches that reect decision mechanis ms, and lea v es re presenting the outcome. It is easy to understand since it mirrors the human thought process. This classi er’ s internal decision-making process is recognized for its tree structure. This can easily process multidimensional data and requires v ery little training time [43]. 3.6. K-near est neighbor The KNN classication method is an easy and clean w ay of classicat ion [44]. The parameters con- sidered for KNN are n neighbors(n=4) with a step size of 1. W e set the parameter weights to distance. 3.7. AdaBoosting AdaBoosting classier is an ensemble strate gy for machine learning that combines weak learner mod- els and produces strong learners. When a machine learning system recei v es weights from training data, it acts as a base learner . W e used sklearn.ensemble import AdaBoostClassier to implement AdaBoost classier . At rst, a randomly traini ng set will be used, and the model will be trained continuall y . In ne xt iteration, mis- classication observ ations are considered to ha v e more weight and a greater lik elihood. This technique will be continued as long as the data in the training database are not entirely aligned with the model [45]. Int J Artif Intell, V ol. 13, No. 4, December 2024: 4380–4387 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4383 3.8. Multi-lay er per ceptr on MLP is a po werful articial neural netw ork that mimics the operation of the human brain and map its collection of inputs and outputs accordingly . F or the e xperiment, we import MLP classier via sklearn.neural netw ork. It’ s parameters include acti v at ion function, distinct classes of decit se v erity , input/output layers, learning rate, and iterations [45]. 3.9. Stochastic gradient descent SGD is a machine learning algorithm that performs the discriminati v e analysis of dif ferential clas - sication with loss functions. The benet of adopting SGD is that it is simple to implement and increases ef cienc y . In order to assist in determining dif ferent loss functions and penalties, a SGD classier is used for the implementation of SGD. The SGD classier is created by importing SGD classier from sklearn.linear model [45]. 3.10. P erf ormance e v aluation This section e xplores a set of e v aluation metrics for assessing the quality of generati v e models. V ari ous metrics ha v e been calculated such as f alse positi v es (FP), true posi ti v es (TP), f alse ne g ati v es (FN), true ne g ati v es (TN), recall, precision, F1-score, accurac y , and area under the curv e (A UC) [46]. An additional instrument for assessing the ef fecti v eness of machine learning classication the confusion matrix. 4. RESUL TS AND DISCUSSION This se ction display the results of the proposed models. Machine learning algorithms ha v e been im- plemented using bigram and trigram TF-IDF based models. The RF and SVM achie v ed the highest results with trigram representation as sho wn in T able 1. All performance metrics ha v e been calculated for each classier . Precision v alues range from 0.73 to 0.84, F1-score v alues v ary from 0.78 to 0.87, recall v aries from 0.81 to 0.90 and the accurac y v alue is between 0.74 and 0.85. F or bigram TF-IDF based models, RF , KNN, and SVM reported accurac y of 0.85, 0.81, and 0.80 respecti v ely . As sho wn in T able 2, A UC scores for bigram models range from 0.72 to 0.84 for all algorithms. RF , KNN, and SVM outperformed all other models resulting A UC of 0.84, 0.80, and 0.79 respecti v ely . As sho wn in T able 1, dif ferent e v aluation metrics ha v e been calculated for all algorithms and tr i- gram models. The precision score v aries from 0.75 to 0.85, F1-score ranges from 0.80 to 0.87, recall v aries from 0.81 to 0.92 and accurac y ranges from 0.76 to 0.85. RF and SVM with trigram model reported accu- rac y of 0.85 and 0.84 respecti v ely . As sho wn in T able 2, A UC scores for trigram models range from 0.74 to 0.84 for all algorithms. RF and SVM outperformed all other models resulting in A UC of 0.84 and 0.83 respecti v ely . Figures 2 and 3 sho w the confusion matrices and A UC for all algorithms with trigram model. Figures 2(a) to 2(h) represent the confusion matrix for Adaboost, DT , KNN, LR, MLP , RF , SGD, and SVM respecti v ely . Figures 3(a) to 3(h) represent the A UC for Adaboost, DT , KNN, LR, MLP , RF , SGD, and SVM respecti v ely . T able 1. Results of the proposed system with eight classiers in percentage with bigram and trigram N-gram Algorit hms Accurac y Precision recall F1-Score T rigram Random for est 0.85 0.83 0.92 0.87 Bigram 0.85 0.84 0.90 0.87 T rigram SVM 0.84 0.85 0.87 0.86 Bigram 0.80 0.81 0.85 0.83 T rigram K-neighbors 0.81 0.84 0.81 0.82 Bigram 0.81 0.83 0.82 0.83 T rigram AdaBoos t 0.78 0.79 0.84 0.81 Bigram 0.79 0.79 0.84 0.81 T rigram S GD 0.81 0.81 0.85 0.83 Bigram 0.77 0.76 0.84 0.80 T rigram Decis ion tree 0.76 0.75 0.86 0.80 Bigram 0.75 0.75 0.81 0.78 T rigram Logistic re gression 0.78 0.76 0.87 0.81 Bigram 0.74 0.73 0.84 0.78 T rigram M LP 0.77 0.75 0.87 0.81 Bigram 0.74 0.74 0.84 0.79 Char acter N-gr am model for toxicity pr ediction (Eman Shehab) Evaluation Warning : The document was created with Spire.PDF for Python.
4384 ❒ ISSN: 2252-8938 T able 2. Results of A UC of the proposed system with eight classiers with bigram and trigram N-gram Algori thms A UC T rigram Random forest 0.8408 Bigram 0.84428 T rigram SVM 0.8353 Bigram 0.7947 T rigram K-neighbors 0.8046 Bigram 0.80265 T rigram AdaB oost 0.7763 Bigram 0.77876 T rigram SGD 0.7985 Bigram 0.756 T rigram Decis ion tree 0.7489 Bigram 0.728 T rigram Logistic re gression 0.76202 Bigram 0.722 T rigram MLP 0.7523 Bigram 0.73073 (a) (b) (c) (d) (e) (f) (g) (h) Figure 2. Confusion matrix of (a) Adaboost, (b) DT , (c) KNN, (d) LR, (e) MLP , (f) RF , (g) SGD, and (h) SVM Int J Artif Intell, V ol. 13, No. 4, December 2024: 4380–4387 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4385 (a) (b) (c) (d) (e) (f) (g) (h) Figure 3. A UC of (a) Adaboost, (b) DT , (c) KNN, (d) LR, (e) MLP , (f) RF , (g) SGD, and (h) SVM The results re v ealed that trigram produced higher performance than bigram. RF , SVM, and KNN classiers outperformed all other classiers. T o our kno wledge, the proposed model resulted in a considerable performance according to the resources that used to train the models. The reported results of the proposed models are close to the reported results in [20], [47]. 5. CONCLUSION In this research, machine learning techniques and N-gram model were inte grated and used to clas sify chemical respiratory toxicity . F or this purpose , dif ferent standard models were used for detailed analysis. W e g athered chemicals associated with respiratory toxicity from v arious databases and the lite rature. This study used SVM, RF , LR, AdaBoost, DT , MLP , SGD, and KNN as classication algorithms. The proposed model ha v e been e v aluated using accurac y , confusion matrix, F1-score precision, and sensiti vity . Results sho wed that the trigram model outperforms the bigram model. On the other hand, RF and SVM achie v ed the highest results in terms of accurac y with trigram representation. F or future w ork, we will study to impro v e the model interpretability and prediction performance using dif ferent AI techniques and molecular properties. In addition, dif ferent representation models can be e xperimented with other learning models. REFERENCES [1] D. D. Martini, “Empo wering phase II clinical t rials to reduce phase III f ailures, ” Pharmaceutical Statistics , v ol. 19, no. 3, pp. 178–186, May 2020, doi: 10.1002/pst.1980. [2] R. K. Harrison, “Phase II and phase III f ailures: 2013–2015, ” Natur e Re vie ws Drug Disco very , v ol. 15, no. 12, pp. 817–818, Dec. 2016, doi: 10.1038/nrd.2016.184. [3] G. Biala et al. , “Research in the eld of drug design and de v elopment, ” Pharmaceuticals , v ol. 16, no. 9, 2023, doi:10.3390/ph16091283. Char acter N-gr am model for toxicity pr ediction (Eman Shehab) Evaluation Warning : The document was created with Spire.PDF for Python.
4386 ❒ ISSN: 2252-8938 [4] F . M. I. Hunter , A. P . Bento, N. Bos c, A. Gaulton, A. Herse y , and A. R. Leach, “Drug safety data curation and modeling in ChEMBL: box ed w arnings and withdra wn drugs, ” Chemical Resear c h in T oxicolo gy , v ol. 34, no. 2, pp. 385–395, Feb . 2021, doi: 10.1021/acs.chemrestox.0c00296. [5] H. Zhang, J.-X. Ma, C.-T . Liu, J.-X. Ren, and L. Ding, “De v elopment and e v aluation of in silico prediction model for drug-induced respiratory toxicity by using na ¨ ıv e Bayes classier method, ” F ood and Chemical T oxicolo gy , v ol. 121, pp. 593–603, No v . 2018, doi: 10.1016/j.fct.2018.09.051. [6] V . V enkatraman, “FP-ADMET : a compendium of ngerprint-based ADM ET prediction models, ” J ournal of Cheminformatics , v ol. 13, no. 1, Dec. 2021, doi: 10.1186/s13321-021-00557-5. [7] H. Zhang, L. Zhang, C. Gao, R. Y u, and C. Kang, “Pharmacophore screening, molecular docking, ADMET prediction and MD simulations for identication of ALK and MEK potential dual inhibitors, ” J ournal of Molecular Structur e , v ol. 1245, Dec . 2021, doi: 10.1016/j.molstruc.2021.131066. [8] Y . Hua et al ., “Drug repositioning: Progress and challenge s in drug disco v ery for v arious diseases, ” Eur opean J ournal of Medicinal Chemistry , v ol. 234, Apr . 2022, doi: 10.1016/j.ejmech.2022.114239. [9] A. B. Deore, J. R. Dhumane, R. W agh, and R. Sona w ane, “The stages of drug disco v ery and de v elopment process, ” Asian J ournal of Pharmaceutical Resear c h and De velopment , v ol. 7, no. 6, pp. 62–67, Dec. 2019, doi: 10.22270/ajprd.v7i6.616. [10] B. Shak er , S. Ahmad, J. Lee, C. Jung, and D. Na, “In silico methods and tools for drug disco v ery , ” Computer s in Biolo gy and Medicine , v ol. 137, Oct. 2021, doi: 10.1016/j.compbiomed.2021.104851. [11] O. Sil akari and P . K. Singh, “ ADMET tools: Prediction and assessment of chemical ADMET properties of NCEs, ” in Concepts and Experimental Pr otocols of Modelling and Informatics in Drug Design , Else vier , 2021, pp. 299–320, doi: 10.1016/B978-0-12- 820546-4.00014-3. [12] W . Zhang, H. Zou, L. Luo, Q. Liu, W . W u, and W . Xiao, “Predicting potential side ef fects of drugs by recommender methods and ensemble learning, ” Neur ocomputing , v ol. 173, pp. 979–987, Jan. 2016, doi: 10.1016/j.neucom.2015.08.054. [13] W . Zhang, X. Y ue, F . Liu, Y . Chen, S. T u, and X. Zhang, “ A unied frame of predicting side ef fects of drugs by using linear neighborhood similarity , ” BMC Systems Biolo gy , v ol. 11, no. S6, Dec. 2017, doi: 10.1186/s12918-017-0477-2. [14] W . Zhang, X. Liu, Y . Chen, W . W u, W . W ang, and X. Li, “Feature-deri v ed graph re gularized matrix f actorization for predicting drug side ef fects, ” Neur ocomputing , v ol. 287, pp. 154–162, Apr . 2018, doi: 10.1016/j.neucom.2018.01.085. [15] P . C.-Reboredo et al ., “ A re vie w on mac hine learning approaches and trends in drug disco v ery , ” Computational and Structur al Biotec hnolo gy J ournal , v ol. 19, pp. 4538–4558, 2021, doi: 10.1016/j.csbj.2021.08.011. [16] Y . Liu et al ., “Experimental study and random forest prediction model of microbiome cell surf ace h ydrophobicity , ” Expert Systems with Applications , v ol. 72, pp. 306–316, Apr . 2017, doi: 10.1016/j.esw a.2016.10.058. [17] P . R. -Fernandez, C. R. Munteanu, J. Dorado, R. M.-Romalde, A. D.-Sanchez, and H. G.-Diaz, “From chemi cal graphs in computer - aided drug design to general mark o v-g alv ez indices of drug-tar get, proteome, drug-parasitic disease, technological, and social-le g al netw orks, ” Curr ent Computer Aided-Drug Design , v ol. 7, no. 4, pp. 315–337, Dec. 2011, doi: 10.2174/157340911798260340. [18] P . Shirv ani and A. F assihi, “Molecular modelling study on p yrrolo[2,3- b ]p yridine deri v ati v es as c-Met kinase inhibitors: a com- bined approach using molecular docking, 3D-QSAR modelling and molecular dynamics simulation, ” Molecular Simulation , v ol. 46, no. 16, pp. 1265–1280, No v . 2020, doi: 10.1080/08927022.2020.1810853. [19] B. S. -Garcia, J. I. B. -Bordils, A. F alc ´ o, M. T . P . -Gracia, G. A. -F os, and P . A.-L ´ opez, “Quantitati v e structure–acti vity relationship methods in the disco v ery and de v elopment of antibacterials, ” WIREs Computational Molecular Science , v ol. 10, no. 6, No v . 2020, doi: 10.1002/wcms.1472. [20] Z. W ang et al ., “In silico prediction of chemical respirat ory toxicity via machine learning, ” Computational T oxicolo gy , v ol. 18, May 2021, doi: 10.1016/j.comtox.2021.100155. [21] Y . Peng, Z. Zhang, Q. Jiang, J. Guan, and S. Zhou, “T OP: A deep mixture representation learning method for boosting molecular toxicity prediction, ” Methods , v ol. 179, pp. 55–64, Jul. 2020, doi: 10.1016/j.ymeth.2020.05.013. [22] J. Chen, Y .-W . Si, C.-W . Un, and S. W . I. Siu, “Chemical toxicity prediction based on semi-supervised learning and graph con v olu- tional neural netw ork, ” J ournal of Cheminformatics , v ol. 13, no. 1, Dec. 2021, doi: 10.1186/s13321-021-00570-8. [23] K. Jag anathan, H. T ayara, and K. T . Chong, “Prediction of drug-induced li v er toxicity using SVM and optimal descriptor sets, ” International J ournal of Molecular Sciences , v ol. 22, no. 15, Jul. 2021, doi: 10.3390/ijms22158073. [24] X. Huang, F . T ang, Y . Hua, and X. Li, “In silico prediction of drug-induced ototoxicity using machine learning and deep learning methods, ” Chemical Biolo gy & Drug Design , v ol. 98, no. 2, pp. 248–257, Aug. 2021, doi: 10.1111/cbdd.13894. [25] H. Feng et al ., “Predicting the reproducti v e toxicity of chemicals using ensemble learning methods and molecular ngerprints, ” T oxicolo gy Letter s , v ol. 340, pp. 4–14, Apr . 2021, doi: 10.1016/j.toxlet.2021.01.002. [26] J. Zhang, U. Norinder , and F . Sv ensson, “Deep learning-based conformal prediction of T oxicity , ” J ournal of Chemical Information and Modeling , v ol. 61, no. 6, pp. 2648–2657, Jun. 2021, doi: 10.1021/acs.jcim.1c00208. [27] Y . Hua, Y . Shi, X. Cui, and X. Li, “In silico prediction of chemical-induced hematotoxicity with machine learning and deep learning methods, ” Molecular Diver sity , v ol. 25, no. 3, pp. 1585–1596, Aug. 2021, doi: 10.1007/s11030-021-10255-x. [28] P . Camus, “The drug-induced respiratory disease website, ” Pneumotox , 2024. [Online]. A v ailabl e: https://www .pneumotox.com/drug/inde x/ [29] M.-C. Cai et al ., “ ADReCS: an ontology database for aiding standardization and hierarchical classication of adv erse drug reaction terms, ” Nucleic Acids Resear c h , v ol. 43, no. D1, pp. D907–D913, Jan. 2015, doi: 10.1093/nar/gku1066. [30] S. Dik, J. Ezendam, A. R. Cunningham , C . A. Carrasquer , H. V . Lo v eren, and E. Rorije, “Ev aluation of in silico models for the iden- tication of respiratory sensitizers, ” T oxicolo gical Sciences , v ol. 142, no. 2, pp. 385–394, Dec. 2014, doi: 10.1093/toxsci/kfu188. [31] J. Jarvis, M. J. Seed, S. J. Stocks, and R. M. Agius, “ A rened QSAR model for predi ction of chemical asthma hazard, ” Occupational Medicine , v ol. 65, no. 8, pp. 659–666, No v . 2015, doi: 10.1093/occmed/kqv105. [32] G. R. V erhe yen, E. Braek en, K. V . Deun, and S. V . Miert, “Ev aluation of in silico tools to predict the skin sensitization potential of chemicals, ” SAR and QSAR in En vir onmental Resear c h , v ol. 28, no. 1, pp. 59–73, Jan. 2017, doi: 10.1080/1062936X.2017.1278617. [33] S. J. Enoch, D. W . Roberts, and M. T . D. C ronin, “Mechanistic cate gory formation for the prediction of respiratory sensitization, ” Chemical Resear c h in T oxicolo gy , v ol. 23, no. 10, pp. 1547–1555, Oct. 2010, doi: 10.1021/tx100218h. [34] M. A. W arne, J. K. Nicholson, J. C. Lindon, P . D. Guine y , and K. P . R. Gartland, “ A QSAR in v estig ation of dermal and respiratory Int J Artif Intell, V ol. 13, No. 4, December 2024: 4380–4387 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4387 chemical sensitizers based on computational chemistry properties, ” SAR and QSAR in En vir onmental Resear c h , v ol. 20, no. 5–6, pp. 429–451, Jul. 2009, doi: 10.1080/10629360903278768. [35] R. C. Bra g a et al ., “Pred-Ski n: A f ast and reliable web application to assess skin sensitization ef fect of chemicals, ” J ournal of Chemical Information and Modeling , v ol. 57, no. 5, pp. 1013–1017, May 2017, doi: 10.1021/acs.jcim.7b00194. [36] P . T omasulo, “ChemIDplus-super source for chemical and drug information, ” Medical Refer ence Services Quarterly , v ol. 21, no. 1, pp. 53–59, Jan. 2002, doi: 10.1300/J115v21n01 04. [37] F . Grisoni, “Chemi cal language models for de no v o drug design: Challenges and opportunities, ” Curr ent Opinion in Structur al Biolo gy , v ol. 79, Apr . 2023, doi: 10.1016/j.sbi.2023.102527. [38] J. A. -Pous et al ., “Randomized SMILES strings impro v e the quality of molecular generati v e models, ” J ournal of Cheminformatics , v ol. 11, no. 1, Dec. 2019, doi: 10.1186/s13321-019-0393-0. [39] “Supervise d learning, ” Scikit Learn , 2023. [Online]. A v ailable: https://scikit-learn.or g/stable/supervised learning.html [40] M. T aha, H. H. Zayed, M. Azer , and M. Gadallah, “ Automated CO VID-19 misinformation checking system using encoder repre- sentation with deep learning models, ” IAES International J ournal of Articial Intellig ence (IJ-AI) , v ol. 12, no. 1, pp. 488-495, Mar . 2023, doi: 10.11591/ijai.v12.i1.pp488-495. [41] C.-C. W u et al ., “Prediction of f at ty li v er disease using machine learning algorithms, ” Computer Methods and Pr o gr ams in Biomedicine , v ol. 170, pp. 23–29, Mar . 2019, doi: 10.1016/j.cmpb .2018.12.032. [42] J. Cerv antes, F . G.-Lamont, L. R.-Mazahua, and A. Lopez, “ A comprehensi v e surv e y on support v ector machine classication: Applications, challenges and trends, ” Neur ocomputing , v ol. 408, pp. 189–215, Sep. 2020, doi: 10.1016/j.neucom.2019.10.118. [43] B. Charb uty and A. Abdulazeez, “Classication based on decision tree algorithm for machine learning, ” J ournal of Applied Science and T ec hnolo gy T r ends , v ol. 2, no. 1, pp. 20–28, Mar . 2021, doi: 10.38094/jastt20165. [44] M. Bansal, A. Go yal, and A. Choudhary , “ A com parati v e analysis of k-nearest neighbor , genetic, support v ector machine, de- cision tree, and long short term memory algorithms in machine learning, ” Decision Analytics J ournal , v ol. 3, Jun. 2022, doi: 10.1016/j.dajour .2022.100071. [45] G. Sambasi v am, J. Amudha v el, and G. Sath ya, “ A predicti v e performance analysis of vitamin D decienc y se v erity using machine learning methods, ” IEEE Access , v ol. 8, pp. 109492–109507, 2020, doi: 10.1109/A CCESS.2020.3002191. [46] M. Mor g an, C. Blank, and R. Seetan, “Plant disease prediction using clas sication algorithms, ” IAES International J ournal of Articial Intellig ence (IJ-AI) , v ol. 10, no. 1, pp. 257-264, Mar . 2021, doi: 10.11591/ijai.v10.i1.pp257-264. [47] K. Jag anathan, H. T ayara, and K. T . Chong, “ An e xplainable supervised machine learning model for predicting respiratory toxicity of chemicals using optimal molecular descriptors, ” Pharmaceutics , v ol. 14, no. 4, Apr . 2022, doi: 10.3390/pharmaceutics14040832. BIOGRAPHIES OF A UTHORS Eman Shehab is currently a Lecturer Assistant in the Department of Computer Science, F aculty of Computers and Artici al Intelligence, Uni v ersity of Sadat City , Egypt, since 2020. She hols a bachelor of computer science. She achie v ed her master’ s de gree from Menoua Uni v ersity . She has w ork ed on se v eral research topics, her research interests are neural netw ork and machine learning. She can be contacted at email: eman.shehab@fcai.usc.edu.e g. Hamada Nay el is an Assistant Professor at the Department of C omputer Science, F aculty of Computers and Articial Intelligence, Benha Uni v ersity , Benha, Egypt. In 2019, he recei v ed his Ph.D. from Mang alore Uni v ersity , India. His research interests include Arabic NLP , biomedical NLP , and social media analysis. He can be contacted at email: hamada.ali@fci.b u.edu.e g. Mohamed T aha is an Associat e Professor at Department of Computer Science, F aculty of Computers and Articial intelligenc e, Benha Uni v ersity , Egypt. He recei v ed his M .Sc. de gree and his Ph.D. de gree in computer science at Ain Shams Uni v ersity , Egypt, i n February 2009 and July 2015. He is the founder and coordinator of ”Netw orking and Mobile T echnologies” program, F aculty of Computers and Articial Intelligence, Benha Uni v ersity . His research interest’ s concern: computer vision (object tracking and video surv eillance sys tems), digital forensics (image for gery detection, document for gery detection, a nd f ak e currenc y detection), image processing (OCR), computer net- w ork (routing protocols and security), augmented reality , cloud computing, and data mining (associ- ation rules mining and kno wledge disco v ery). He has contrib uted more than 30+ technical papers in international journals and conferences. He can be contacted at email: mohamed.taha@fci.b u.edu.e g. Char acter N-gr am model for toxicity pr ediction (Eman Shehab) Evaluation Warning : The document was created with Spire.PDF for Python.