Predicting Drug Sensitivity of Cancer Cell Lines to BET Inhibitor OTX015 Using Machine Learning Approaches

1. Introduction

Cancer remains a major global public health concern, ranking as the second leading cause of death worldwide. Projections indicate a dramatic increase in cancer cases by 2050, attributed to population growth, aging, and lifestyle factors [1]. However, advancements in early detection, treatment, and supportive care have significantly improved survival rates, with approximately 70% of cancer patients surviving at least five years after diagnosis, and 11% surviving for 25 years or more [2]. Despite progress in traditional treatments such as radiotherapy, surgery, and chemotherapy, advanced biotechnological approaches, including gene therapy, stem cell therapy, epigenetic-based therapies, and precision medicine, are increasingly recognized for their potential in cancer diagnosis and treatment [3].

The Bromodomain and Extra-Terminal (BET) protein family, particularly BRD2, BRD3, BRD4, and BRDT, plays a central role in transcriptional regulation and chromatin remodeling by recognizing acetylated histones. Overactivation of BET proteins is implicated in various cancers, including hematologic malignancies, neuroblastomas, breast cancer, and prostate cancer [4,5]. BET inhibitors (BETi) like OTX015, JQ1, and I-BET762 have demonstrated significant efficacy in preclinical and clinical studies by downregulating oncogenic transcription factors such as MYC, BCL-2, and E2F [6,7].

Machine learning (ML) has emerged as a powerful tool to accelerate drug discovery and precision medicine. Supervised learning, in particular, offers a promising approach for predicting drug responses by recognizing patterns within complex biological data [8,9]. The use of programming languages such as R, alongside comprehensive repositories like Bioconductor, facilitates the development of robust models to predict drug efficacy based on genomic data [10].

In this study, we aimed to develop predictive models for estimating the half-maximal inhibitory concentration (IC₅₀) of OTX015 across various cancer cell lines, leveraging gene expression profiles and regression-based ML techniques. Among the models tested, the Support Vector Machine (SVM) consistently achieved the lowest Mean Absolute Error (MAE), outperforming other models across multiple datasets. These findings suggest that ML models, particularly SVM, hold great potential for enhancing drug discovery and personalized medicine. Future research should focus on refining these models, expanding datasets, and integrating diverse machine learning approaches to improve prediction accuracy.

2. Results

This study employed five regression-based machine learning models to predict the half-maximal inhibitory concentration (IC₅₀) values of the compound OTX015 across various cancer cell lines. The selected models were Elastic Net (EN), K-Nearest Neighbor (KNN), Neural Networks (NNET), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGB). The prediction performances of these models were assessed based on their Mean Absolute Error (MAE) scores, which serve as a key metric for evaluating prediction accuracy.

2.1. The Impact of Data Dimensions on Model Performance

To investigate the impact of data dimensionality on model performance, the main dataset was divided into five subsets. Each subset contained a different number of genes, specifically 10, 25, 50, 75, and 100, selected based on their correlation with IC₅₀ values. This approach allowed for the evaluation of whether increasing or decreasing the number of genes influences the performance of prediction models. Table 1 presents the MAE scores achieved by each model across different data dimensions.

The performance of the OTX015 compound across different data dimensions is shown in Table 1.

The SVM model consistently demonstrated the best performance across all datasets, achieving the lowest MAE scores. This suggests that the SVM model is particularly suitable for predicting IC₅₀ values, even when the number of genes varies.

2.2. Evaluation of Error Dimensions of Models

Although MAE scores provide a general assessment of model performance, it is important to examine the distribution of errors across different cancer cell lines to ensure robustness. Individual errors for each model across all data groups are depicted in Figure 1.

Additionally, examining the frequency of errors allows for the identification of models that produce low-degree errors more frequently than high-degree errors. As illustrated in Figure 2, the SVM model showed the highest concentration of errors close to zero, indicating consistent predictive accuracy. This analysis confirms that the SVM model is particularly effective in minimizing prediction errors across various cell lines.

2.3. Comparison of Predicted and Actual IC₅₀ Values

To further evaluate model performance, predicted IC₅₀ values were compared with actual values. The comparison graphs in Figure 3 illustrate the ability of each model to generalize predictions from training data. The results show that the SVM model is capable of producing predictions that closely align with actual values, demonstrating its generalization capability.

2.4. Performance of the SVM Model in Single Cell Line Predictions

Given the superior performance of the SVM model, additional evaluations were conducted to assess its predictive accuracy for individual cell lines. Each cell line in the validation set of the Data of 10 was considered as a single sample, and predictions were generated accordingly. The results for 76 cell lines are presented in Table 2.

The SVM model exhibited a low average error rate (MAE = 1.143), highlighting its robust performance even when applied to single cell lines. This finding demonstrates the suitability of the SVM model for predicting drug sensitivity on an individual cell line basis, enhancing its potential application in personalized medicine.

3. Discussion

The findings of this study highlight the efficacy of machine learning models, particularly the Support Vector Machine (SVM), in predicting the half-maximal inhibitory concentration (IC₅₀) of the compound OTX015 in various cancer cell lines. The use of gene expression profiles for prediction aligns well with current trends in precision medicine, where genomic data is utilized to tailor therapeutic approaches for individual patients [11,12].

The performance of the models across different data dimensions demonstrated that increasing the number of genes included in the dataset generally improved prediction accuracy, as reflected by lower Mean Absolute Error (MAE) scores. However, the improvement was not strictly linear, indicating that certain subsets of genes may provide more relevant information for predicting IC₅₀ values than others. This observation is consistent with the idea that the quality of features is more important than their quantity, a key concept in feature selection and model optimization [13,14].

The comparison of individual errors and error distributions across models showed that the SVM model consistently outperformed other models, achieving the lowest average errors and exhibiting a high concentration of errors near zero. This suggests that the SVM model is robust in minimizing prediction errors across various cell lines. Additionally, the strong performance of the Elastic Net model, which incorporates both L1 and L2 regularization, indicates that hybrid methods may also hold promise for enhancing prediction accuracy [15,16].

A critical observation in this study is the performance of the SVM model when applied to single cell line predictions. Despite training on a general dataset, the SVM model demonstrated strong predictive capabilities for individual cell lines, further supporting its suitability for personalized medicine applications. This finding is particularly significant, as personalized predictions can guide experimental studies towards prioritizing high-potential compounds and optimizing therapeutic strategies.

The results of this study are in agreement with previous research efforts aimed at developing predictive models for drug sensitivity using genomic data. For instance, Chiu et al. (2019) successfully applied neural networks and random forests to predict drug efficacy based on mutation and expression profiles of cancer cells [17]. Similarly, Su et al. (2019) demonstrated the potential of deep forest models to predict drug sensitivity from gene expression data [18]. Furthermore, Ding et al. (2019) highlighted the value of integrating machine learning algorithms with omics data to improve prediction accuracy and therapeutic efficacy [19].

Despite the promising results, certain limitations must be acknowledged. First, this study focused exclusively on predicting IC₅₀ values for a single compound, OTX015. While this approach allowed for a detailed evaluation of prediction models, it limits the generalizability of the findings to other BET inhibitors or unrelated drug classes. Expanding the analysis to include additional compounds would provide a more comprehensive evaluation of model performance. Second, the training datasets used in this study were derived from existing literature, which may lack diversity in terms of cancer cell lines, experimental conditions, and data quality. The reliance on publicly available datasets could introduce biases that may not accurately reflect real-world scenarios. Addressing this limitation would require integrating more diverse datasets, including those generated from patient-derived samples and primary tumor cells. Third, while the models demonstrated strong predictive capabilities on existing data, their performance on real-world patient-derived samples remains unexplored. Testing the models against clinical datasets or organoid models derived from patient tissues would be essential to validate their practical applicability in precision medicine. Finally, the current study utilized gene expression data as the sole feature set for prediction. Integrating other omics data types, such as proteomics, metabolomics, and epigenomics, could enhance prediction robustness by providing a more comprehensive representation of the molecular mechanisms influencing drug sensitivity.

Overall, the results of this study emphasize that the SVM model is the most appropriate choice among the models evaluated, due to its superior generalization capability and ability to provide consistent predictions across diverse datasets. Future research should focus on expanding the dataset to include additional cell lines and drugs, optimizing feature selection processes, and integrating various machine learning approaches to enhance prediction robustness.Moreover, the integration of biological pathway information and transcriptomic profiles could provide additional insights into the molecular mechanisms underlying drug sensitivity, further enhancing the applicability of machine learning models in precision oncology.

The findings of this study contribute to the growing body of research advocating the use of machine learning for predicting potential drug candidates and optimizing cancer treatment strategies. Continued efforts to refine predictive models and incorporate additional datasets will be essential to advance personalized medicine and improve therapeutic outcomes for cancer patients.

4. Materials and Methods

In this study, five different regression-based machine learning models were developed using the R programming language and R Studio to predict the IC₅₀ values of the bromodomain inhibitor OTX015 in cancer cell lines:

• IC₅₀ Values: The IC₅₀ values of OTX015 in cancer cell lines were obtained from the CancerRxGene (https://www.cancerrxgene.org/) data source. This data provides a crucial measure of the impact of OTX015 on cancer cell lines.
• Gene Expression Data: The gene expression levels of cancer cells were obtained from the Sanger Institute's Cancer Cell Lines Project (https://cancer.sanger.ac.uk/cell_lines) data source. These data allow for a detailed examination of the gene expression profile in cancer cells.

4.1. Comparison of Predicted and Actual IC₅₀ Values

The collected IC₅₀ values and gene expression data underwent a series of preprocessing steps before being used for model training, testing, and validation. The preprocessing was conducted using the R programming language, employing various statistical and machine learning packages. The following steps were performed:

• Data Integration: IC₅₀ values and gene expression data were merged based on cancer cell lines. This step helped determine in which cancer cells IC₅₀ values were determined for the OTX015 compound and in which cells gene expression levels were measured.
• Data Processing: Missing data were checked, and any incomplete data were removed from the dataset. Following the literature (Table 1), cell lines with IC₅₀ values of 200 µM or higher were considered inactive and were excluded from the dataset [20].
• Feature Engineering: To determine the features used for model development, feature engineering was performed on the gene expression data. Only significant gene expression levels were selected. Each gene in the dataset was subjected to Pearson correlation with the IC₅₀ value. The relationships were ranked from highest to lowest based on the absolute value of the correlation, irrespective of whether the correlation was positive or negative. Genes with significant correlations to IC₅₀ values were selected in sets of 10, 25, 50, 75, and 100 to evaluate the impact of data dimensions on the models, resulting in five different new datasets.
• Data Scaling: Rescaling the data often improves model performance by reducing error rates, depending on the nature of the problem, data, and models. Therefore, rescaling was applied to all five models used in this study. This contributed to better model performance and lower error rates.
• Creation of Training, Testing, and Validation Data: To evaluate the model's performance on a dataset it had not seen before, the dataset was randomly split before model training. 80% of the data (314 samples) was used for training and testing, and 20% (76 samples) was used for validation. The validation data set was reserved for assessing the model’s performance and was not used during the training or optimization stages. Cross-validation was performed using the training data. In cross-validation, the value of k was set to 5, creating five different training and testing sets during model training.

4.2. Model Training, Utilization, and Evaluation

In this study, five different regression-based machine learning models were implemented using the R programming language and relevant libraries to predict the IC₅₀ values of the bromodomain inhibitor OTX015 in cancer cell lines based on gene expression profiles. Each model was developed using its respective R package, leveraging their built-in functions for model training and evaluation. The selected models, along with their corresponding R libraries, are as follows:

• K-Nearest Neighbor (KNN) – implemented using the kknn package
• Extreme Gradient Boosting (XGB) – implemented using the xgboost package
• Elastic Net (EN) – implemented using the glmnet package
• Neural Networks (NNET) – implemented using the nnet package
• Support Vector Machine (SVM) – implemented using the e1071 package

These five regression models were trained separately using the same dataset and their performances were compared using the same validation set. Model training and evaluation were conducted in R Studio, utilizing the caret package for workflow standardization, hyperparameter tuning, and cross-validation.

During model training, hyperparameter optimization was performed for key parameters influencing model performance, ensuring that each model was fine-tuned for optimal accuracy. The optimal hyperparameters were determined based on the Mean Absolute Error (MAE) calculated using Equation (1)

$$MAE={\displaystyle \frac{1}{n}}{{\displaystyle \sum }}_{i=1}^n \left|x_i-x\right|$$

(1)

MAE was used to compute the mean of the absolute differences between the predicted and actual values. The parameters with the lowest MAE scores were selected as the optimal hyperparameters to maximize model performance. The models were then retrained with these optimal parameters and the same training set, and these retrained models were considered the best-performing models.

After optimization, the performance of the best models was evaluated using the validation set, which comprised 20% of the initial data that was not used during training. The MAE score was used for this evaluation. The MAE scores facilitated the comparison of model performances across different data sizes as well as among the five different machine learning models. The MAE scores obtained by the models on the validation sets are presented in Table 2. The entire workflow is illustrated in Figure 4.

5. Conclusions

This study demonstrated that the Support Vector Machine (SVM) model provides the most accurate predictions of IC₅₀ values for the BET inhibitor OTX015 in cancer cell lines compared to other models. Its strong performance across varying gene expression datasets highlights its potential for drug discovery and personalized oncology.However, the study's limitations include its focus on a single drug, reliance on literature-derived datasets, and lack of validation with real-world patient-derived samples. Future research should include diverse datasets, integrate additional omics data, and validate predictions using clinical and organoid-based systems. The SVM model's reliable performance makes it a promising tool for drug response prediction, contributing to the advancement of precision oncology and improving therapeutic strategies.