https://orcid.org/0009-0000-3097-0428
Figures
Abstract
Cancer treatment has become one of the biggest challenges in the world today. Different treatments are used against cancer; drug-based treatments have shown better results. On the other hand, designing new drugs for cancer is costly and time-consuming. Some computational methods, such as machine learning and deep learning, have been suggested to solve these challenges using drug repurposing. Despite the promise of classical machine-learning methods in repurposing cancer drugs and predicting responses, deep-learning methods performed better. This study aims to develop a deep-learning model that predicts cancer drug response based on multi-omics data, drug descriptors, and drug fingerprints and facilitates the repurposing of drugs based on those responses. To reduce multi-omics data’s dimensionality, we use autoencoders. As a multi-task learning model, autoencoders are connected to MLPs. We extensively tested our model using three primary datasets: GDSC, CTRP, and CCLE to determine its efficacy. In multiple experiments, our model consistently outperforms existing state-of-the-art methods. Compared to state-of-the-art models, our model achieves an impressive AUPRC of 0.99. Furthermore, in a cross-dataset evaluation, where the model is trained on GDSC and tested on CCLE, it surpasses the performance of three previous works, achieving an AUPRC of 0.72. In conclusion, we presented a deep learning model that outperforms the current state-of-the-art regarding generalization. Using this model, we could assess drug responses and explore drug repurposing, leading to the discovery of novel cancer drugs. Our study highlights the potential for advanced deep learning to advance cancer therapeutic precision.
Citation: Mohammadzadeh-Vardin T, Ghareyazi A, Gharizadeh A, Abbasi K, Rabiee HR (2024) DeepDRA: Drug repurposing using multi-omics data integration with autoencoders. PLoS ONE 19(7): e0307649. https://doi.org/10.1371/journal.pone.0307649
Editor: Amgad Muneer, The University of Texas, MD Anderson Cancer Center, UNITED STATES OF AMERICA
Received: May 13, 2024; Accepted: July 9, 2024; Published: July 26, 2024
Copyright: © 2024 Mohammadzadeh-Vardin et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The code and sample data for DeepDRA can be found on the GitHub repository: https://git.dml.ir/taha.mohammadzadeh/DeepDRA.
Funding: KA was supported by the Iran National Science Foundation (INSF) under Grant No. 99025741, and HRR was also partially supported by the Iran National Science Foundation (INSF) under Grant No. 96006077.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Cancer remains one of the most challenging diseases in modern medicine, with treatments ranging from surgery to chemotherapy and targeted drug therapy. A pharmacological intervention that targets cancer presents fewer side effects than more invasive therapies, but it is also costly, time-consuming, and complex to develop [1–3]. As a result, drug repurposing has emerged as a compelling alternative to traditional drug development [4].
Many clinical methodologies contributed to the advent of drug repurposing. Computational techniques have rapidly eclipsed these traditional approaches, offering new insight by decoding complex relationships between drug data and cancer multi-omics, such as transcriptomics, proteomics, genomics, and epigenomics [5–9]. Early experiments with classical machine learning in drug response prediction have yielded promising results [10–20]. With the ascent of deep learning, its use in cancer drug response prediction has surged, paralleling trends in computational biology [21]. It has been shown that deep learning models can be more accurate than classical models when it comes to predicting drug responses [22–25]. Deep learning’s ability to discern intricate patterns in large datasets makes it one of the most powerful tools for drug response prediction [16,26–29]. Furthermore, deep learning has made it possible to investigate both individual drug response parameters within cancer contexts [29,30] as well as the effect of drug combinations within cancer contexts [4–6]. Furthermore, drug rankings have emerged as one of the key areas of investigation in the cancer drug response analysis [31,32].
Using deep learning techniques to predict drug responses has become increasingly important in recent years due to the capability of handling multi-dimensional feature space and identifying hidden connections between these features. Menden et al. [18] pioneered the use of a feed-forward neural network with a single hidden layer to predict IC50 values. Sharifi-Noghabi et al. [33] proposed a multi-omics late integration approach with triplet loss and a multi-omics data integration approach for enhanced predictions. In recent years, the DeepDRK model has been introduced, and it relies on kernels to extract features from drug and omics data. The present approach, presented by Wang et al. [34], employed a classifier to predict drug response while leveraging the trained model to repurpose the drug. It has also been reported that recommender systems can also be used to predict drug response [35].
Several works have used Convolutional Neural Networks to predict how a patient will respond to a particular medication. In pioneering work in the field, twin neural networks (tCNNs) have been employed to extract features from drug and omics data, representing a novel approach that has been used in the field for many years [36]. There has also been a growing interest in graph-driven drug response prediction (GraphDRP) as a result of the advent of Graph Neural Networks (GNNs) [37]. Liu et al. [38], introduced DeepCDR, a machine learning (GNN) approach to drug response prediction that pioneered the field of predicting drug response. More recent methods have utilized transformers for the task [39,40]. One notable contribution to this field was DeepTTA, which used transformers to extract features and represented a novel approach to drug response prediction [39]. There is another transformer-based method called iBT-Net, which uses transformers to create embeddings of omics data, as well as a simple multilayer perceptron (MLP) to obtain embeddings of drug data to generate a classifier that predicts drug response and uses these embeddings in a classification model [40].
Although these models have made significant progress, they often fail to generalize to different cancer datasets. When trained on specific datasets, such models perform poorly on external datasets. As models struggle to generalize outside specific datasets, the conventional development of predictive algorithms within a single study has frequently led to inflated estimates of prediction performance [29]. Further, a limited variety of data types has constrained the representation of cancer and drugs, highlighting the urgent need for multi-omic integration to enhance representation quality.
As a result, we propose a simpler, autoencoder-based model for generating a feature set for cell lines and drug embeddings using diverse datasets. Based on these embeddings, a classifier can then predict drug responses. In addition to improving generalization, the simplicity of this model enables it to integrate diverse data sources and minimize data noise, thereby enhancing predictions. The following advantages are associated with this approach:
- Comparing autoencoders with more intricate deep-learning models, autoencoders have shown superior generalization abilities.
- By integrating data types, the model was able to produce better results than when using a single omic dataset.
- Using autoencoders for dimensionality reduction improved the model’s performance.
- The model correctly predicted Drug-cell line pairs even though they were not included in the dataset and were unrecognized by the model.
This paper has three sections: the first delineates the datasets employed, the second continues with the problem formulation and explains the proposed solution, the next presents and analyzes experimental results and discusses their correlation with clinical research in detail, and the final section thoroughly discusses these findings.
Results
We conducted seven experiments to evaluate our model’s performance and generalization capabilities. These experiments covered various scenarios and datasets to assess the model’s robustness thoroughly. In the first experiment, we conducted an ablation study by removing the autoencoder parts of the model to examine their contributions to the final results. The results showed that these autoencoders played a crucial role in the model’s final output.
In the second experiment, we tested the model’s outcome in single omic data with multi-omic integration in generalization, focusing on the data integration hypothesis. Following this ablation study, we started to compare our model with previous methods.
Our third experiment employed a mixed dataset methodology, combining GDSC and CTRP datasets. We used these data as training and testing data, conducted 5-fold cross-validation, and compared our results with DeepDRK. The encouraging results indicated that our model is viable.
In the fourth experiment, we trained the model on a combination of datasets from GDSC and CTRP and tested it on the CCLE dataset to see if it could apply its learned patterns to external datasets. We compared the results of this experiment to those of DeepDRK. We also benchmarked our model against other works in the field, such as DeepTTA [39] and iBT-Net [40], to assess its performance.
We conducted additional experiments to understand the model’s abilities better. In the fifth experiment, we used 5-fold cross-validation to measure the model’s internal consistency as we trained and tested it on the GDSC dataset. In the sixth and seventh experiments, we trained on the GDSC dataset and tested it on the CCLE and CTRP datasets to evaluate its adaptability and efficacy across different datasets.
Our experimental analysis confirmed that our model is resilient and accurate in predicting drug effects. It has the potential to make significant contributions to precision medicine and therapeutic research. The insights from our comparative studies further emphasize the model’s competitive stance in drug response modeling.
Ablation study
An experiment was conducted to determine the contribution of autoencoder components to the overall effectiveness of a model for drug-cell line interaction classification. Initially, the model was evaluated without autoencoders, using only direct embedding data. As measured by the Area Under the Precision-Recall Curve (AUPRC), the result was a modest 0.74, indicating the need for improvement.
Subsequent trials incorporated the Drug Autoencoder, which led to a slight improvement in performance, with the AUPRC increasing to 0.76. This improvement demonstrated the positive impact of the Drug Autoencoder on the model’s performance.
Then, the Cell Line Autoencoder was evaluated, and it produced a remarkable increase in performance, resulting in an AUPRC of 0.91. This finding highlighted the pivotal role of the Cell Line Autoencoder in refining a feature space from high-dimensional cell line embeddings, which is essential to the model’s architecture.
Finally, the model was tested using both the Cell Line and Drug Autoencoders, and the results showed superior performance, indicating the synergistic effect of using both autoencoders. The results are presented in Table 1.
Further, we sought to quantify the impact of data integration by comparing the model’s performance on single omics data versus multi-omics data. The comparative analysis revealed that models utilizing multi-omics data outperformed those based on single omics data in mixed-set testing scenarios, as shown in Table 2.
Method comparison
During the initial evaluation, our model was trained on a combined dataset that included expression (exp), mutation (mut), and copy number variation (CNV) data, as well as drug descriptors and drug fingerprints. This dataset helped establish a performance baseline and served as the platform for our model’s first testing phase. Our model showed outstanding performance in this primary assessment with 5-fold cross-validation, achieving an AUPRC of 0.99 (Table 3). We conducted a comparative analysis against the DeepDRK model using an identical dataset to provide context for these results. The DeepDRK model achieved an AUPRC of 0.97, indicating that our proposed model has better results robustness (Fig 1).
(a) DeepDRA showed improvements in AUC results compared to the DeepDRK model in the mixed-set experiment on the GDSC+CTRP dataset. (b) AUPRC results were reported using the same test set as in (a). (c) DeepDRA and previous models were compared using.
We conducted additional evaluations to determine the capacity of our model to generalize in different scenarios. We performed a cross-dataset analysis where we trained our model on a combined dataset of CTRP and GDSC and then tested it on the CCLE dataset. Our model demonstrated robust generalization capabilities, as evidenced by an AUPRC of 0.87 88 (Table 4). To compare the results, we also subjected the DeepDRK model to the same testing regimen, and it returned an AUPRC of 0.85, which was significantly lower than the performance of our model (Fig 1).
These comparative experiments validate the strength and robustness of our model and also highlight its superior ability to generalize across diverse datasets. This attribute is crucial for precision medicine applications.
The model’s performance was rigorously evaluated under various other experimental conditions in these investigations. Our third experiment involved training the model on the GDSC dataset and testing it on the same dataset using 5-fold cross-validation. The model’s proficiency was reflected in an AUPRC of 0.98 (Table 5), a metric that attests to its outstanding predictive accuracy within the GDSC. This performance signifies a notable enhancement over the results achieved in the prior studies.
The model trained on GDSC data was tested against the CCLE dataset to evaluate cross-dataset generalization. Our model results in an AUPRC of 0.73 (Table 6). This outcome affirms the model’s adeptness in extrapolating learned patterns to an external dataset, an advancement that surpasses the generalization capabilities showcased in the preceding research efforts.
The final experiments involved training the model on the GDSC dataset and testing the CTRP dataset. Within this setting, the model manifested results commensurate with established models, thereby underscoring its steadfast performance across distinct datasets. Such consistent efficacy highlights the model’s adaptability and reaffirms its potential as an invaluable tool for drug response prediction (Table 7).
The results of our comprehensive experiments show the model’s capacity to function effectively across diverse analytical scenarios. These accomplishments hold significant promise, potentially paving the way for substantial contributions to precision medicine and therapeutic strategy optimization.
To evaluate the model’s generalization, experiments were performed on patient data. Due to the different domains in cell-line data and patient data, it’s normal to observe weak results when testing on actual patient data. Training and testing the model on patient data is preferable to achieve better results. However, patient omic and response data are scarce, and deep learning models require more data to train properly. Due to these challenges, the model was trained on the CTRP+GDSC dataset and tested on TCGA patient data downloaded from the TCGA portal. The results were then compared to DeepDRK, as shown in Table 8. Although the results are low, our models performed better than DeepDRK on patient data.
Drug repurposing
After training our model with the combined CTRP and GDSC datasets, we designed another experiment on breast cancer for further analysis. Our study aimed to evaluate the effectiveness of the full spectrum of drugs within the dataset against this specific type of cancer. Intriguingly, the model identified several effective drugs not previously cataloged in the dataset as sensitive to or resistant to breast cancer.
During validation, the model was tested on curated subsets of cancer cell lines, focusing primarily on breast cancer. Following model predictions for all possible drug-cell line pairs, we ranked the drugs according to their predicted efficacy scores. Those drugs that achieved scores above 0.5 were considered to have a probable positive effect on the cancer type. The high-ranking drugs in our model are commonly used in cancer treatment, followed by specific drugs that we suggest as potential breast cancer therapies. As shown in the following table (Table 9), these drugs, and the relevant references that support their effectiveness are listed.
For example, our model indicated that NVP-TAE684 is sensitive to breast cancer, a finding confirmed by recent studies [41]. Similarly, Olaparib, which lacked binarized response data in our database, is now recognized as a breast cancer therapeutic [42]. Dabrafenib also received a high score in the model, which reflects recent research confirming its efficacy against breast cancer [43]. Bleomycin also received a high score for treatment potential despite being binarized as ’unknown’ in our training data, which is consistent with prior investigations [44]. In the following studies, COL-3 was suggested as a potential treatment, a suggestion that was substantiated by recent findings [45]. Our model listed and recommended the following drugs based on their respective scores: LRRK2-IN-1, ML162, gefitinib, JQ-1, tigecycline, and neratinib [46–51].
Our model also identified BRD-A71883111 and ML050 as potential therapeutic agents in addition to the aforementioned drugs. This presents an opportunity for groundbreaking research since these drugs have not yet been explored in the literature. Without empirical data on these compounds, rigorous clinical testing is required to evaluate their efficacy and safety profiles. The recommendation of these drugs adds a new dimension to the ongoing search for innovative breast cancer treatments. It highlights the importance of integrating computational models into the early stages of drug discovery.
Materials and methods
This section aims to discuss the formulation of the problem and the implementation and evaluation of the model.
Dataset
Our data acquisition concentrates on three primary datasets in the current study. The Genomics of Drug Sensitivity in Cancer (GDSC) [63] stands out as a comprehensive repository, encompassing comprehensive multi-omics data related to cancer research. This dataset is notably rich, containing critical information on gene mutations, gene expression, copy number variations (CNV), and methylation profile, each contributing to the complexity of oncological molecular characterization. Despite its lack of multi-omics data, the Cancer Therapeutics Response Portal (CTRP) [64,65] is an invaluable source of drug response data. In contrast, the Cancer Cell Line Encyclopedia (CCLE) [66,67] provides comprehensive multi-omics data, including gene expression, mutation profiles, and copy number variation (CNV) data. We complemented CTRP’s lack of multi-omics data by leveraging CCLE’s data to enhance drug response analysis.
Preprocessing
During the first step of our data preprocessing, we carefully examined all the missing values in our datasets. Whenever we came across NaN (Not a Number) values, we used a standard imputation strategy to replace them with zeroes. This method works well for our datasets, as it helps to keep any distortion of the underlying data distributions to a minimum. Subsequently, we proceeded to extract relevant features from two comprehensive multi-omics datasets. The critical task here was identifying and extracting intersecting features across these datasets. This helped us train and test our model on different datasets.
Cell-line data
The GDSC dataset contains a vast amount of information about cancer cell lines. It includes 17,738 unique gene expression features, 23,189 features related to mutations, 14,729 features related to methylation, and 21,879 features related to copy number variations. This information can be used to analyze genomic alterations, epigenetics, and structural genomic variances that may be important in the development and progression of cancer. The CCLE dataset comprises 19,221 gene expression features and 19,534 mutation features. The CNV data has 59,267 features, which provides a comprehensive source for understanding the genomic structural diversity of cancer cell lines. Since the CTRP dataset lacks multi-omics data, the CCLE cell-line data was adopted to integrate vital information with CTRP’s drug response data.
A meticulous curation process is undertaken to optimize the datasets for model training. This process involves aligning the three common modalities across the GDSC and CCLE datasets to refine the dataset and eliminate irrelevant or unnecessary data. The outcome is a set of intersecting features that includes 15,963 gene expression features, 17,671 mutation features, and 23,874 CNV features, culminating in a comprehensive feature set of 57,508 for the combined cell-line dataset. By interlinking data from the GDSC and CCLE, we established a foundation for exploring the complex interplay between genetic, epigenetic, and transcriptional elements. Such a systematic approach is essential for unraveling the nuances of cancer biology and drug sensitivity, propelling the frontiers of precision medicine and therapeutic innovation (Table 10).
Drug data
To thoroughly study drug responses, it is essential to have access to comprehensive pharmacological data. To this end, we have compiled drug response information from three different datasets: GDSC, CTRP, and CCLE. This compilation has resulted in creating a repository of drug data, laying a solid foundation for conducting in-depth pharmacological analysis. To ensure consistency and accuracy in our analysis, we obtained the SMILES representations for each compound from the PubChem database [68]. These representations provide a standardized chemical notation, which simplifies subsequent computational tasks.
Using the RDKit cheminformatics software, we extracted various molecular descriptors for each drug in our dataset (Landrum). These descriptors encapsulate a range of chemical properties, from structural attributes to physicochemical factors, that indicate the drug’s biological activity and pharmacodynamics. We also extracted Morgan’s Fingerprint data for these drugs and added it to our model. This extraction process resulted in a dataset consisting of 704 unique drug entities. Each entity is defined by a multi-dimensional feature set that reflects its distinct pharmacological profile. We took great care in curating and extracting comprehensive features from drug data, which are essential to this research. We significantly improved the reliability and accuracy of the drug prediction model.
Drug response
A crucial step in our preprocessing was the binarization of drug response data. The aim was to standardize various response metrics, such as Area Under the Curve (AUC) and half maximal Inhibitory Concentration (IC50), into binary categories that indicate ’resistant’ or ’sensitive’ responses. This binarization process facilitated a uniform analysis framework that can be used for different types of drug responses encountered in our research. In the CTRP dataset, we consider AUC values lower than 6 to be sensitive, while values higher than 16 are considered resistant. For the GDSC dataset, we define drug response values lower than 0.2 as sensitive and values higher than 0.99 as resistant. Finally, for the CCLE dataset, we used the ActArea measurement for binarization. We consider act area values higher than 2 sensitive and values lower than 0.5 resistant. In the final preprocessing stage, we compiled a comprehensive list of drugs from the primary drug response datasets. For each drug, we systematically retrieved its corresponding Simplified Molecular Input Line Entry System (SMILES) [69] representation.
To train our model, we integrated drug response information from two datasets, GDSC and CTRP. This created a large merged dataset with over 28,000 cell-line drug pairs. This dataset was divided into 10,007 pairs labeled as ’sensitive’ and 18,758 pairs labeled as ’resistant’ (Table 11).
Before training our classifier, we carefully processed and classified the drug response data to establish a strong foundation. By converting the continuous drug response measurements into binary values, we were able to develop a robust classifier that could generalize well across various cell lines. This approach aligns with standard practices in the field and offers insights into the mechanisms that govern drug sensitivity and resistance. These insights can be invaluable in advancing precision medicine and developing new therapeutic interventions.
Problem formulation
Our research utilizes diverse inputs, including multi-omics data from cell lines (such as genomics, transcriptomics, and epigenomics), drug descriptors, and fingerprints. The objective of our model is to predict a sensitivity or resistance score, ranging from 0 to 1, for a given drug-cell-line pair. Specifically, we have a dataset represented as a for i ranging from 1 to N, representing the cell-line data, corresponding to the drug descriptor, and denoting the response value associated with this pair. To make predictions, our approach integrates mutation, gene expression, copy number variation, methylation, drug descriptors, and fingerprints.
As part of our approach, we aimed to improve the model’s performance by ensuring that the features were distributed across the four types of omic data utilized in our study. This illustrates our proposed approach’s structure. During our initial phase, we focused on collecting and preprocessing omics datasets. We converted drug response measurements into binary classes and created a processed dataset that could be fed into our model. Next, we obtained multi-omics data for each cell line and used this information to train our cell-line autoencoder. The autoencoder was responsible for distilling the data into a feature space that captured the essential attributes of the multi-omics information.
We have created a system to predict drug responses in cell lines. To do this, we assembled the SMILES for the drugs we are studying and used RDKit to generate a set of molecular descriptors and fingerprints. These drug embeddings were put into our Drug Autoencoder, which created a feature space for each drug entity. We then combined the cell-line data with the drug data to create matched pairs according to drug response pairs. Finally, these pairs were used as input for the final classification model to predict drug responses (Fig 2). We provide further explanation of our approach in the following sections.
Initially, data is collected and preprocessed, cell-line modalities are combined, and the drug embeddings are derived using RDKit. As a final step, autoencoders are utilized to extract features, and then drug response is predicted using a multilayer perceptron (MLP). The figure illustrates these sequential steps.
Model implementation and evaluation
In this study, we aimed to improve the generalization capabilities of models within the limited dataset context by using autoencoder architectures. To train the autoencoder and classifier simultaneously, we designed a multi-task model, and its loss function was a composite of the autoencoder and classifier losses. This dual-faceted approach showed the model’s adeptness in handling mixed datasets. Our research objective was to achieve robust generalization, particularly under the constraints of limited data availability. By refining the integration strategies and identifying the most efficient autoencoder architectures, we significantly solved the generalization issue, leading to superior model performance, especially in cross-dataset evaluations. These results have important implications for precision medicine and therapeutic research, as they underscore the potential to enhance our understanding of drug response dynamics.
The autoencoder utilized in this study is designed with a solitary, hidden layer that has 256 units and uses the Rectified Linear Unit (ReLU) activation function. The size of the input data determines the input layer of the autoencoder. The autoencoder employs Mean Squared Error (MSE) as the loss function during training. The model includes the MLP and is trained with a composite loss function that sums the MSE and MLP losses. This multi-task training approach leverages the autoencoder’s feature extraction capabilities and the MLP’s classification strength, creating an effective learning method.
The MLP used in this study consisted of a single hidden layer with 128 units. ReLU and Sigmoid functions were used for the hidden and output layers. The size of the autoencoders’ latent layers determined the size of the input layer. The MLP was trained for 25 epochs using the AdamDelta optimizer with a learning rate of 0.01. The model was trained using a Core i3-10300 CPU and a 3060 Ti GPU. Care was taken to avoid overfitting during the training process by splitting the data into training, validation, and test sets in a ratio of 0.7, 0.1, and 0.2, respectively. Checking these three parts’ accuracy helped avoid overfitting during the training process.
For the loss function, we use MSE loss in autoencoder; the loss function in autoencoder is simple like this:
For the loss function, we use MSE loss in autoencoder; the loss function in autoencoder is simple like this:
(1)
And MLP has Binary Cross Entropy loss:
(2)
But in end-to-end integration, our loss is like this where Y is for autoencoder and y is the final class:
(3)
(4)
Hyperparameter optimization was performed through various experiments, and it was determined that an autoencoder with a hidden layer of 256 nodes produced better results than hidden layers with 1024, 512, and 128 nodes. A latent layer with 50 nodes was the best choice among 200, 100, 50, and 25 nodes. For the MLP, a hidden layer consisting of 128 nodes was optimal. Different learning rates were tested, with rates of 0.005, 0.01, and 0.05. It was determined that a learning rate of 0.01 produced the best results. The combination of autoencoder loss and classifier loss with different factors was also tested. Factors a, b, and c were used. The values 2, 1, and 0.5 were tested for each factor, but it was discovered that using a factor of 1 for all losses was the best.
The model code is available on GitHub for model implementation in practical systems and clinical use. First, the model should be trained on the input data. Clinical patient data is preferred over cell-line data for training and achieving better results, but both datasets can be used for training. The model can be trained on different hardware, and using a GPU is not necessary for the training step. After training, the saved model can be used with any patient omics data and all combinations of drugs to find the best drugs for that patient.
Discussion
We investigated by gathering cell line and response data from three leading cancer databases: GDSC, CTRP, and CCLE. We preprocessed at the initial stage to make the data fit for integration. Our approach involved utilizing four different types of omics data in combination with two categories of drug data, which helped improve prediction accuracy. We ensured that the omics features were consistent across all datasets and eliminated any cases of incomplete data. Moreover, we used the DeepDRK methodology to binarize drug responses, setting thresholds to enable analysis of various datasets and response metrics.
Although there is a current trend in developing complex architectures such as GNNs and Transformers for model development, we prioritized generalizability, especially in cross-dataset validations. For this reason, we designed a streamlined architecture consisting of two main Autoencoders—one for omic data and the other for drug data. We used Mean Squared Error (MSE) loss to reconstruct data in these Autoencoders. The resulting latent representations were then concatenated to create the input for a binary classification classifier. The classifier used binary cross-entropy as its loss function, and the overall loss was calculated by adding the losses of both autoencoders with the loss of the classifier.
Although complex architectures such as GNNs and Transformers are currently trending for model development, we focused on developing a streamlined architecture that prioritizes generalizability, especially in cross-dataset validations. Our architecture consists of two main Autoencoders—one for omics data and the other for drug data—using Mean Squared Error (MSE) loss to reconstruct data. We then concatenated the resulting latent representations to create the input for a binary classification classifier. The classifier used binary cross-entropy as its loss function, and the overall loss was calculated by adding the losses of both autoencoders with the loss of the classifier.
Our research on drug repurposing has led us to compare our latest model with other models in drug response prediction, such as iBT-Net. Our model has consistently outperformed these in various metrics. For example, when trained on 80% of the GDSC dataset and tested on the remaining 20%, our model achieved an AUPRC of 0.98, significantly higher than iBT-Net’s 0.90. Similarly, upon training with GDSC and testing on CCLE, our model achieved an AUPRC of 0.72, compared to iBT-Net’s 0.68, and an AUPRC of 0.73 against CTRP, where iBT-Net scored 0.67.
To validate the clinical relevance of our model, we selected cell lines specific to particular cancers. We tested all plausible drug-cell line combinations not included in our training dataset. Our model’s scoring of these combinations revealed a correlation with established cancer treatments, underscoring the model’s predictive validity. Interestingly, we identified certain drugs that were unique to specific cancers. These drugs have either been recently recognized for their efficacy or are yet to be explored, suggesting potential new avenues for cancer therapy.
In our work, we encountered some limitations that require attention. When we used binarized response data, we experienced some data loss, as the data class was undefined. Additionally, we could not conduct clinical experiments due to incomplete data modalities, which prevented us from testing our model on patient data with all the modalities used in the model.
Based on our findings, we need to focus more on data preparation and integration rather than relying solely on complex models, which may not generalize well. In the future, we plan to improve the Autoencoders’ denoising capability, include additional omics data such as PPRA and RNA-seq, and increase the model’s robustness by training it with datasets that exhibit incomplete modalities. Adding a KEGG pathway or complete survival analysis will be very helpful and should be done in future work.
References
- 1. Baptista D., Ferreira P.G., and Rocha M., Deep learning for drug response prediction in cancer. Briefings in bioinformatics, 2021. 22(1): p. 360–379. pmid:31950132
- 2. Debela D.T., et al., New approaches and procedures for cancer treatment: Current perspectives. SAGE open medicine, 2021. 9: p. 20503121211034366. pmid:34408877
- 3. Yang Y., et al., MDICC: novel method for multi-omics data integration and cancer subtype identification. Briefings in Bioinformatics, 2022. 23(3): p. bbac132. pmid:35437603
- 4. Mathur S. and Sutton J., Personalized medicine could transform healthcare. Biomedical reports, 2017. 7(1): p. 3–5. pmid:28685051
- 5. Chen J. and Zhang L., A survey and systematic assessment of computational methods for drug response prediction. Briefings in bioinformatics, 2021. 22(1): p. 232–246. pmid:31927568
- 6. Tanoli Z., Vähä-Koskela M., and Aittokallio T., Artificial intelligence, machine learning, and drug repurposing in cancer. Expert opinion on drug discovery, 2021. 16(9): p. 977–989. pmid:33543671
- 7. He Z., et al., Integrative analysis identified CD38 as a key node that correlates highly with immunophenotype, chemoradiotherapy resistance, and prognosis of head and neck cancer. Journal of Cancer, 2023. 14(1): p. 72. pmid:36605482
- 8. Jiang Z., et al., Recent advance in biological responsive nanomaterials for biosensing and molecular imaging application. International Journal of Molecular Sciences, 2022. 23(3): p. 1923. pmid:35163845
- 9. Wang C., et al., Andrographolide regulates H3 histone lactylation by interfering with p300 to alleviate aortic valve calcification. British Journal of Pharmacology, 2024. pmid:38378175
- 10. An X., et al., Representation of molecules for drug response prediction. Briefings in Bioinformatics, 2022. 23(1): p. bbab393. pmid:34571534
- 11. Azuaje F., Computational models for predicting drug responses in cancer research. Briefings in bioinformatics, 2017. 18(5): p. 820–829. pmid:27444372
- 12.
Ballester P.J., et al., Artificial intelligence for drug response prediction in disease models. 2022, Oxford University Press. p. bbab450.
- 13. Caponigro G. and Sellers W.R., Advances in the preclinical testing of cancer therapeutic hypotheses. Nature reviews Drug discovery, 2011. 10(3): p. 179–187. pmid:21358737
- 14. Firoozbakht F., Yousefi B., and Schwikowski B., An overview of machine learning methods for monotherapy drug response prediction. Briefings in Bioinformatics, 2022. 23(1): p. bbab408. pmid:34619752
- 15. Garnett M.J., et al., Systematic identification of genomic markers of drug sensitivity in cancer cells. Nature, 2012. 483(7391): p. 570–575. pmid:22460902
- 16. Güvenç Paltun B., Mamitsuka H., and Kaski S., Improving drug response prediction by integrating multiple data sources: matrix factorization, kernel and network-based approaches. Briefings in bioinformatics, 2021. 22(1): p. 346–359. pmid:31838491
- 17. Lamb J., et al., The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease. science, 2006. 313(5795): p. 1929–1935.
- 18. Menden M.P., et al., Machine learning prediction of cancer cell sensitivity to drugs based on genomic and chemical properties. PLoS one, 2013. 8(4): p. e61318. pmid:23646105
- 19. Rafique R., Islam S.R., and Kazi J.U., Machine learning in the prediction of cancer therapy. Computational and Structural Biotechnology Journal, 2021. 19: p. 4003–4017. pmid:34377366
- 20. Sharifi-Noghabi H., et al., Drug sensitivity prediction from cell line-based pharmacogenomics data: guidelines for developing machine learning models. Briefings in bioinformatics, 2021. 22(6): p. bbab294. pmid:34382071
- 21. Shen B., et al., A systematic assessment of deep learning methods for drug response prediction: from in vitro to clinical applications. Briefings in Bioinformatics, 2023. 24(1): p. bbac605. pmid:36575826
- 22. Hostallero D.E., Li Y., and Emad A., Looking at the BiG picture: incorporating bipartite graphs in drug response prediction. Bioinformatics, 2022. 38(14): p. 3609–3620. pmid:35674359
- 23. Jia P., et al., Deep generative neural network for accurate drug response imputation. Nature communications, 2021. 12(1): p. 1740. pmid:33741950
- 24. Sakellaropoulos T., et al., A deep learning framework for predicting response to therapy in cancer. Cell reports, 2019. 29(11): p. 3367–3373. e4. pmid:31825821
- 25. Zhu Y., et al., TGSA: protein–protein association-based twin graph neural networks for drug response prediction with similarity augmentation. Bioinformatics, 2022. 38(2): p. 461–468. pmid:34559177
- 26. Chiu Y.-C., et al., Deep learning of pharmacogenomics resources: moving towards precision oncology. Briefings in bioinformatics, 2020. 21(6): p. 2066–2083. pmid:31813953
- 27. Su R., et al., SRDFM: Siamese Response Deep Factorization Machine to improve anti-cancer drug recommendation. Briefings in bioinformatics, 2022. 23(2): p. bbab534. pmid:35043144
- 28. Wang H., et al., GADRP: graph convolutional networks and autoencoders for cancer drug response prediction. Briefings in Bioinformatics, 2023. 24(1): p. bbac501. pmid:36460622
- 29. Xia F., et al., A cross-study analysis of drug response prediction in cancer cell lines. Briefings in bioinformatics, 2022. 23(1): p. bbab356. pmid:34524425
- 30. Adam G., et al., Machine learning approaches to drug response prediction: challenges and recent progress. NPJ precision oncology, 2020. 4(1): p. 19. pmid:32566759
- 31. Daoud S., et al., Q-rank: reinforcement learning for recommending algorithms to predict drug sensitivity to cancer therapy. IEEE Journal of Biomedical and Health Informatics, 2020. 24(11): p. 3154–3161. pmid:32750950
- 32. Gerdes H., et al., Drug ranking using machine learning systematically predicts the efficacy of anti-cancer drugs. Nature communications, 2021. 12(1): p. 1850. pmid:33767176
- 33. Sharifi-Noghabi H., et al., MOLI: multi-omics late integration with deep neural networks for drug response prediction. Bioinformatics, 2019. 35(14): p. i501–i509. pmid:31510700
- 34. Wang Y., et al., DeepDRK: a deep learning framework for drug repurposing through kernel-based multi-omics integration. Briefings in Bioinformatics, 2021. 22(5): p. bbab048. pmid:33822890
- 35. Suphavilai C., Bertrand D., and Nagarajan N., Predicting cancer drug response using a recommender system. Bioinformatics, 2018. 34(22): p. 3907–3914. pmid:29868820
- 36. Liu P., et al., Improving prediction of phenotypic drug response on cancer cell lines using deep convolutional network. BMC bioinformatics, 2019. 20(1): p. 1–14.
- 37. Zhou J., et al., Graph neural networks: A review of methods and applications. AI open, 2020. 1: p. 57–81.
- 38. Liu Q., et al., DeepCDR: a hybrid graph convolutional network for predicting cancer drug response. Bioinformatics, 2020. 36(Supplement_2): p. i911–i918.
- 39. Jiang L., et al., DeepTTA: a transformer-based model for predicting cancer drug response. Briefings in bioinformatics, 2022. 23(3): p. bbac100. pmid:35348595
- 40. Zhan Y., et al., iBT-Net: an incremental broad transformer network for cancer drug response prediction. Briefings in Bioinformatics, 2023. 24(4): p. bbad256.
- 41. Wang J., et al., Reversal effect of ALK inhibitor NVP-TAE684 on ABCG2-overexpressing cancer cells. Frontiers in oncology, 2020. 10: p. 228. pmid:32175279
- 42. Tutt A.N., et al., Adjuvant olaparib for patients with BRCA1-or BRCA2-mutated breast cancer. New England Journal of Medicine, 2021. 384(25): p. 2394–2405. pmid:34081848
- 43. Seo T., et al., Response to dabrafenib and trametinib of a patient with metaplastic breast carcinoma harboring a BRAF V600E mutation. Case Reports in Oncological Medicine, 2020. 2020. pmid:32206360
- 44. Kullberg M., Mann K., and Anchordoquy T.J., Targeting Her-2+ breast cancer cells with bleomycin immunoliposomes linked to LLO. Molecular pharmaceutics, 2012. 9(7): p. 2000–2008. pmid:22621404
- 45. Brisson B.K., et al., Tumor-restrictive type III collagen in the breast cancer microenvironment: prognostic and therapeutic implications. Research Square, 2023. pmid:37090621
- 46. Chemmalar S., et al., Busting the Breast Cancer with AstraZeneca’s Gefitinib. Advances in Pharmacological and Pharmaceutical Sciences, 2023. 2023. pmid:38090376
- 47. Guo L., et al., Neratinib for HER2-positive breast cancer with an overlooked option. Molecular Medicine, 2023. 29(1): p. 134. pmid:37803271
- 48. Huang S.-W., et al., Cancer as an infectious disease: A different treatment alternative using a combination of tigecycline and pyrvinium pamoate–An example of breast cancer. Journal of Microbiology, Immunology and Infection, 2022. 55(1): p. 51–59. pmid:33610508
- 49. Jung K., et al., TM4SF4 and LRRK2 are potential therapeutic targets in lung and breast cancers through outlier analysis. Cancer Research and Treatment: Official Journal of Korean Cancer Association, 2021. 53(1): p. 9–24.
- 50. Singh B., et al., Sensitization of resistant cells with a BET bromodomain inhibitor in a cell culture model of deep intrinsic resistance in breast cancer. Cancers, 2023. 15(7): p. 2036. pmid:37046697
- 51. Yang F., et al., Ferroptosis heterogeneity in triple-negative breast cancer reveals an innovative immunotherapy combination strategy. Cell Metabolism, 2023. 35(1): p. 84–100. e8. pmid:36257316
- 52. Lochmann T.L., et al., Venetoclax is effective in small-cell lung cancers with high BCL-2 expression. Clinical cancer research, 2018. 24(2): p. 360–369. pmid:29118061
- 53. Luo J., Shen L., and Zheng D, Association between vitamin C intake and lung cancer: a dose-response meta-analysis. Scientific reports, 2014. 4(1): p. 6161. pmid:25145261
- 54. Chang H.-W., et al., Therapeutic effect of repurposed temsirolimus in lung adenocarcinoma model. Frontiers in pharmacology, 2018. 9: p. 778. pmid:30087612
- 55. Zhao N., et al., Methylation-derived inflammatory measures and lung cancer risk and survival. Clinical epigenetics, 2021. 13(1): p. 1–12.
- 56. Chen M., et al., MYH9 is crucial for stem cell-like properties in non-small cell lung cancer by activating mTOR signaling. Cell Death Discovery, 2021. 7(1): p. 282. pmid:34635641
- 57. Kim J.H., et al., Dexamethasone inhibits hypoxia-induced epithelial-mesenchymal transition in colon cancer. World Journal of Gastroenterology: WJG, 2015. 21(34): p. 9887.
- 58. Lee P.-C., et al., Development of triazole-based PKC-inhibitors to overcome resistance to EGFR inhibitors in EGFR-mutant lung cancers. American Journal of Cancer Research, 2023. 13(10): p. 4693–4707. pmid:37970360
- 59. Lendeckel U., et al., The Role of the Ectopeptidase APN/CD13 in Cancer. Biomedicines, 2023. 11(3): p. 724.
- 60. Salehi B., et al., Liposomal cytarabine as cancer therapy: from chemistry to medicine. Biomolecules, 2019. 9(12): p. 773. pmid:31771220
- 61. Schuetze S.M., et al., Results of a phase II study of sirolimus and cyclophosphamide in patients with advanced sarcoma. European journal of cancer, 2012. 48(9): p. 1347–1353. pmid:22525224
- 62. Barbie D.A., et al., Phase 1B study of momelotinib combined with trametinib in metastatic, kirsten rat sarcoma viral oncogene homolog-mutated non–small-cell lung cancer after platinum-based chemotherapy treatment failure. Clinical lung cancer, 2018. 19(6): p. e853–e859. pmid:30087028
- 63. Yang W., et al., Genomics of Drug Sensitivity in Cancer (GDSC): a resource for therapeutic biomarker discovery in cancer cells. Nucleic acids research, 2012. 41(D1): p. D955–D961. pmid:23180760
- 64. Basu A., et al., An interactive resource to identify cancer genetic and lineage dependencies targeted by small molecules. Cell, 2013. 154(5): p. 1151–1161. pmid:23993102
- 65. Seashore-Ludlow B., et al., Harnessing connectivity in a large-scale small-molecule sensitivity dataset. Cancer discovery, 2015. 5(11): p. 1210–1223. pmid:26482930
- 66. Barretina J., et al., The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity. Nature, 2012. 483(7391): p. 603–607. pmid:22460905
- 67. Ghandi M., et al., Next-generation characterization of the cancer cell line encyclopedia. Nature, 2019. 569(7757): p. 503–508. pmid:31068700
- 68. Kim S., et al., PubChem 2023 update. Nucleic acids research, 2023. 51(D1): p. D1373–D1380. pmid:36305812
- 69. Weininger D., SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. Journal of chemical information and computer sciences, 1988. 28(1): p. 31–36.