Contents II. Introduction
Since cancer is a common human genetic disease caused due to irregular growth of human cells. The human genetic system is so complex that makes it really challenging to treat cancer. There is no universal medicine that works for every patient, each patient responds to the drug in a different way. Personalized medicine uses human genomic profiles or proteins to thwart and diagnose disease. Genomic information is used to observe an individual’s responses to drugs. If a gene variant is associated with specific drug response in a patient, doctors then make a decision based on genetics by adjusting the prescribed amount of drug or picking a different drug. The complex mechanism of drug action and the high heterogeneity of cancer lessens the response rate of most anti-cancer drugs. Drug sensitivity states the concept that the bacteria cannot be produced if the drug is present and demonstrates that the drug or antibiotic is effective for those bacteria or cell lines [1] . In this research, XAI is discussed as a method that can be used in the diagnosis and analysis of drugs. The proposed approach is presented with the intention of attaining transparency, accountability, and model improvement in drug sensitivity prediction on cancer cell lines. In particular, Explainable Artificial Intelligence (XAI) has been recognized as a practical working method for concluding the relevance of important features when making predictions using Machine Learning (ML) models having high local fidelity. Thus XAI findings could lead to accurate clinical predictions. The response of the drugs being different is factorized into signaling pathway drug target feature [2] . It is an important measure for determining the drug as being sensitive or resistant. The association between drugs and human genomics profile is unveiled by executing HTS-High throughput screening and is open-sourced, and accessible in the form of pharmacogenomics datasets like Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Cell Line Encyclopedia (CCLE). In a dataset of Cancer Cell Line Encyclopedia (CCLE), the compilation of chromosomal profiles of 947 human cancer cell lines and profile of 24 anticancer drugs around 500 cancel lines to support personalized medicine for cancer patients. What makes it extra helpful in drug response prediction is that the cell line in CCLE and Genomics of Drug Sensitivity in Cancer (GDSC) datasets have acquired the data from various types of cancer tissues. Some of which are lungs, brain, kidney, breast. The main idea is to predict the drug responses of cancer cell lines using GDSC and CCLE datasets and assess the drug as being sensitive or unaffected in terms of IC50 value. In our research, we modeled an architecture inspired by deep learning for drug response motivated by the expeditious development of deep learning technology. The prediction of cancer cell lines is achieved by incorporating chemical profiles of compounds and genomic profiles of cell lines. The approach involved the prediction of the response values IC50, the model comprises a neural network, which worked in an unsupervised way to extract cell lines features from gene expression data. The dimension of gene expression data is extremely massive, hence the Pearson correlation was used to reduce dimensions. Afterward, to make drug sensitivity data of the particular cell line-compound pairs, the chemical features of compounds were incorporated into the model. For evaluation 10-fold cross-validation technique was used. Coefficient of determination and Root Mean Square Error (RMSE) were the loss functions. The Shapley values were calculated which can evaluate feature importance for a specific prediction for any model. However, XAI is a niche focused on enhancing the explainability and clarity of AI algorithms. The popular XAI methods which include ‘Local Interpretable Model-Agnostic Explanations’ (LIME) and ‘SHAPely Additive Explanations’ (SHAP) have proven fruitful in the interpretation of black box models. A lot of methods have been proposed in recent studies, such as modular relations among genomics features and models built on deep neural networks [3] . In this research, first the gaps were investigated. We then came up with a more improved algorithm that can efficiently address the gaps of earlier research. The main focus was to translate critical information which was produced during prediction so that personalized medicine can be further improved and medical practitioners can use that translatable information [4] more effectively. Moreover, the model reveals that deep learning [5] can significantly promote the study of drug sensitivity prediction. To learn the molecular basis of drug sensitivity [6] , gene expressions and cancer cell lines with diverse genomics background is studied. Multiple concepts of AI have also been effectively implemented for personalized drug discoveries in earlier years. Quite extensive research has been done on these data sets, for instance, anticancer drug repositioning [7] Cancer pathogenesis analysis [8] etc. Though, extensive experiments have proved that modeling cell line features alongside chemical features may bring great results for drug sensitivity predictions. In this literature review, the majority of the discussed approaches have predicted drug response by using machine learning algorithms. In [9] Aman proposed a method for mapping nonlinear interconnection among the response of drugs and gene expressions of cell lines by utilizing multi-task and ensemble learning. In [10] ensemble technique was suggested. Which was built on rotation forest, which is a technique based on feature extraction for producing classifier ensembles. [11] Liu’s combination model accomplished higher prediction accuracy as well as great interpretation ability. The ensemble learning method by Mehmat [12] gave promising results. He also proposed drug activity signatures and cell line sensitivity by mixing multiple databases and observed its effect on the performance of the response variable. As compared to traditional machine learning algorithms of SVM and Random Forest [13] their model accomplished significantly improved performance. Even though the DL model is effectively used in numerous applications, the use of large and complex datasets restricts its use in integrating genomic feature sets to predict the drug response. To handle this challenge [14] suggested a deep neural network model built on mutation information and the expression profile of the cancer cells. The model included three networks (mutation encoder, expression, and feed-forward network), where mutation encoders and the expressions were mutually responsible for dimensionality reduction. The reduced dataset was then put into the FFNN for predicting drug response based on the IC50 values. Towards Explainable Anticancer Compound Sensitivity Prediction via Multimodal Attention-based Convolutional Encoders [15] They proposed the architecture of anticancer compound sensitivity for explainable prediction by utilizing a multimodal attention-based convolutional encoder. [15] In this study, the authors have worked on presenting a novel framework for patient classification and drug reprofiling by using exploratory data mining and network analysis. This is a step toward building advanced Explainable AI systems. Manica, Ali [16] AI Enables Explainable Drug Sensitivity Screenings. The IBM Researchers established PacMan an insilico stage for screening compounds which is based on the latest developments in AI for computational biochemistry. In their findings, the genes with the highest attention weights were the main participants in the development and treatment of the disease.
