Recently, efforts are exerted on cancer treatment prediction based on the biomarkers related to the tumor. Gene expression and mutation profiles are the most used biomarkers for cancer prediction. Machine learning and deep learning algorithms have been used to predict drug response. The recent research show that the performance of deep learning models is better than the performance of machine learning based one. In this paper, we introduce the use of Convolutional Neural Network (CNN) models to predict different drugs response. DeepInsight algorithm used to convert the input data to images to be more suitable as input to the CNN. We proposed 3 different pretrained CNNs-models (InceptionV3, Xception, EfficientNetB7) with alternatives in their settings in the training process and modification in their architectures to be able to predict the drug response using IC50 regression values. Those models are selected due to their efficiency for ImageNet applications.the proposed modified Xception model achieves the best accuracy over the 2 others. At first, the whole data input passes through DeepInsight which converts the gene expression data and mutation data to images. Dimension reduction is then applied using the helper technique inside the DeepIsignt. Comparative analysis with other Deep models, shows that the proposed approach improve the prediction accuracy in a range between 14% and 22% as a reduction in mean squared error (MSE)
15.00 Jean-Claude Guédon (Université de Montreal), Universal, not Global, science: why Open Access is crucial to the good health of scientific communication. Since the WWII, the landscape of scientific communication has quietly, yet deeply, changed. The advent of the Science Citation Index in the ’70’s, the growth of commercial publishers, and, finally, the fundamental affordances of digitization and the Internet have converged to create an oligopolistic structure to scientific communication. Scientific communication has thus been globalized in ways that are not all that different from economic globalization. Open Access may be a way to correct this situation, but only if some further conditions are also respected and implemented. The presentation will explore some of these conditions to reach the goal of universality for science. Giovanni Destro Bisol (Università di Roma “La Sapienza”), The emerging complexity of data sharing: lessons from human genetics and genomics In this lecture, I present an overview of data sharing dynamics with a focus on Human Evolutionary Anthropology. I will first discuss three different views on Open data: “Share or die”, “I just don’care” and “Tension between secrecy and openness”. Then, I will talk about the complexity surrounding open data dynamics in a relatively small research field, Human evolutionary Genetic and Genomics. Thereafter, I will show some results of empirical studies on the data sharing behavior in human genetics and other scientific disciplines, and describe two cases of good practice (Forensic Genetics and Paleogenetics). Finally, I will discuss the implications of arguments and evidence shown here for the future initiatives of the Associazione Italiana per la Scienza Aperta.
Several studies have underscored the importance of ICT governance for organizations that wish to leverage ICT capabilities to achieve strategic goals. This is also true for higher education institutions. However, there is still a lack of consolidated view on factors associated with ICT governance effectiveness in higher education. Therefore, we conducted a systematic literature review of the factors related to ICT governance effectiveness to address this gap. Our finding suggests that three groups of factors are associated with ICT governance effectiveness: Relational mechanism (Office of the CIO, internal communication, and leadership), ICT governance structure (ICT strategy committee, relationship managers, and steering committees), and ICT governance processes (strategic ICT planning, framework, and performance measurement).
Traditionally, pathologists have used a light microscope to examine tissue sections mounted on glass slides to diagnose and categorize brain tumors. The outcomes of this method may not always be perfectly correct and are labor- and time-intensive. The procedure of classifying tumors could be made better using the computer-aided method. In this paper, we present an automated procedure that combines the following three feature extraction techniques to classify histopathological images of brain tumors: color histogram, Hu invariant moments, and scale-invariant hybrid image descriptor (RSHD). Color normalization and augmentation techniques are used for preprocessing the histopathological images before the feature extraction phase. In order to classify brain tumor images, three feature sets were extracted. Next, the generated features were fused together to form a combined feature set fed into the machine learning classifier. For differentiating between different sub-types of brain cancer, the XGBoost model performs best in terms of classification accuracy, precision, F1 score, and recall. XGBoost obtained 97.5% for recall, 94.1% for F1 score, 93.1% for precision, and 92.7% for accuracy. The proposed method produced effective results, outperforming multiple state-of-the-art investigations.
The numerically challenging problem of generating the Jordan canonical form is attacked exclusively for a class of zero-one square matrices with the property that each column has at most one nonzero element. The approach is based on the generation and analysis of the directed graph having the matrix in question as its adjacency matrix. The computational investigation is carried out by developing a MATLAB experimental toolbox. Fortunately, it has been found that this toolbox strikingly outperformed the built-in MATLAB function “jordan” since the former successfully executed for square matrices of order up to 1000 while the latter lingered for matrices of order 60. Since the Collatz matrix generated by the Collatz function is a special case of the zero-one investigated matrix, the obtained results are applicable to this important matrix which appears in some problems in signal and image processing. The open problems related to Collatz conjecture are computationally addressed and some empirical results emerged. An empirical formula is devised for the number of connected components in the Collatz graph.
