Hello, I’m Kemal!

In my master’s thesis, I developed two novel deep learning models, VAE-DeepDep and MAE-DeepDep, to improve the performance of the existing DeepDEP model in predicting gene dependencies in cancer cells. VAE-DeepDep, utilizing variational autoencoders, demonstrated the highest prediction accuracy, while MAE-DeepDep, based on masked autoencoders, also outperformed the original model. Extensive datasets, including TCGA, DepMap, and MSigDB, were used for training and validation, ensuring robust evaluation of the models. Key parameters such as the beta-value in VAE-DeepDep and the mask ratio in MAE-DeepDep were optimized, with minimal performance differences observed across variations. Input dropout analysis revealed that fingerprint data contributed the most to model performance, emphasizing its importance in cancer dependency prediction. Synthetic lethality analyses of four genes provided insights into potential therapeutic targets, further advancing the applicability of the models in cancer research.

In this study, three different language models were run in two different propaganda detection tasks and the results were compared. The models used are DeBERTaV3, LSTM with BERT Word Embeddings (LBWE) and SVM. By selecting these models, a transformer-based model, a model whose word embedding is provided by a transformer-based model but uses LSTM in classifica tion, and a traditional machine learning model that works with word2vec word embeddings were compared. The first task to which these models are applied is to classify whether the sentence contains propaganda or not. The second task is to classify which type of propaganda the sentences containing propaganda contain. In the results obtained, while DeBERTaV3 showed the best performance in the first task, the LBWE model showed the best performance in the second task. In both tasks, SVM achieved the lowest F1 and accuracy scores.

In this project, the GDE3 and NSGA-II algorithms were applied and compared on the Multi-Objective Capacitated Facility Location Problem (MOCFLP). The study aimed to minimize logistics costs and CO2 emissions, revealing that while the GDE3 algorithm performed faster, the NSGA-II algorithm generally achieved more diverse and successful results. The project is significant as it marks the first application of GDE3 to MOCFLP and provides detailed insights into the performance of both algorithms.

In this project, stock portfolio optimization was performed using a Genetic Algorithm (GA) on 15 selected S&P 500 stocks to maximize the Sharpe ratio. The portfolio, tested with 2023 stock prices, achieved a return approximately 2.55 times higher than the S&P 500 index, demonstrating successful results. Additionally, the effects of crossover rate, mutation rate, and population size on convergence time were analyzed, revealing that reducing the mutation rate significantly increases the time to reach the optimal solution.

In this project, six different connection patterns were tested to determine the most efficient design for high-unit-number homeostats, focusing on adaptation time, complexity, and robustness. The analysis showed that sparse connectivity, importance-based connectivity, and the Watts-Strogatz model performed similarly with lower complexity than the fully connected pattern, which was the most complex. Sparse connectivity was the most robust for 4-7 and 9 units, importance-based connectivity excelled at 8 units, and the Watts-Strogatz model was the most robust at 10 units, while importance-based connectivity outperformed others in adaptation time for high unit numbers.