Academic Thesis

(Machine Learning in Drug Discovery)

In my thesis on target identification in drug discovery using machine learning, I focused on identifying viable molecular targets specifically for COX-1 and COX-2 enzymes. These enzymes are crucial in inflammatory processes, making them primary targets for developing anti-inflammatory drugs. Traditional methods for identifying these targets are costly and time-consuming, so my research aimed to develop a data-driven machine learning approach to streamline this process.

Data Collection and Preparation
I curated a comprehensive dataset from sources like ChEMBL and the Protein Data Bank (PDB), which included molecular descriptors, structural information, and known interactions between COX enzymes and various compounds. Key features included hydrophobicity, molecular weight, topological polar surface area, and sequence-based descriptors, providing essential data points for predictive modeling.
To prepare the data for analysis, I addressed class imbalance with SMOTE (Synthetic Minority Over-sampling Technique) to balance interactions versus non-interactions. I also applied standardization and normalization techniques to ensure consistency across the dataset and reduce noise.
Exploratory Data Analysis (EDA) Through EDA, I explored correlations among descriptors, refining highly correlated features like molecular weight and surface area to avoid redundancy. Visualizing descriptor distributions allowed me to observe trends specific to COX-1 and COX-2 interactions, guiding feature selection and engineering to capture biologically relevant patterns.

Model Evaluation
Model evaluation was conducted using the following metrics:
Precision Recall F1 Score ROC-AUC Interpretability and Feature Importance To ensure interpretability, I used SHAP (SHapley Additive exPlanations) values, which helped pinpoint the molecular descriptors most influential in identifying COX-1 and COX-2 interactions. Important features like hydrophobicity and polarity consistently aligned with known biological insights into COX binding, supporting the relevance of the model’s predictions.

Results and Conclusions
After fine-tuning, my models demonstrated high precision in predicting COX-1 and COX-2 targets, significantly reducing the need for exhaustive experimental testing. This work underscored the potential for machine learning to efficiently identify therapeutic targets, advancing drug discovery efforts for COX-specific treatments.

Internship

(Machine Learning Engineer @ EdenLibrary)

During my internship, I focused on advanced computer vision tasks, honing my skills in both model experimentation and dataset management. I began by working with YOLOv5 and YOLOv8 models, initially testing YOLOv5’s performance on large vehicle datasets under various conditions like low light and nighttime, which helped me identify key model limitations. This led to the use of CVAT for manual annotation and dataset refinement, ensuring better training data for challenging environments. Through detailed experimentation, I compared the performance of different dataset sizes, gaining insights into how image quantity affects model accuracy.

To address domain adaptation challenges, I implemented Style Transfer using the StyleShot tool, successfully changing images’ lighting conditions to evaluate model robustness across diverse settings. This adaptation allowed me to train and test models using images taken at different times of the day, with notable results recorded on TensorBoard. I also explored PhyCV for physics-based image processing, which enhanced image quality by removing excess light or shadows, providing cleaner inputs for training.

On the research side, I delved into domain generalization, transfer learning, and style transfer techniques, evaluating Visual-Language Models (VLMs) like Kosmos-2 and Florence-2 for potential in accelerating the annotation process. I further experimented with SAM2 on Ultralytics, exploring its segmentation and detection capabilities through bounding box prompts, enhancing the accuracy and speed of object detection.

Throughout the internship, I documented each phase and troubleshooting step, and I wrapped up by creating a comprehensive documentation of the entire workflow. This hands-on experience with machine learning tools and model fine-tuning broadened my expertise in deploying and optimizing computer vision systems across complex conditions.

Paris Amerikanos Machine Learning / Data Science / PhysicsMachine Learning / Data Science / Physics
Paris managed Tasos directly.
"I had the pleasure of mentoring Tasos during his 2-month internship at Eden Library, where he worked as a Machine Learning Engineer.
He quickly adapted to our projects and exceeded our expectations within the first week. Tasos successfully ran and completed three computer vision experiments, showing strong technical skills and great attention to detail. He’s easy to work with, communicates well, and was always eager to learn. His potential in ML & R&D, is clear, and I’m confident he’ll excel in future roles. I highly recommend him for any opportunity in the field!"