Due to recent technological advances, single-cell assays are now firmly accepted as state-of-the-art for a wide range of biomedical projects; RNA-focused assays represent a cost-effective and quick proxy for both DNA and protein evaluations, as well as containing intrinsic expression signals. Clustering is pivotal in scRNA-seq for partitioning cells into distinct groups, setting the stage for analyses such as trajectory inference of cell-cell interactions. Recently, community detection algorithms emerged as superior for identifying transcriptomic-driven subpopulations. Despite their enhanced performance, all cutting-edge community detection techniques are inherently stochastic and crisp; iterative runs result in variable partitions and interpretations even when applied to identical inputs. This variability highlights the need to replace crisp assignation with a probabilistic/fuzzy approach, i.e., fuzzy clustering, where cells can concurrently be associated with multiple clusters. Inspired by findings from Gribben et al. \citep{Gribben2024} on biphenotypic cells in metabolic dysfunction-associated diseases, in this project I introduce \textit{flufftail}, a comprehensive R package designed for single-cell data analysis through the angle of fuzzy clustering that aims to characterise fuzzy cells/genes and subsequently uncover insights on mechanisms that drive cellular plasticity and GRN dynamics. \textit{flufftail} exploits the variability of standard clustering approaches by proposing a fuzzy community-detection clustering through repeated stochastic clustering. I also developed a new methodology for identifying key genes (major regulatory hubs) that drive biological transitions through fuzzy gene module clustering. Furthermore, \textit{flufftail} presents a new approach for characterising gene regulatory network (GRN) dynamics and the evolution of regulatory interactions for major regulatory hub genes at different stages of the pseudotime ordering of cells. Importantly, a Shiny interface was developed, allowing interactive analysis with the package and democratising it to researchers from diverse backgrounds. This marks the first exploration/exploitation of GRN dynamics across different points (or bins) of pseudotime in single-cell data, advancing our understanding of the mechanisms controlling cellular plasticity, transdifferentiation, and potential therapeutic targets.

Main technologies used:

R Seurat R Shiny C++

ClustAssessPy offers a data-driven approach for optimizing parameter values across all stages of graph-based community detection clustering in single-cell datasets. This Python package is a lighter adaptation of ClustAssess (R), incorporating its main functions and can be used by the Scanpy community to guide robust clustering through data-driven selection in all community detection clustering steps:

• Dimensionality Reduction: Selection of the base embedding (UMAP vs PCA) and the number and type of features (e.g., highly-variable vs most abundant).
• Graph Type: Choice of graph type for the adjacency matrix (nearest neighbors vs shared nearest neighbors) and the number of neighbors.
• Clustering: Identification of the most stable algorithm (Leiden or Louvain) and the appropriate resolution value.

Main technologies used:

Python NumPy ScanPy C++

Cancer is a leading cause of death worldwide, causing nearly one in six deaths. Resistance to therapy continues to be a primary cause of cancer treatment failures, resulting in a high number of cancer-related deaths. Recent technological advancements and the availability of a vast amount of omics data have ushered us into a new era of precision medicine, a paradigm shift away from one-size-fits-all cancer treatment toward patient-specific medicine. The main aim of this project was the in-depth statistical analysis of the NUC-7738 drug using multiple omics datasets and the development of machine learning (ML) models that can be used to predict in silico whether an individual would be sensitive or resistant. Through data analysis and visualisation, tumour origins that are likely to be sensitive and resistant to the drug were identified. Extensive correlation analysis identified genes that highly correlate with drug efficacy across different omics datasets (gene expressions, alternative polyadenylation, proteomics, gene effect, mutations). Furthermore, through a range of feature selection and ML methods, predictive models were created that can be used to predict drug response given a patient’s molecular characteristics. A multi-omics model that incorporates all the models’ predictions was also attempted, with modest success. Clustering cancer cell lines based on their alternative polyadenylation usage yielded two clusters with varied drug sensitivity. Ensemble clustering of the gene effect dataset revealed genes with similar dependency profiles, which is especially beneficial when it comes to identifying alternative druggable targets for genes of interest that are not conventionally druggable. Finally, correlation analysis when the drug was combined with two other promising drugs was performed. This showcased that when the drug was used in combination with Paclitaxel, the HINT1 gene, which is required to activate NUC- 7738, had a significantly stronger correlation with drug efficacy.

Main technologies used:

Python SciKit-learn Tensorflow SciPy NumPy Seaborn Matplotlib

Important note:

All of the code I have written is in a private repository, since the dataset is owned by NuCana and cannot be shared. Click below to explore one of the eight Jupyter notebooks I wrote, as a sample.

This project involved creating a context-free grammar (CFG) for a subset of English, focusing on parsing declarative, conditional, and interrogative sentences. The initial CFG was designed to generate and parse sentences while assigning appropriate syntactic structures. To improve grammatical accuracy, number agreement was implemented, ensuring consistency between nouns and verbs (e.g., singular/plural forms). Additionally, subcategorization features were introduced to specify valid argument structures for verbs, eliminating ungrammatical outputs like "Arthur loves thinks." The grammar was incrementally developed, tested on positive sentences (e.g., "Arthur loves Guinevere") and negative examples (e.g., "Guinevere love the horse") to ensure correctness and robustness. Recursive rules allowed handling of nested phrases and multiple adjectives dynamically. Testing included both hand-drawn parse trees for validation and automated parsing for complex sentences. While the grammar captures a limited subset of English, it demonstrates the principles of CFG engineering, number agreement, and subcategorization. Reflections highlight areas for potential improvement, such as better handling of tense and more complex structures. This project showcases practical applications of computational linguistics in building dynamic, linguistically-informed grammars.

Main technologies used:

Python NLTK

Following a discussion with a physicist friend, I created a simple website that allows you to use quantum random numbers to escape determinism!
The inability to predict the outcome of a process is defined as randomness. When rolling a die, for example, it is impossible to predict what number will come up. It is important to note, however, that randomness is not inherent in the process. A dice roll obeys the laws of mechanics and can be predicted if the initial conditions are known precisely and computing power is sufficient. Similarly, when a computer chooses an option at random, the numbers remain deterministic because they are generated using a seed value, which is required to initialize the algorithm. If the same seed value is used, the same sequence of "random" numbers will be generated. While this is adequate for some applications, the outcome can be predicted if the seed value and algorithm are known. Processes that follow precise laws are referred to as deterministic.
Laplace famously stated that if an intelligence had complete knowledge of the universe at a given instant and knew the laws of nature, it would be able to predict all future and past events. This means that everything we do in life, including you reading this text, is determined by the initial conditions of the universe.
Later, physicists discovered quantum phenomena, which are fundamentally random and involve elementary particles such as electrons and photons. No matter how much we know about the initial conditions, the outcome of these processes cannot be predicted. However, quantum phenomena exist only at the microscopic level and have no effect on processes at the human scale.
Here's where our website comes in. Quantum random number generators use the inherent randomness of quantum processes to generate a sequence of random numbers that, in theory, cannot be predicted. We use this to quantum randomly select one of the options you enter, allowing you to bring quantum randomness to the macroscopic scale and theoretically enjoy experiences and memories that are truly free of determinism.

Main technologies used:

HTML/CSS JavaScript Node.js Firebase

My prototype is a smart physical device that assists students in studying using the Pomodoro technique. I have been using this technique to divide my study sessions into small, highly focused chunks separated by short breaks since I was fifteen years old. I have always wished for a system that would allow me to track this habit while also visualising and rewarding my progress in a meaningful way. This prototype is based on this concept and provides a complete system to accomplish that.

Features:

• Its main component is powered by an Arduino Uno and includes a holographic timer and alarm system (such the name Holodoro).
• Semi-autonomous plant watering system that is watering the user’s favourite plant based on their studying habits, powered by an Arduino micro.
• Wireless component on this system located outside of the room that turns LED lights red whenever the user is studying, and green whenever they are having a break (do-not disturb feature).
• System is mainly controlled using voice commands (NLP).
• Uses Firebase for its authentication and database purposes, and provides a web interface using p5.js.
• Other smart features include: measuring plant growth over time and generating a visual comparison. Automatic pause of a study session whenever the user leaves the computer, powered by ML.

Main technologies used:

HTML/CSS JavaScript ml5.js Firebase p5.js Arduino