Xavi Loinaz
Email · LinkedIn · Google Scholar · GitHub
Hi there! I'm an associate computational biologist at the Broad Institute of MIT and Harvard and Massachusetts General Hospital where I'm really grateful to be co-mentored by Profs. Gad Getz and Esther Rheinbay. My work currently focuses on computational cancer genomics, with a special focus on structural variation. Previously, I studied Computational Biology and Engineering at Brown, where I'm very lucky to have been advised by Prof. Ritambhara Singh.
I'm also currently applying for PhD programs where I have the support of the NSF CSGrad4US fellowship ($159,000 total)!
SVelfie (Structural Variants enriched with likely functional/impactful events, pronounced "SVEL-fee") is a statistical method to infer structural variant (SV) driver genes in cancer genome cohorts by analyzing the enrichment of likely functional effects based off breakpoint arrangements. This is an alternative approach to traditional approaches focusing on SV breakpoint frequencies, and it can potentially be used to complement such traditional methods. SVelfie recapitulates known drivers and yields novel candidates, and was applied in a Nature Genetics paper that was accepted in principle (see publication below).
This study presents a comprehensive genomic analysis of multiple myeloma (MM) and its precursor stages, identifying candidate drivers from mutations, copy number alterations, and structural variants from 1,030 patients. Using these findings, an “MM-like” score was developed that predicts disease progression, risk classification, and subclonal dynamics, offering insights to improve early intervention and monitoring strategies. The method I developed, SVelfie (see one above) was used to infer structural variant drivers integrated into this MM-like score, and we also found specific mutational aetiologies associated with such SV candidate drivers (see above graphic).
This paper develops machine learning models to classify new cancer samples into predefined TCGA molecular subtypes using multi-omic data from 8,791 tumor samples across 26 cancer cohorts. By validating the models with external datasets and providing containerized versions of the top-performing classifiers, this work attemps to enable more accessible and accurate application of molecular subtyping in clinical and research settings.
Introduces GC-MERGE, a graph convolutional model that integrates long-range genomic interactions and local regulatory factors (specifically histone modifications) to predict gene expression while capturing the 3D structure of the genome. By applying the model to multiple cell lines, the study demonstrates state-of-the-art predictive performance and interprets the biological mechanisms behind its predictions, offering potential insights into the local and long-range regulatory factors influencing gene expression.
Introduces XL-MERGE, a modified version of GC-MERGE with three different mechanisms for improvement, outperforming GC-MERGE for all three tested cell lines in terms of AUROC.
I worked on two different research projects: one observing protein crystallization in the presence of gold nanoparticles, and one on creating gadolinium-oxide MRI contrast agents for characterizing glioma. For the latter, I determined a way to synthesize and demonstrate desired PAMPS-LA polymer surface coating of desired size, an issue that had been unresolved in the lab for approximately a year.
A brief article I wrote on evaluating different Sabermetric pitching statistics for predicting future pitcher performance.
Back in high school, I was a research intern in the neurosurgery department of Kaiser Permanente at Redwood City, California. I developed software to automatically parse video of deep brain stimulation based off the importance of certain parts of the surgical procedure. The image above is a snpashot of the type of video content I worked with.
