Research
Questions, methods, and current collaborations
My research focuses on deep learning for structured scientific data. I currently work with RNA secondary structure, combining predictive models, generative models, and evaluation strategies designed to measure generalization.
Together, these directions ask how a model can learn useful relationships between sequence and structure, represent plausible alternatives, and remain reliable on less familiar data.
Generating RNA Structural Ensembles
RNA can adopt alternative secondary structures, so a single prediction may leave out relevant variability. Structural ensembles make it possible to study diversity, consensus, and uncertainty directly.
I develop conditional discrete diffusion models with two-dimensional neural architectures for sampling RNA secondary structures. My work includes reproducible training and sampling pipelines and ensemble-level evaluation using consensus, energy, and structural-diversity criteria.
Evaluating Generalization in RNA Prediction
Strong performance does not always imply broad generalization. When closely related sequences appear across training and test sets, evaluation can overstate how well a model handles less familiar RNA families.
I study family-aware and sequence-similarity evaluation regimes across RNA datasets. My contribution focuses on reproducible comparison of classical and deep-learning predictors and on identifying when apparent predictive performance depends on the composition of the evaluation split.
Biomolecular Ensembles Beyond RNA
Through the IDPfun2 Marie Skłodowska-Curie Staff Exchanges collaboration, I also apply generative modeling to ordered and intrinsically disordered protein ensembles. This work connects computational sampling with energy-landscape interpretations grounded in statistical physics.
Research Practice
Across these projects, I use reproducible pipelines, document data partitions and evaluation decisions, and compare models under clearly defined conditions. The goal is to separate improvements in the model from effects introduced by the data or evaluation setup.