Research
Deep Learning and Machine learning methods for biological systems, with a focus on graph representation learning and physics-informed models.
Research Interests
- Graph representation learning
- Computational biology & systems biology
- Protein dynamics and structure–function relationships
- Knowledge graphs and ontology-aware embeddings
- Scientific machine learning / physics-informed neural networks
Current Research
Physics-Informed Graph Networks for Protein Dynamics
Proteins are naturally represented as contact graphs between residues. Combining this structure with quantum-inspired neural PDE operators, we can predict per-residue flexibility and normal modes directly — bypassing the computational cost of molecular dynamics simulation for many use cases.
This work is implemented in the LGNM server.
Deep Learning for Biomarker Identification
I build knowledge graphs from Gene Ontology data and Gene Ontology annotations, then apply GNN's architectures with attention-based interpretability. The goal is not just prediction accuracy but the ability to surface which molecular subnetworks a model relies on — a prerequisite for clinical translation.
This work is implemented in the CABIgoWEB server.
Open Questions I Think About
- How do we make ontological hierarchies (like the Gene Ontology DAG) truly differentiable, so gradients can flow through them meaningfully?
- How do we fuse biological networks with fundamentally different noise profiles — say, high-confidence physical PPIs and noisy co-expression?
- Oversmoothing in biological GNNs: at what depth does message-passing on PPI networks become uninformative, and what structural priors resist it?
- Physics-informed vs. purely data-driven losses: when does the physics prior help, and when does it constrain the model into worse solutions?
- Evaluation bias: how much of biomarker discovery performance is really just recovering well-annotated, well-studied genes?