About Me
I’m a postdoctoral researcher at the Genome Institute of Singapore (A*STAR), working at the intersection of AI and genomics. I was initially trained as a physicist, before transitioning to computer science applied to biology. Most of my work has focused on developing an AI approach to genome assembly—a process of reconstructing a complete genome sequence from millions of short, noisy sequencing reads. More recently, I became interested in genomic foundation models and exploring whether they learn anything meaningful, or whether they’re just pattern-matching on sequence without capturing real biology.
Outside of work, I keep busy with hiking, rock climbing, and weightlifting. Recently, I also added underwater rugby to the mix, because apparently the rest wasn’t enough.
- Genomic Foundation Models
- Virtual Cell
- Geometric Deep Learning
PhD in Computer Science
University of Zagreb, Croatia
Integrated BSc and MSc in Physics
University of Zagreb, Croatia
My main body of work centers on genome assembly, the computational problem of reconstructing a complete genome sequence from raw sequencing reads. I developed GNNome, a graph neural network framework that reframes assembly as a learning problem on assembly graphs, achieving substantial improvements over state-of-the-art methods on human and other genomes. More recently this has extended to diploid assembly, where the challenge is to resolve not just one but two distinct haplotypes simultaneously.
Alongside this, I’ve become increasingly interested in genomic foundation models and their limitations: do they actually learn meaningful biology? In recent work we show—through entropy and disagreement analysis—that DNA language models behave fundamentally differently from their text counterparts, raising sharp questions about what pretraining on raw sequence actually captures. This field suffers from a lack of honest benchmarks and rigorous failure analysis, which is a direction I intend to keep working on.
