Scrambling in the Tree of Life
Self-supervised learning of genome-rearrangement patterns from pairwise whole-genome alignments.
As genomes diverge they rearrange: inversions, translocations, duplications, fissions and fusions. This project asks whether a self-supervised model can learn a meaningful geometry of these rearrangements directly from alignment structure, without labelled topologies. Supervised by Prof. Nicholas Luscombe and Dr. Charles Plessy.
Approach
- Data. Each pairwise whole-genome alignment becomes a 5D point cloud of aligned blocks: genomic coordinates, strand and alignment quality.
- Model. A Set Transformer models interactions between blocks, and Pooling by Multihead Attention maps each variable-size set to one embedding.
- Objective. VICReg: invariance across views, variance against collapse, covariance to decorrelate dimensions.
- Scale. PyTorch DDP on 4× A100 GPUs on OIST’s Saion cluster, with BF16, Flash Attention and
torch.compile.
Status
Ongoing. I am currently testing whether the learned clusters reflect genomic topology rather than domain identity.