Poster #42: Tan Long Nguyen B.Comp.

PangenomeFM: Graph-Native Pretraining for Reusable Human Pangenome Representations

Long Nguyen Tan¹, Tomoya Kanno¹˒², Yang Zhao¹, Matthew Jensen³˒⁴, Mark Gerstein³˒⁴, Yufeng Shen⁵˒⁶, Ting Wang⁷˒⁸, Xinghua Shi¹˒² ¹ Department of Computer and Information Sciences, College of Science and Technology, Temple University, Philadelphia, PA, USA ² Institute for Genomics and Evolutionary Medicine, Temple University, Philadelphia, PA, USA ³ Department of Molecular Biophysics and Biochemistry, Yale University, New Haven, CT, USA ⁴ Program in Computational Biology and Bioinformatics, Yale University, New Haven, CT, USA ⁵ Department of Systems Biology, Columbia University, New York, NY, USA ⁶ Department of Biomedical Informatics, Columbia University, New York, NY, USA ⁷ Department of Genetics, Washington University School of Medicine, Saint Louis, MO, USA ⁸ McDonnell Genome Institute, Washington University School of Medicine, Saint Louis, MO, USA

Human pangenome graphs represent genomic diversity through alternative haplotype paths and population-scale patterns of branching and reconvergence, but most genomic foundation models learn primarily from nucleotide sequence and do not directly model this graph organization. We present PangenomeFM, a graph-native self-supervised framework for learning reusable representations from human pangenome graphs. During pretraining, observed connections between genomic segments are masked, and the model reconstructs them using genomic position and the remaining graph neighborhood, without requiring biological labels or nucleotide sequence as encoder inputs. After pretraining, the learned graph representations are frozen and evaluated across chromosomes excluded from training, across independently released pangenome resources, and in downstream genomic prediction tasks. We study candidate cis-regulatory element classification and structural-variant breakpoint classification to determine whether population graph topology provides information beyond genomic coordinates and sequence-based representations. Our analyses suggest that graph-derived representations provide complementary information to sequence models, with particularly clear utility for tasks closely related to genome structure. Cross-resource experiments further examine how well learned graph representations transfer between different pangenome constructions, while topology-based controls distinguish reusable neighborhood information from simple connectivity shortcuts. PangenomeFM demonstrates a framework for learning directly from the native organization of population pangenomes and motivates future genomic foundation models that jointly integrate sequence, graph structure, and haplotype information.