Genomic language model for variant effect prediction that scores deleterious mutations from a single DNA sequence, with no alignment at inference.
No providers recorded yet. Browse all providers
Genomic language models have long underperformed at a task their protein counterparts handle well: deciding which single-nucleotide changes in a mammalian genome are functionally disruptive. PhyloGPN, from Yun S. Song's group at UC Berkeley and presented at RECOMB 2025, closes much of that gap by changing what the network is asked to predict. Rather than predicting masked or next nucleotides, it predicts the parameters of a nucleotide substitution process at each position, and is trained to maximize the likelihood of the corresponding column of a 447-species whole-genome alignment under a phylogenetic tree.
The design choice that matters is where the alignment enters. In GPN-MSA, the same lab's earlier model, the alignment is an input feature, so scoring a position requires an alignment over the training species — awkward for other organisms and for regions that align poorly to human. GPN-Star, the lab's later phylogeny-aware model, likewise consumes alignment columns at inference. PhyloGPN puts the alignment only in the loss. At prediction time it reads a single 481 bp window of DNA and emits per-site evolutionary rate parameters, making it usable wherever a sequence exists.
Modeling the tree explicitly also removes a compromise its predecessor had to make. Because closely related genomes share alleles by inheritance rather than by constraint, GPN-MSA excluded most primates from training to stop the model from copying them. PhyloGPN accounts for that shared ancestry in the likelihood itself, so primate genomes — the ones carrying information about recent constraint in humans — can stay in.
PhyloGPN is a convolutional network of roughly 83 million parameters, adapted from the CARP protein model and ultimately from ByteNet. Forty residual blocks with dilated, reverse-complement-equivariant convolutions give it a 481 bp receptive field. Training data pairs each position of GRCh38 with its aligned column from the Zoonomia Consortium's 447-placental-mammal whole-genome alignment and a per-site tree pruned to the species present. The loss conditions the phylogenetic likelihood on the human reference base, so the model cannot trivially rely on the central nucleotide, and is optimized through a numerically stable upper bound. Training ran 18 epochs on four NVIDIA A100 GPUs with AdamW.
On ClinVar pathogenic-versus-benign classification, log-likelihood ratios from PhyloGPN reach an AUROC of 0.96, against 0.64 for Caduceus, 0.61 for Nucleotide Transformer and 0.49 for HyenaDNA, matching GPN-MSA. It outperforms all three of those baselines on 24 of 25 deep mutational scanning assays. On the BEND benchmark suite it reaches 0.98 AUROC on Disease VEP — 0.21 above the next-best genomic language model — plus 0.86 on chromatin accessibility, 0.81 on histone modification and 0.95 on CpG methylation. It trails on gene finding, where the 481 bp window is limiting and PhyloGPN-X recovers an MCC of 0.69, and on Expression VEP.
The single-sequence interface makes PhyloGPN practical for prioritizing candidate variants in rare-disease and population genetics pipelines, including in regions where alignment to the human reference is unreliable and alignment-consuming scorers cannot be applied. Its per-position embeddings serve as features for supervised models of chromatin accessibility, histone marks and CpG methylation, giving functional-genomics groups a pretrained representation that needs no alignment infrastructure to use. Weights and the model implementation are published together on HuggingFace, so scoring or fine-tuning requires no retraining.
PhyloGPN's central argument is that the obstacle for genomic language models was the training signal rather than model scale or architecture: an 83M-parameter convolutional network supervised by phylogenetic likelihood matches alignment-dependent scorers and far exceeds much larger sequence-only models on deleteriousness prediction. The framework generalizes beyond this checkpoint, since richer substitution models and larger alignments slot into the same loss. Its main limitation follows from its training data — the Zoonomia alignment covers placental mammals densely but lacks the more distant vertebrates GPN-MSA uses, and PhyloGPN accordingly falls short of GPN-MSA on coding variant pathogenicity and on regulatory variants from OMIM. Its release is also more fragmented than its siblings'. The MIT-licensed songlab-cal/gpn repository contains only a five-line loader for PhyloGPN that defers to trust_remote_code; the 258-line reverse-complement-equivariant architecture, modeling_phylogpn.py, sits in the HuggingFace repository beside the weights under a bare cc tag that names no specific Creative Commons license — the only non-MIT declaration anywhere in the lab's HuggingFace organization, where GPN-MSA and all nineteen GPN-Star checkpoints are tagged MIT. Neither release includes training or evaluation code for PhyloGPN, so the reported numbers are not reproducible end to end.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.