This track collection shows scores from GPN-Star, a genomic language model that predicts how strongly natural selection constrains each base of the genome, and how damaging each possible single-base substitution is likely to be. GPN-Star learns from whole-genome alignments of many species together with their evolutionary tree, so it combines the cross-species conservation signal used by PhyloP and PhastCons with the surrounding sequence context. Scores are available for every possible single-base substitution, in coding and non-coding sequence alike.
Three GPN-Star models were trained for the human genome, each on an alignment that covers a different span of evolution:
The authors found that the best model depends on the question. In their benchmarks the vertebrate model performed best on pathogenic missense variants (ClinVar) and on somatic cancer variants (COSMIC). The mammal model performed best on pathogenic non-coding variants (OMIM and HGMD) and on fine-mapped GWAS variants. The primate model showed the highest heritability enrichment for complex traits. Like PhyloP and PhastCons, GPN-Star measures evolutionary constraint over a given span of time rather than pathogenicity as such, so a highly constrained score does not by itself mean a variant causes disease. Other variant effect predictors are in the Deleteriousness Predictions collection.
The -LLR graphs for all three models are in one track, and each model has its own sequence logo track. When the collection is turned on, only the mammal model (M) is shown. The vertebrate (V) and primate (P) -LLR graphs can be turned on from the -LLR track's configuration page, and their logos from the configuration page of this collection. The label of every graph includes its model letter.
-LLR graphs. There are three possible substitutions at each position, so the scores are split across four graphs, one per alternate allele. The graph labeled "(M) Mutation: A" shows the score for changing the reference base to an A, and so on. At each position the graph matching the reference base carries no substitution and is shown as zero. The score is the calibrated log-likelihood ratio (LLR) of the alternate versus the reference allele, displayed with its sign flipped (-LLR), so that higher bars mean a substitution the model considers more damaging. The default view is scaled from -3 to 10; values outside that range are clipped at the edge of the graph, and the range can be changed on the track configuration page. In dense mode the graphs are shown in grayscale. When expanded, they use these colors:
| Positive -LLR: the alternate allele is less likely than the reference, suggesting a constrained position or a damaging change | |
| Negative -LLR: the alternate allele is more likely than the reference |
The scores are calibrated so that a substitution behaving like those at neutral sites scores about 0.
Sequence logo. At each position the logo stacks the four bases, with each letter drawn in proportion to how strongly the model favors it. The total height of the stack shows how constrained the position is: a tall stack, up to a maximum of 2, means the model is confident about which base belongs there, while a flat or empty column means the position looks neutral. The letters are always stacked in the same order, so the base the model favors most is the tallest letter, not necessarily the top one. In tall, constrained columns this is almost always the reference base; in short columns, where no base is strongly preferred, it often is not. The logo is meant for viewing: it shows which positions are constrained, but it levels off at a height of 2, so a substitution scored at -5 and one scored at -10 look almost the same. Use the -LLR graphs to compare how damaging specific substitutions are. The letters use the standard nucleotide colors:
| A | |
| C | |
| G | |
| T |
Individual scores are only visible when zoomed in to a few hundred bases. At wider zoom levels each pixel summarizes many bases, and the graphs show the average with whiskers for the range.
Scores are only available on the main chromosomes. The mitochondrial genome and unplaced, unlocalized or alternate sequences have no scores.
GPN-Star is a transformer model trained to predict masked bases in a multiple-species whole-genome alignment, using the species tree to weight related species by their evolutionary distance. The LLR of a substitution is the log ratio of the predicted probabilities of the alternate and reference alleles. To remove the effect of mutation rate, the authors subtracted the average LLR at neutral sites (positions within ancestral repeats with near-zero PhyloP and PhastCons scores) with the same five-base context. The three human models have 200 million parameters each and were trained on the UCSC multiz100way alignment (V), on the Cactus 447-way alignment of 447 mammals (M), and on the 243 primates in that alignment (P). See Ye, Benegas et al. (2026) for details.
The bigWig files were downloaded from the GPN-Star scores dataset on Hugging Face, revision 47e7f051113abab49f04f43f9107cae2cbbfd34d. For each model we used the four signed LLR files (one per alternate allele, with the reference allele set to zero) and the four sequence logo files. The LLR values are the authors' mutation-rate calibrated scores, rounded to three decimal places. The logo heights were computed by the authors from the same calibrated scores: the four base probabilities are a softmax of the LLRs, with the reference base at zero, and each letter's height is its probability times (2 - H), where H is the entropy in bits. The values were not modified at UCSC; the browser flips the sign of the LLR at display time. We checked a random sample of positions against the authors' Parquet tables to confirm that the scores line up with the right base. The download and checking scripts are in makeDb/scripts/gpnStar, the steps are documented in doc/hg38/gpnStar.txt, and the track settings are in trackDb/human/hg38/gpnStar.ra.
The data can be explored interactively with the Table Browser or the Data Integrator. From scripts, the data can be accessed through our API, for example track=gpnStarLlrMA for the mammal model's scores for substitutions to A.
For automated download and analysis, the bigWig files can be downloaded from
our download
server, in one subdirectory per model (v100, m447, p243). The
files llr_A.bw to llr_T.bw hold the LLR scores and A.bw to
T.bw the logo heights. Scores for a region can be extracted with our tool
bigWigToBedGraph, which can be compiled from the source code or downloaded as a
precompiled binary for your system. Instructions for downloading source code and binaries can be
found here. For example:
bigWigToBedGraph -chrom=chr17 -start=43044294 -end=43044394
https://hgdownload.soe.ucsc.edu/gbdb/$db/gpnStar/m447/llr_A.bw stdout
The full-precision scores, including a separate calibrated entropy score per position that is not shown here, are available as Parquet tables from the GPN-Star scores dataset. Note that those tables use one-based positions, while the bigWig files and the Genome Browser use zero-based start coordinates. The trained models and the code are available from the GPN-Star collection on Hugging Face and from GitHub.
Thanks to Yun Song, Gonzalo Benegas, Chengzhong Ye and colleagues at the University of California, Berkeley for making these scores available and for providing a track hub.
Ye C, Benegas G, Albors C, Li JC, Prillo S, Fields PD, Clarke B, Song YS. Predicting genome-wide functional constraints with GPN-Star. Nature. 2026 Sep 9. DOI: 10.1038/s41586-026-11005-5; PMID: 42717086