When a new genome is annotated, the step that decides where the genes are is usually a "combiner." A combiner takes three kinds of evidence and turns them into one gene model: statistical gene predictions from the DNA alone, alignments of known proteins from other species, and alignments of RNA reads from the organism itself. Today's combiners treat a protein match as a flat score. A hit to a mouse protein and a hit to a sponge protein count the same if their alignment scores match, and ten hits from one clade count the same as ten hits spread across the tree of life. That throws away the most useful thing about homology evidence, which is where on the tree of life the support comes from. A real exon under long-term selection has support that follows vertical descent from a deep ancestor. A transposable-element fragment or a pseudogene tends to draw scattered, patchy hits. A lineage-specific gene draws hits only from close relatives, or none at all.
We propose to represent the set of protein hits for each candidate exon or intron as a cloud of points in hyperbolic space. Hyperbolic space is a curved geometry in which trees fit with almost no distortion in very few dimensions. We re-center the space so the target genome sits at the origin. Two quantities of the hit cloud then summarize its phylogenetic profile in a target-independent way: its center, which tells us how deep and in which direction the support lies, and its spread, which tells us whether the hits are consistent with inheritance down a tree. A small learned scorer turns these profiles, together with ordinary RNA-read and ab initio features, into per-candidate scores. We then feed those scores into EVidenceModeler's existing dynamic-programming decoder in place of its hand-set weights. The claim is narrow and testable. Most of the gain should appear on exactly the genes where current combiners fail: poorly conserved genes and genes with low expression. And the gain must come from the geometry of the profile, not merely from adding phylogenetic information.
Core Claim
Multi-evidence combiners plateau at the gene level. In the nematode assessment nGASP, combiners reached a median gene-level sensitivity of 78% and a specificity of only 42%, and that paper names poorly conserved orthologs as a leading cause of failure. mGene.ngs shows the other half of the problem. Spliced RNA reads account for 91.6% of its improvement over DNA-only prediction, which amounts to a transcript F-score of 51.8% versus 49.9%. Where reads are absent, which in practice means low-expression genes, it falls back toward ab initio accuracy.
Homology is the natural stream to fill that gap. Yet in EVidenceModeler homology enters as a user-weighted, per-evidence-type score, and the comparative approach of human–mouse exon prediction uses static pairwise alignment scores. Our claim is that a phylogenetically structured summary of homology support separates three kinds of candidate that flat scores confuse: true exons, decoys derived from repeats or pseudogenes, and genuine lineage-specific exons. We further claim that hyperbolic geometry is a good way to compute this summary, because it is compact, differentiable, and invariant to the choice of target species. The combiner can then be trained on annotated genomes and transferred to unannotated ones.
Research Program
The informant set is a large panel of reference proteomes spanning eukaryotes, with their species taxonomy.
Embedding the taxonomy. We embed the taxonomy once into the Lorentz model of hyperbolic space, the sheet-of-a-hyperboloid coordinate system that avoids the numerical blow-up of the Poincaré ball near its edge. We use the distortion-minimizing approach of Poincaré embeddings. That paper supplies the method: it shows that large taxonomies embed in two to ten hyperbolic dimensions with little distortion. Bridging ML and algorithms gives us a way to check distortion against combinatorial constructions.
Re-centering on the target. For each target genome, we apply the hyperbolic isometry (the distance-preserving map) that moves the target species to the origin. After this move, an informant's distance from the origin equals its evolutionary distance from the target, and its direction encodes which clade it belongs to. This is what makes a scorer trained on worm, human, mouse and rice applicable to a coral or a lizard.
Summarizing the hit cloud. For a candidate exon or intron with protein hits from informants at re-centered positions $x_i$ and alignment-derived weights $w_i \ge 0$, we compute the Lorentzian centroid
$$\mu = \frac{\sum_i w_i x_i}{\sqrt{\left|\langle \textstyle\sum_i w_i x_i,\ \sum_i w_i x_i\rangle_L\right|}},$$
where $\langle\cdot,\cdot\rangle_L$ is the Minkowski inner product that defines the hyperboloid. We also compute the weighted mean hyperbolic distance of the hits from $\mu$, which we call the dispersion $\delta$.
These two quantities have direct readings. The distance of $\mu$ from the origin measures how far back the support reaches. Its direction says which lineages support the candidate. When $\delta$ is large relative to the value expected if the same number of hits were spread along a single subtree, the support is patchy, which is the signature we expect from repeat-derived and pseudogene decoys. This use of geodesics bending toward common ancestors echoes the phylogenetic inference work in hyperbolic variational SMC. That work infers trees, whereas we take the tree as given and use the geometry to summarize evidence over it.
Scoring and decoding. A small permutation-invariant network over the hit set produces the final homology representation for each candidate. It takes $\mu$, $\delta$, and the per-hit points mapped to the tangent space at the origin. Its output is concatenated with RNA-read features (spliced-read counts, coverage) and ab initio scores, which all stay Euclidean. This split is deliberate. Hyperbolic multimodal taxonomies found that hyperbolic geometry helped DNA-barcode retrieval but hurt image retrieval, where the features are not tree-like. That result constrains our design: only the evidence that actually lives on a tree goes into curved space.
The scorer is trained with a per-candidate correctness loss against gold-standard exons and introns. Its scores then replace EVM's fixed evidence weights, and EVM's dynamic program still chooses the highest-scoring gene path. We do not claim to recover exons that no upstream tool proposed. EVM's own analysis rules that out for any combiner, and we keep that ceiling.
Relation to other hyperbolic biology work. Other hyperbolic biology models embed the data items themselves: codons inside amino-acid cones in HyperHELM, expression spots and image tiles in DELST, and ligands in HypSeek. We instead embed the provenance of evidence. Because of that, we need no entailment cones, whose sensitivity to their angle and margin settings DELST reports as a practical burden.
Baselines
Naive reference. The simplest reference point is the best single ab initio predictor run alone, together with EVM using its default evidence-type weights on identical inputs. On the rice benchmark of 1,058 cDNA-verified genes, EVM with full evidence reports 92% of exons and 62% of complete gene structures correct (EVM). On nGASP, combiners have a reported median gene-level sensitivity of 78% and specificity of 42%, and the best isoform-level results are about 66% sensitivity and 56% specificity (nGASP). For the ENCODE regions, the file contains only the qualitative statement that EVM matched or exceeded Ensembl, so no reported number was found on Lacuna for EVM on human ENCODE.
We rerun EVM ourselves. It gets the same ab initio outputs, the same protein alignments against the same informant panel, and the same RNA alignments. Its weights are grid-searched on the same training genes our scorer sees. Beating it means a statistically significant gene-level F-score gain on held-out loci of each benchmark, with no loss in base-level accuracy.
Strongest learned competitor. This is a gradient-boosted or small neural combiner. It sees every raw feature we use: per-hit alignment scores, counts of supporting species, and the patristic distance (distance along the tree) from the target to the nearest and farthest hit. It also sees explicit tree statistics computed directly on the taxonomy: the depth of the lowest common ancestor of the hits and a parsimony-style count of how many independent gains the hit pattern would require. Its output feeds the same EVM decoder. This is the competitor that could erase our claim, because it holds the same phylogenetic information in handcrafted form. We run it with the same training data, the same tuning budget in trials, and the same feature extraction. We also run mGene.ngs-style SVM training, the reference for RNA-heavy discriminative combining, on the nematode data, where it reports a transcript F-score of 51.8%.
Ablations that isolate the key ingredient. First, we replace the hyperbolic taxonomy embedding with a Euclidean one of the same dimension and train it the same way. Second, we give the Euclidean embedding eight times the dimensions. Third, we remove the dispersion term. Fourth, we drop the re-centering and use absolute coordinates. If the Euclidean ablation matches us at equal dimension, the geometry claim fails, whatever the absolute gains.
Why not just…?
Why not just add phylogenetic distance and hit counts as features to a learned combiner? This is the most serious threat, and we treat it as the main competitor rather than a straw man. Scalar features can capture how far away the support is. They cannot easily tell ten hits spread down one subtree from ten hits scattered across unrelated clades, and that distinction is the one we expect to separate true exons from repeat decoys. The experiment settles the question directly. We stratify held-out candidates by whether they overlap annotated repeats or known pseudogenes. If the explicit tree-statistics baseline matches the hyperbolic scorer on that stratum, then the useful content is phylogenetic, not geometric. In that case we would report the tree-aware combiner as the result and drop the geometric claim.
Why not just run explicit tree algorithms or a graph network over the taxonomy for every candidate? An explicit tree computation would be exact where our embedding is approximate. The cost is that every candidate requires a pass over a taxonomy with thousands of nodes, and a genome produces millions of candidates. Our summary costs time proportional to the number of hits, and the re-centering step makes transfer to new targets automatic. The exact tree statistics are already included in the strongest baseline. So the comparison reports both accuracy and wall-clock cost, and a tie in accuracy at much lower cost is the honest best case for the geometric version.
Why not just tune EVM's weights more carefully, or rely on RNA reads as mGene.ngs does? Weight tuning cannot change the fact that one weight applies to an entire evidence type, whatever the phylogenetic makeup of the hits. RNA reads are absent exactly where they are most needed: mGene.ngs gains almost nothing over DNA-only prediction once spliced reads are removed. Our low-expression experiment withholds RNA evidence from the lowest expression deciles. It will show whether structured homology recovers accuracy there while tuned EVM and the SVM do not.
Why not just use a larger Euclidean embedding of the taxonomy? The volume argument predicts that a Euclidean embedding needs many more dimensions to keep deep leaves apart, and Poincaré embeddings showed low-dimensional Euclidean embeddings of taxonomies collapse. The eight-times-wider Euclidean ablation tests whether capacity alone closes the gap. If it does at modest dimension, the practical case for curvature is weak, and we will say so.
Proposed Experiments and Decisive Checks
Benchmarks and splits. We train and evaluate with leave-one-genome-out splits across four gold standards. The first is the nGASP 10 Mb C. elegans test region scored against WormBase ref1 for sensitivity and ref2 for specificity at base, exon, gene, and isoform levels. The second is the rice set of 1,058 cDNA-verified genes. The third is human and mouse loci from GENCODE. The fourth is C. briggsae with the mGene.ngs RNA-Seq setup. Holding out whole genomes tests transfer, which is the regime that matters for the non-model assemblies in the file, such as the reef coral and the bearded dragon. Those assemblies lack gold standards, so we use them only for qualitative case studies, such as how many gene models lie in repeat regions and how many lineage-specific models the method keeps.
Decisive checks. There are three. First, on the stratum of poorly conserved genes and the stratum with RNA evidence removed, the hyperbolic scorer must beat tuned EVM, and it must beat the tree-statistics combiner at equal tuning budget. Second, the equal-dimension Euclidean ablation must lose on repeat-overlapping decoys and on lineage-specific exons. We will confirm that this is where the gain sits by checking whether $\delta$ alone discriminates decoys from true exons. Third, performance on a held-out genome must not drop more than the within-genome variation. This tests whether re-centering actually delivers transfer.
Primary metrics. These are gene-level and exon-level sensitivity and specificity as defined by nGASP. We report isoform-level metrics too, but make no claim about them, since the decoder, and hence isoform assembly, is unchanged.
A negative result would still be useful. If explicit tree statistics capture all the gain, annotation pipelines gain a cheap, interpretable homology feature, and the hyperbolic question is settled for this setting.
References
- nGASP - the nematode genome annotation assessment project
- Next generation genome annotation with mGene.ngs
- Automated eukaryotic gene structure annotation using EVidenceModeler and the Program to Assemble Spliced Alignments
- Human and mouse gene structure: comparative analysis and application to exon prediction
- Poincaré Embeddings for Learning Hierarchical Representations
- Bridging ML and algorithms: comparison of hyperbolic embeddings
- Variational Combinatorial Sequential Monte Carlo for Bayesian Phylogenetics in Hyperbolic Space
- Hyperbolic Multimodal Representation Learning for Biological Taxonomies
- HyperHELM: Hyperbolic Hierarchy Encoding for mRNA Language Modeling
- DELST: Dual Entailment Learning for Hyperbolic Image-Gene Pretraining in Spatial Transcriptomics
- Learning Protein-Ligand Binding in Hyperbolic Space
- GENCODE 2025: reference gene annotation for human and mouse
- Re-annotation improved large-scale assembly of the reef-building coral Acropora intermedia
- A near-complete genome assembly of the bearded dragon Pogona vitticeps provides insights into the origin of Pogona sex chromosomes