A protein's 3D contacts can be read from evolution. Across a family of related sequences, two residues that touch tend to mutate together, and statistical models of these paired mutations recover contacts without any solved structure. The method fails for orphan proteins, fast-evolving viral proteins and designed variants, because they have few relatives. The usual response is to find more relatives or to substitute a language-model prior. We argue that the count of relatives is the wrong target. When a family has twelve sequences and nine of them come from one recent clade, their shared mutations look like co-evolution but are only inherited history. The quantity that matters is how many independent lineages the alignment samples.
We propose to index protein sequence space in hyperbolic geometry, a curved space whose volume grows exponentially outward and so embeds trees with little distortion. The index is trained to reproduce evolutionary trees. Distance from the origin then encodes how specialized a sequence is, and distance between points encodes divergence. We use the index for two jobs. The first is retrieval: picking relatives for a shallow query that spread across separate subtrees instead of piling into one clade. The second is estimation: a continuous per-sequence independence weight that replaces the fixed sequence-identity threshold used in classical coupling analysis. The central claim is testable. At a matched number of retrieved sequences, lineage-diverse retrieval should raise contact precision on low-homology targets, and the gain should come from the curved geometry itself, not from the retrieval step alone.
Background and what is actually open
Direct coupling analysis (DCA) fits a global statistical model to a multiple sequence alignment (MSA, sequences lined up column by column). A global model can separate direct couplings from indirect correlations passed along chains of contacts, which simple pairwise correlation cannot (Using evolutionary couplings to predict contacts and build structures). Gaussian DCA turns this into a covariance-matrix inversion and reaches about 95% true-positive rate on the top 10 contacts over 50 Pfam families (Fast and Accurate Multivariate Gaussian Modeling of Protein Families). It needs deep alignments, though, and its covariance becomes rank-deficient when few sequences are available. This paper supplies our cheapest estimator and marks where it breaks.
Below roughly 30 sequences, co-mutation across a recent clade mimics compensatory mutation (Protein 3D Structure Computed from Evolutionary Sequence Variation). This observation motivates the proposal. It locates the failure in phylogenetic dependence, not in sample count alone.
Three lines of work attack the shallow regime by adding capacity or priors. MSA Transformer is pretrained on 26M alignments and reads contacts from 8–16 sequences (MSA Transformer). Contact-Distil distills a deep-MSA teacher into a shallow-MSA student fed with language-model embeddings (Contact-Distil). The Co-evolution Transformer learns to down-weight noisy homologs inside a given alignment (Co-evolution Transformer for Protein Contact Prediction). All three take the alignment as given. None of them changes which sequences enter it, and none models the tree those sequences come from.
Protriever does change what enters: it learns a retriever over 62M UniRef50 sequences end to end (Protriever). Its index is a flat inner-product space, however, and it is optimized for fitness prediction, not for lineage independence. It is our closest prior work and our strongest retrieval baseline.
FastCherries and SiteRM show that modeling phylogeny correctly, even with sites treated independently, beats large language models on variant-effect prediction (Ultrafast classical phylogenetic method beats large protein language models). We read this as evidence that phylogeny is underexploited, and we reuse its fast tree construction for training supervision.
Hyperbolic embeddings supply the geometry. Poincaré and Lorentz embeddings reproduce trees in very few dimensions, with norm tracking generality (Poincaré Embeddings; Learning Continuous Hierarchies in the Lorentz Model). We use the Lorentz model because its closed-form distances avoid the boundary instabilities noted in Hyperbolic Neural Networks. The Lorentz paper also marks a constraint we must respect: hyperbolic space distorts cyclic, non-tree data, and a single origin assumes one root. Protein space has many roots, so the index is trained per superfamily region, not as one global hierarchy.
Core Claim
For a shallow query, a retrieved alignment containing $k$ sequences from $k$ independent lineages yields better couplings than $k$ sequences from one clade. A tree-trained Lorentz index finds such sets better than Euclidean embedding search does, because its norm and geodesic structure expose clade membership directly. A tree built afresh on 20 sequences can tell you they are redundant. It cannot tell you where in 62M sequences to find non-redundant ones.
Research Program
The index. We build an encoder from frozen ESM-2 features to the Lorentz hyperboloid. It is trained on deep Pfam families, whose trees we build with FastCherries. Training minimizes the mismatch between hyperbolic distance and patristic (along-tree) distance between leaves. A secondary term aligns norm with root-to-leaf depth under a midpoint rooting.
The retriever. We index UniRef50 in the resulting space. For a query $q$, candidates within a divergence radius are selected greedily to maximize effective sequence count, and the selected sequences are then aligned to the query with a profile HMM.
The estimator. The same geometry defines each sequence's independence weight:
$$w_s = \Big(\sum_{t} \exp\!\big(-d_L(z_s, z_t)/\tau\big)\Big)^{-1}.$$
Here $z_s$ is the Lorentz embedding of sequence $s$, $d_L$ is the Lorentz geodesic distance, the sum runs over all sequences $t$ in the alignment including $s$, and $\tau$ is a bandwidth. This is a smooth, tree-aware version of the classical rule that counts every sequence above 80% identity to another as one. The effective count is $M_{\text{eff}} = \sum_s w_s$. Weighted covariances feed Gaussian DCA and adabmDCA (adabmDCA 2.0), and the retrieved alignments are also passed unchanged to MSA Transformer.
The role of the 1-WL limit. Graph networks that pass messages only between neighboring nodes are bounded by the 1-WL test, a color-refinement procedure for telling graphs apart. Such networks cannot count triangles or cliques (How Powerful are Graph Neural Networks?; Weisfeiler and Lehman Go Topological). This matters for the residual phylogenetic noise. Clade-driven noise tends to appear as dense cliques of spurious couplings among the columns where a clade differs from everyone else, while real packing produces sparse, locally triangulated clusters. Telling the two apart requires counting cliques. We therefore design any learned refiner of the coupling map to operate on residue pairs, not on single residues, and we test that choice directly. This role is a design constraint, not the main contribution.
Baselines
Native alignment (naive reference). The naive reference point is the query's native shallow alignment with standard 80%-identity weighting, run through the same Gaussian DCA, adabmDCA and MSA Transformer pipelines. Any retrieval method must first beat what the target already has.
Protriever-style Euclidean retrieval (strongest competitor). The strongest competitor is Euclidean dense retrieval in the style of Protriever: an ESM-2 embedding with an inner-product index over the same UniRef50 snapshot. It is the only existing learned retriever over this corpus and the direct alternative to our index. Its reported figure is a ProteinGym Spearman of 0.479, versus 0.470 for PoET and 0.439 for EVE. We found no reported Lacuna result for it on contact prediction, so we will run it ourselves. For fairness, both retrievers return the same number of sequences $k$, use the same alignment step, the same downstream predictors, and the same tuning budget for $\tau$ and the retrieval radius. Beating it means higher Top-L long-range contact precision on CAMEO-L at equal $k$ for the majority of $k \in \{8, 16, 32\}$.
Euclidean retrieval with diversity reranking. To separate the retrieval idea from the geometry, we add Euclidean retrieval followed by identity clustering and maximal-marginal-relevance reranking. This baseline captures "just ask for diversity" without curvature.
Curvature ablation. The key ablation trains the identical encoder with the identical tree-distance loss, but in a Euclidean target space of the same dimension. Any gain over this baseline is attributable to curvature alone.
Oracle tree weighting. A FastCherries tree built on each retrieved set supplies classical tree-based weights. This oracle-style estimator tells us whether the hyperbolic weight $w_s$ adds anything beyond a real tree once the sequences are fixed.
Published contact-prediction reference points. We report published numbers against existing contact predictors. On CAMEO-L, Contact-Distil reports a 13.3% improvement over AlphaFold2, reaching 37.4% for single sequences, both relative figures. No absolute CAMEO-L precision is reported on Lacuna, so we will rerun its released protocol. MSA Transformer reports 57.4% Top-L long-range precision on trRosetta, against 41.1% for ESM-1b. The Co-evolution Transformer reports 51.6% on CASP14 free-modeling targets.
Why not just…?
Why not just reweight the native alignment by sequence identity? Reweighting corrects redundancy among sequences you already have but cannot add lineages you lack. With under 30 sequences from one clade, reweighting shrinks $M_{\text{eff}}$ toward one and exposes the problem without solving it. The native-alignment baseline measures exactly this ceiling.
Why not just use Euclidean retrieval plus diversity reranking? This is the real threat. Clustering can remove near-duplicates in any metric. What it cannot easily do is distinguish "different" from "independent": two sequences can be far apart in Euclidean space yet sit in the same deep subtree, because peripheral leaves crowd together in flat space (Poincaré Embeddings). If the reranking baseline and the curvature ablation match our method on CAMEO-L, we will conclude that diversity-aware retrieval, not hyperbolic geometry, is the contribution, and report it that way.
Why not just build a proper tree with FastCherries? For a fixed set of sequences, a real tree may weight as well as $w_s$, and the oracle baseline tests this. A tree cannot drive retrieval over 62M sequences without first retrieving them, which is precisely where our index acts. If the oracle matches $w_s$, we keep hyperbolic retrieval and drop hyperbolic weighting.
Why not just skip alignments and use a language model or distillation? MSA Transformer's 57.4% versus ESM-1b's 41.1% shows that explicit evolutionary input still matters (MSA Transformer). Contact-Distil is complementary rather than competing, since its student can consume our retrieved alignments. We therefore test it both with native inputs and with our retrieved inputs.
Why not just use simplicial message passing on the contact map? Simplicial networks such as Message Passing Simplicial Networks add expressive power but also clique-lifting cost. Pair-level refiners, such as the 2D convolutions in the Co-evolution Transformer, already see triangles through sequence-ordered pair maps. Our refiner ablation compares a node-level GIN refiner, a pair-level refiner and a simplicial refiner on the same couplings, which settles whether higher-order lifting is worth its cost here.
Proposed Experiments
Synthetic check. The first decisive check has known answers. We sample sequences from Potts models with adabmDCA along simulated trees of controlled clade imbalance, so both the true couplings and the true phylogeny are known. At $N \in \{8, 16, 32\}$, we measure coupling-recovery precision for all weighting schemes, including an oracle that uses the true tree. The prediction is that hyperbolic weighting approaches the oracle, identity weighting degrades as clade imbalance grows, and the Euclidean curvature ablation falls in between. Index quality is reported separately as tree-distance distortion on held-out Pfam families, following the Lorentz model paper.
Real low-homology benchmark. The second check is CAMEO-L (339 proteins with 10 or fewer native homologs), with trRosetta and CASP14 free-modeling targets as secondary sets. All comparisons are at matched $k$, using Top-L and Top-L/5 long-range precision. We stratify results by the clade concentration of each target's native alignment. The claim predicts gains concentrated in clade-dominated targets. Uniform gains would instead suggest the method is simply retrieving more or better sequences, which the Euclidean baselines should also capture.
Refiner check. The third check is the pair-versus-node refiner comparison described above. On synthetic data, we score how well each refiner removes injected clade cliques while keeping true contacts.
Transfer check. As a transfer check, we feed hyperbolically retrieved sets to the PoET reader on ProteinGym and compare against Protriever's 0.479 Spearman. This check is secondary and would not by itself support or refute the contact claim.
References
- Using evolutionary couplings to predict contacts and build structures
- Fast and Accurate Multivariate Gaussian Modeling of Protein Families: Predicting Residue Contacts and Protein-Interaction Partners
- Protein 3D Structure Computed from Evolutionary Sequence Variation
- MSA Transformer
- Contact-Distil: Boosting Low Homologous Protein Contact Map Prediction by Self-Supervised Distillation
- Co-evolution Transformer for Protein Contact Prediction
- Protriever: End-to-End Differentiable Protein Homology Search for Fitness Prediction
- Ultrafast classical phylogenetic method beats large protein language models on variant effect prediction
- Poincaré Embeddings for Learning Hierarchical Representations
- Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry
- Hyperbolic Neural Networks
- adabmDCA 2.0 - a flexible but easy-to-use package for Direct Coupling Analysis
- How Powerful are Graph Neural Networks?
- Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks