OmniScore

A scoring function trained on protein-ligand poses does not tell you whether a predicted antibody-antigen interface is right. The two problems share a substrate — atoms in space — but little else: the labels differ, the interfaces differ in size and chemistry, and supervision is unevenly distributed across them. So the field has accumulated one scoring function per interaction type, each encoding its own assumptions about what a complex looks like.

OmniScore, from Arontier with Seoul National University and Hanoi University of Science and Technology, answers this by moving the shared work into pretraining. It learns a single geometry-aware representation of biomolecular structure once, then attaches small task-specific heads to it — the backbone stays frozen when a downstream scorer is fit. Each input structure is abstracted into blocks drawn from {residue, nucleotide, atom}, so a protein chain, an RNA strand and a ligand all enter the same encoder while macromolecules keep their biological units and small molecules keep atom-level granularity.

The design places it between the specialized scoring functions it competes with — BioScore and EquiScore are the nearest catalog neighbours — and structure tokenizers like AminoAseed. Unlike a scoring function tied to a predefined interaction potential, OmniScore's discrete tokens are not committed to a scoring objective, which is what lets the same frozen encoder serve residue-level prediction tasks it was never finetuned for.

#Key Features

• One backbone, frozen across tasks: The pretrained encoder and quantizer stay fixed during finetuning; only a lightweight head is optimized, so a new scoring task adds parameters rather than retraining the representation.
• Two coupled views of a structure: An equivariant graph transformer reads a block-level heterogeneous graph, and a bidirectional Mamba encoder reads the aligned atom-wise sequence, merging local geometry with long-range context.
• Discrete geometric tokens: Residual finite scalar quantization compresses continuous embeddings into a compact discrete latent space, from which a Mamba decoder reconstructs 3D coordinates.
• Physically grounded pretraining: Beyond coordinate recovery and masked-token repair, the backbone is trained to predict Rosetta energies, DockQ, binding affinity and per-residue solvent exposure, which stabilized training and improved downstream transfer.
• Transfer beyond scoring: Frozen residue embeddings, average-pooled from atomic tokens, drop directly into residue-level functional-site classification and property regression.

#Technical Details

The architecture is an autoencoder: equivariant graph transformer, then bidirectional Mamba encoder, then a residual finite scalar quantizer, then a Mamba decoder. The reported configuration uses three residual quantization layers with levels [5,5,3,3]; variants span 24M to 124M parameters across 512- and 1024-dimensional groups. Pretraining draws on PDBBind v2020 (15,758 training complexes), antibody/nanobody-antigen crystal structures plus Chai-1 and Boltz-2 regenerations from the authors' IgPose work, CATH (17,225), RNA3DB (10,094) and 100,000 NablaDFT small molecules, along with monomer and ligand subsets extracted from the complexes.

On antibody-antigen and nanobody-antigen quality assessment, OmniScore led all four reported metrics, its 111M variant reaching Pearson 0.909, Spearman 0.848, ROC-AUC 0.990 and PR-AUC 0.990. On StructTokenBench, the 84M variant averaged 71.76% AUROC over 15 functional-site classification settings — ahead of ESM-3 (69.24%) and FoldSeek (51.90%), and just behind AminoAseed (72.43%) — with its largest margins on BindShake (77.75% vs 69.61%) and CatBio-Fold (77.24% vs 65.95%). Physicochemical-property regression is weaker overall (32.50% average Spearman against AminoAseed's 38.08%), though OmniScore led all four flexibility settings. On CASF-16, scoring Pearson r sat at 0.633–0.637 and ranking Spearman at 0.593–0.623: above AutoDock Vina and RTMScore, below IGModel, GenScore and the PIGNet variants.

#Applications

The immediate use is triaging predicted complexes. Antibody discovery pipelines that generate many candidate Ig-antigen poses need a filter that ranks them by expected quality, and that is where OmniScore's results are strongest. The same checkpoint scores and ranks protein-ligand poses without swapping in a separate model, and its frozen residue embeddings serve as structural features for binding-site, catalytic-site and conservation prediction. Because the heads are small and the backbone is fixed, adding a scoring target means fitting an MLP or a mixture-density head rather than retraining.

#Impact

OmniScore's contribution is evidence that geometry-aware pretraining across heterogeneous structures yields a representation useful beyond the objective it was built for — the frozen embeddings compete with dedicated protein tokenizers on residue-level tasks despite never being trained for them. The authors are candid about the limits: protein-ligand scoring is on par with, not ahead of, single-task methods, and the conclusions are framed as holding within the evaluated benchmarks. Reconstruction remains poor for nucleic-acid-containing complexes and small-molecule ring geometry, the prediction heads are described as proof of concept, and the model outputs relative rather than calibrated absolute scores. A useful negative result sits alongside these: the checkpoints reconstructing best were not the ones transferring best, so RMSD misleads as a model-selection signal here. The preprint has not been peer reviewed, and no code or weights have been released.