OmniScore: Universal Scoring of Diverse Biomolecular Complexes via Equivariant Geometry-Aware Discrete Representation Learning
Scoring biomolecular complexes is central to structure assessment and drug discovery, yet the complexes themselves vary widely in pose, size, and molecular composition. A scoring function tuned for one interaction type rarely carries over to another, and most existing methods compound the problem by leaning heavily on task-specific labels. We introduce OmniScore, a universal structure-based…
Scoring the intricate arrangements of biomolecular complexes is crucial for assessing their structure and aiding drug discovery. However, the complexity of these complexes, which vary in pose, size, and molecular composition, makes it challenging for scoring functions to be universally applicable. Most existing methods exacerbate this issue by relying heavily on task-specific labels, which limits their universal utility.
Introducing OmniScore, a pioneering universal structure-based framework designed to learn a shared geometry-aware representation of these diverse complexes just once and then adapt it seamlessly to various downstream scoring tasks through lightweight heads. OmniScore ingeniously combines a graph view and a sequence view of each structure, encodes their three-dimensional geometry, and compresses these representations into a compact latent space. This compact space can be readily reused by both a reconstruction module and prediction heads.
The backbone of OmniScore is pre-trained on an extensive array of diverse datasets, encompassing complexes, monomers, and small molecules. This pre-training is guided by complementary objectives, such as coordinate recovery, repair of corrupted input tokens, predicting molecular identity, and anchoring the representation in structure-level physical quantities.
The results demonstrate OmniScore's superior performance across several benchmark evaluations. In antibody-antigen and nanobody-antigen quality assessments, OmniScore outperformed all reported metrics compared to state-of-the-art baselines. Remarkably, its frozen residue embeddings matched the performance of the state-of-the-art protein-tokenization method, achieving an average functional-site accuracy of 71.8% on a standard residue-level benchmark.
Moreover, OmniScore's performance on protein-ligand scoring and ranking benchmarks was on par with methods specifically designed for those tasks.
These findings suggest that geometry-aware pretraining, as implemented in OmniScore, can provide a reusable scoring backbone for tasks that hinge on interfacial and residue-level structure, across the evaluated settings.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
