Unified Protein Embedding Model

Comparing protein three-dimensional structures at scale is a persistent bottleneck in structural biology. Established metrics such as the template modeling score (TM-score) and the local distance difference test (lDDT) require explicit structural superposition or per-residue distance calculations, which become prohibitively expensive when screening the millions of predicted structures now produced by tools like AlphaFold. This model addresses that gap by learning protein embeddings whose vector distances directly approximate these structural similarity scores, allowing comparisons to be made in embedding space without alignment.

Developed by Jerry Xu, Shaojun Pei, and Gil Alterovitz at Harvard Medical School and released as a bioRxiv preprint in October 2025, the model couples a protein language model with a Siamese neural network architecture. Rather than predicting a single similarity value, it is trained with a dual objective that captures both global fold-level agreement, via TM-score, and fine-grained per-residue accuracy, via lDDT. This combination lets a single embedding space serve two complementary purposes: rapidly flagging proteins with globally different folds while remaining sensitive to subtle, localized structural changes.

#Key Features

• Dual-objective training: A combined loss function jointly regresses TM-score and lDDT, so a single embedding captures both whole-chain fold similarity and residue-level structural detail.
• Siamese architecture: Paired protein inputs are encoded by weight-shared network towers, producing embeddings whose distances reflect structural similarity between the two proteins.
• Alignment-free comparison: Structural similarity is estimated directly from embedding distances, avoiding the superposition and per-residue distance computations that make classical metrics slow at scale.
• Dual sensitivity: The model reliably detects large global differences in fold while remaining responsive to small, localized conformational or structural changes.

#Technical Details

The system pairs protein language model representations with a Siamese neural network that maps each protein to a fixed-length embedding. Training minimizes a composite loss combining the mean squared error against reference TM-score and lDDT values, encouraging the embedding geometry to encode both global and local structural relationships simultaneously. Evaluated on two held-out datasets, the model achieved TM-score mean absolute errors of 0.0741 and 0.0583, and lDDT mean absolute errors of 0.0788 and 0.0038, indicating that embedding distances track the reference structural metrics closely across both global and local regimes.

#Applications

The model is suited to any workflow that requires comparing large numbers of protein structures quickly. Use cases include structure-based similarity search across predicted-structure databases, clustering of conformational states, and screening for proteins that share a global fold while differing in localized regions. Because comparisons reduce to distance calculations in embedding space, the approach scales to large collections where running pairwise TM-align or lDDT would be computationally impractical, benefiting structural bioinformaticians and protein engineers triaging predicted models.

#Impact

By encoding two widely used structural metrics into a shared embedding space, this work connects sequence-derived protein representations with quantitative measures of structural similarity, offering a fast alternative to alignment-based comparison. The dual TM-score and lDDT objective is notable for preserving sensitivity to both coarse fold changes and fine local differences within one model, a combination often handled by separate tools. The method is described in a bioRxiv preprint and has also appeared in Briefings in Bioinformatics. As a comparison and retrieval tool it depends on the quality of the underlying reference structures and metrics, and its accuracy is characterized on the two evaluation datasets reported by the authors.