Abstract
In the field of antibody engineering, an essential task is to design a novel antibody whose paratopes bind to a specific antigen with correct epitopes. Understanding antibody structure and its paratope can facilitate a mechanistic understanding of its function. Therefore, antibody structure prediction from its sequence alone has always been a highly valuable problem for de novo antibody design. AlphaFold2, a breakthrough in the field of structural biology, provides a solution to predict protein structure based on protein sequences and computationally expensive coevolutionary multiple sequence alignments (MSAs). However, the computational efficiency and undesirable prediction accuracy of antibodies, especially on the complementarity-determining regions (CDRs) of antibodies limit their applications in the industrially high-throughput drug design. To learn an informative representation of antibodies, we employed a deep antibody language model (ALM) on curated sequences from the observed antibody space database via a transformer model. We also developed a novel model named xTrimoABFold to predict antibody structure from antibody sequence based on the pretrained ALM as well as efficient evoformers and structural modules. The model was trained end-to-end on the antibody structures in PDB by minimizing the ensemble loss of domain-specific focal loss on CDR and the frame-aligned point loss. xTrimoABFold outperforms AlphaFold2 and other protein language model based SOTAs, e.g., OmegaFold, HelixFold-Single, and IgFold with a large significant margin (30+% improvement on RMSD) while performing 151 times faster than AlphaFold2. To the best of our knowledge, xTrimoABFold achieved state-of-the-art antibody structure prediction. Its improvement in both accuracy and efficiency makes it a valuable tool for de novo antibody design and could make further improvements in immuno-theory.
Introduction
In the world of drug discovery, antibodies are the "guided missiles" of the immune system. To design a new antibody, scientists need to know its 3D structure—specifically the complementarity-determining regions (CDRs), the loops that actually grab onto a target (antigen). While AlphaFold2 (AF2) revolutionized protein folding, it has two major drawbacks for antibody engineering. First, the "MSA Bottleneck": AF2 relies on Multiple Sequence Alignments (MSAs)—searching through massive databases to find similar sequences to see how they co-evolved. This is slow and computationally expensive. Second, the "Antibody Exception": Antibodies don't evolve like "normal" proteins; they undergo rapid, targeted mutations to fight specific threats. Consequently, MSAs for antibodies are often sparse or unreliable, leading to poor structural predictions in the critical CDR loops.
Methods
xTrimoABFold is an end-to-end folding pipeline that replaces the MSA search with a more efficient starting point. Its architecture consists of three main pillars:
- Antibody Language Model (ALM): Instead of a general Protein Language Model (PLM), the team pretrained a transformer specifically on antibody sequences from the Observed Antibody Space (OAS). This model learns the specific "grammar" of antibodies, allowing it to generate rich initial representations of how residues interact.
- Cross-Modal Template Searching: To provide a "warm start," the authors developed a fast algorithm that searches for similar known structures using both sequence similarity and coarse-grained structural priors. This provides structural context in milliseconds rather than the minutes or hours required for MSA generation.
- CDR Focal Loss: Since CDR loops are the hardest parts to predict but the most important for function, the authors implemented a domain-specific "focal loss" during fine-tuning. This forces the neural network to pay extra attention to the coordinates of these difficult regions during training.
The core folding engine uses the Evoformer and Structure Module architectures pioneered by AlphaFold2, but optimized to work with single-sequence inputs and templates rather than MSAs. The model was trained on a curated Antibody Data Bank (ADB) containing 501K protein structures and a specialized subset of 19K high-resolution antibody structures.
Key Results
xTrimoABFold sets a new state-of-the-art (SOTA) for antibody structure prediction across several metrics:
- Accuracy: It achieved an average RMSD of 1.96 Å, significantly outperforming AlphaFold2 (3.13 Å) and antibody-specific models like IgFold. This represents a 37% improvement over AF2.
- Speed: Because it avoids the MSA search, it is 151 times faster than AlphaFold2, making it suitable for high-throughput drug design.
- CDR3 Performance: The model showed particular strength in predicting the CDR3 loop, which is notoriously variable and difficult to model, achieving significantly lower error rates than all baselines.
The researchers found that specialization trumps scale; a smaller language model trained specifically on antibodies (ALM) outperformed massive general-purpose protein models. Additionally, by using attention weights from the pretrained ALM to initialize the "pair representation," the model effectively "hallucinates" the co-evolutionary information that MSAs usually provide.
Conclusion
xTrimoABFold represents a significant leap in antibody modeling by combining domain-specific language models with efficient structural modules. While currently specialized for antibody monomer structures, its success demonstrates that MSA-free approaches can outperform traditional methods in both speed and accuracy for specialized protein classes. Future work aims to extend this approach to general protein structure prediction and complex antibody-antigen docking interactions.
Summary updated