TL;DR
This paper introduces "Fusion," a novel computational tool that accurately assembles full-length antibody sequences from mass spectrometry data. Fusion is the first method capable of simultaneously sequencing multiple antibodies from a single sample, significantly increasing throughput and reducing costs for antibody research and development.
What Problem Does This Paper Address?
Understanding the precise amino acid sequence of antibodies is critical for their function, from antigen binding to therapeutic applications. However, determining these sequences, especially de novo (without a known reference), is a significant challenge. Current methods struggle with accuracy, robustness, and, crucially, throughput. Assembling sequences from a mixture of antibodies is particularly difficult, limiting the ability to analyze multiple antibodies simultaneously. This bottleneck hinders large-scale antibody research and the development of biosimilars.
Methods
Fusion employs a four-stage approach:
- Peptide Preprocess: Filters out low-confidence peptide fragments and amino acids from the initial de novo sequencing output to improve data quality.
- Template Search: Matches the processed peptide reads against germline antibody gene sequences (from the IMGT database) to identify potential templates for assembly. This stage uses a mass-based alignment algorithm that accounts for potential amino acid mass ambiguities.
- Split into K-mers: Breaks down the peptides into overlapping short sequences (k-mers, specifically 7-mers) to build a de Bruijn graph. This stage also incorporates a novel weighting scheme for k-mers based on their confidence scores and matches to template sequences.
- Assembly: Constructs a de Bruijn graph and uses a beam search strategy (with a beam width of 20) to extend contigs and identify the most likely complete antibody sequences. Fusion incorporates a sliding window and Levenshtein distance to differentiate between sequences in mixtures.
Datasets
The study utilized several datasets for evaluation:
- Human COVID-19 neutralizing antibody datasets: BD5514LH, S2P6LH (at 50, 100, and 200 µg concentrations).
- Murine COVID-19 neutralizing antibody datasets: 85F7, 36H6, and 2B4.
- Commercial monoclonal antibody: AFS98 (with unknown sequence).
- Mixtures: Combinations of two or three of the murine antibodies (85F7, 36H6, 2B4) were also used.
The paper also mentions using the Casanovo model, which was pre-trained on the MassIVE KB v1 dataset.
Key Insights and Ideas
- Integrating Template Information is Key: Fusion's success hinges on its ability to leverage both de novo peptide data and germline templates. This hybrid approach overcomes limitations of purely de novo or template-based methods.
- Beam Search for Robust Assembly: The beam search strategy allows Fusion to explore multiple assembly paths simultaneously, increasing the chances of finding the correct sequence, especially in complex mixtures.
- High-Throughput for Antibody Mixtures: The most significant breakthrough is Fusion's capability to assemble sequences from mixtures of antibodies. This dramatically improves efficiency compared to sequencing each antibody individually.
- Biological Validation Matters: The successful in vivo validation of the AFS98 antibody, whose sequence was determined by Fusion, provides strong evidence for the algorithm's accuracy and practical utility.
Key Results
- High Accuracy: Fusion achieved 100% accurate coverage for human antibodies and over 99% for murine antibodies, outperforming existing methods like Stitch and PEAKS AB.
- Robustness: Fusion demonstrated consistent performance across different sample quantities and batches, unlike Stitch, which showed variability.
- High Throughput: Fusion is the first method to enable high-throughput sequencing of multiple antibodies from a single bottom-up mass spectrometry experiment.
- AFS98 Antibody: Fusion successfully assembled the sequence of the AFS98 antibody, and subsequent biological validation showed its effectiveness in clearing macrophages.
- Mixture Assembly: Fusion accurately assembled sequences from mixtures of two and three monoclonal antibodies, achieving the same accuracy as single-antibody assemblies.
Limitations and Future Work
- Template-Independent Assembly: While Fusion significantly improves de novo assembly, the paper acknowledges that achieving complete de novo assembly without any templates, especially for novel organisms or highly divergent sequences, remains a significant challenge.
- Further Refinement: The authors aim to further refine the Fusion algorithm to enhance its utility for even more extensive antibody research and development initiatives.
- Adaptability: The paper suggests that Fusion could theoretically be adapted to work with de novo peptide sequencing software beyond Casanovo.
Summary updated