Predictability of antigen binding based on short motifs in the antibody CDRH3

Lonneke Scheffer, Eric Emanuel Reber, Brij Bhushan Mehta, Milena Pavlovic, Maria Chernigovskaya, Eve Richardson, Rahmad Akbar, Fridtjof Lund-Johansen, Victor Greiff, Ingrid Hobæk Haff, Geir Kjetil SandveBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces a novel method to identify short, position-specific amino acid motifs within antibody CDRH3 sequences that can accurately predict antigen binding. Their approach, validated on experimental and simulated data, demonstrates that these motifs can form the basis of a simple classifier that rivals or surpasses complex deep learning models, suggesting that antigen binding for some antibodies is largely dictated by these specific short sequences.

What Problem Does This Paper Address?

Adaptive immune receptors (AIRs), like antibodies and T-cell receptors, are crucial for recognizing and neutralizing foreign threats. A central challenge in immunology is understanding the rules that govern how these receptors bind to specific antigens. This knowledge is vital for developing new therapeutics and diagnostics. While it's widely accepted that the complementarity-determining region 3 (CDR3) of these receptors plays a key role in antigen recognition due to its high variability, the precise mechanisms by which CDR3 sequences dictate binding remain largely unknown. Many studies hypothesize that short "motifs" within the CDR3 are responsible for this specificity, but this has not been systematically investigated across diverse and potentially gapped sequences. This paper aims to rigorously test this hypothesis by developing a method to discover such motifs and assess their predictive power for antigen binding.

Methods

The core of this work is a novel method for discovering short, position-specific amino acid motifs within antibody CDRH3 sequences that are highly predictive of antigen binding. The method operates in several key steps:

  1. Dataset Splitting: The data is divided into training, validation, and test sets to ensure unbiased evaluation of motif performance.
  2. Recall Threshold Determination: For motifs that meet a high precision threshold on the training set, a recall threshold is determined. This step is crucial for filtering out rare motifs that might overfit to the training data and fail to generalize. They use a smoothed combined validation-precision curve to achieve this.
  3. Motif Discovery: A candidate set of motifs (up to 5 amino acids long, allowing for gaps) is generated using an efficient algorithm analogous to the Apriori algorithm for itemset mining. Motifs are filtered based on a precision threshold (e.g., >0.8 or >0.9) on the training set.
  4. Independent Evaluation: The final set of discovered motifs is evaluated on an independent test set to provide an unbiased estimate of their predictive performance.

A key innovation is the consideration of gapped and dispersed motifs, moving beyond the assumption of contiguous amino acid sequences. The method is integrated into the immuneML platform for reproducible analysis.

Datasets

The study utilizes two primary experimental datasets:

  • Mason et al. dataset: A combinatorial mutagenesis dataset of HER2 binder and non-binder antibody heavy chain sequences.
  • Mehta dataset: An independently generated experimental replicate of the Mason et al. dataset, also focusing on HER2 binders.

Both datasets are characterized by a fixed CDRH3 length and other variable regions, with variability primarily concentrated in 10 consecutive amino acid positions of the CDRH3.

Additionally, two simulated datasets were generated to validate the method:

  • One dataset with implanted ground truth motifs to test motif recovery.
  • Another dataset without any true binding motifs to assess the method's ability to avoid capturing noise.

Key Insights and Ideas

  • Short motifs can be highly predictive: The study strongly supports the hypothesis that short, specific motifs within the CDRH3 are indeed powerful predictors of antigen binding, even when other parts of the antibody sequence vary significantly.
  • Gaps and dispersion matter: Unlike previous approaches that focused on contiguous motifs, this work demonstrates the importance of considering dispersed motifs with multiple gaps. These gapped motifs appear to be more prevalent and predictive in the context of antibody-antigen binding.
  • Simplicity can outperform complexity: A straightforward motif-based classifier, using the discovered motifs as features, achieved performance comparable to or better than a sophisticated deep learning model (CNN) on independent test data. This suggests that the underlying "immune signal" for binding might be simpler than initially assumed by some complex models.
  • Motifs capture interaction effects: The learned motifs are not just based on the abundance of individual amino acids but rather on specific combinations of amino acids at particular positions, indicating the capture of main and interaction effects.

Key Results

  • The method successfully identified 11,336 and 792 motifs (for training precision thresholds of 0.8 and 0.9, respectively) from the Mason dataset.
  • These discovered motifs demonstrated high generalizability to the independently generated Mehta dataset, retaining strong predictive power.
  • A simple motif-based classifier, using a subset of just 178 motifs, outperformed a deep learning model (Mason CNN) on the independent Mehta test set in terms of accuracy and balanced accuracy.
  • The method was able to recover eight out of ten implanted ground truth motifs on simulated data, provided they were not too rare.
  • On simulated data without true binding motifs, the method correctly identified no generalizable motifs, demonstrating its robustness against noise.

Limitations and Future Work

  • Dataset specificity: The current study focuses on HER2 binding antibodies. Future work should explore how these findings generalize to other antigens and antibody types.
  • Motif definition: The authors acknowledge that exploring different definitions of "short motifs" (e.g., position-independent, flexible gap sizes, grouping amino acids by properties) could further enhance understanding.
  • Computational feasibility: Investigating a wider range of motif parameterizations may require more sophisticated algorithmic solutions due to computational limitations.
  • Data generation bias: The Mason dataset's design, while useful, introduced some skew in amino acid distributions. Further investigation into how data generation processes impact motif learning is warranted.
  • Broader application: The methodology is presented as a framework that can be extended to T-cell receptors and other adaptive immune receptor repertoires, especially as more large-scale datasets become available.

Summary updated

BibTeX
@article{scheffer2024predictability,
  author = {Lonneke Scheffer and Eric Emanuel Reber and Brij Bhushan Mehta and Milena Pavlovic and Maria Chernigovskaya and Eve Richardson and Rahmad Akbar and Fridtjof Lund-Johansen and Victor Greiff and Ingrid Hobæk Haff and Geir Kjetil Sandve},
  title = {{Predictability of antigen binding based on short motifs in the antibody CDRH3}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae537},
}

Loading related content...

Figure