Humanization of antibodies using a machine learning approach on large-scale repertoire data

Mark Chin, Claire Marks, Charlotte M. DeanebioRxiv 2021 · DOI

TL;DR

This paper introduces Hu-mAb, a novel machine learning tool that significantly improves antibody humanization by predicting and suggesting mutations to reduce immunogenicity. By leveraging a massive dataset of antibody sequences and advanced Random Forest classifiers, Hu-mAb offers a faster, more systematic alternative to traditional trial-and-error methods, maintaining therapeutic efficacy.

What Problem Does This Paper Address?

Monoclonal antibodies (mAbs) are a cornerstone of modern medicine, treating a wide range of diseases. However, many therapeutic mAbs are initially derived from non-human sources, like mice. These "foreign" antibodies can trigger an immune response in humans, known as immunogenicity. This response can reduce the drug's effectiveness and even cause adverse side effects. To overcome this, researchers "humanize" these antibodies by replacing parts of the non-human sequence with human counterparts. Historically, this process has been a laborious, iterative cycle of experimentation, often involving guesswork to restore lost efficacy after humanization. This paper tackles the challenge of making antibody humanization more efficient, predictable, and less reliant on costly, time-consuming experimental approaches.

Methods

The core of this work lies in developing sophisticated machine learning classifiers to predict the "humanness" of antibody sequences. The researchers trained Random Forest (RF) models on a vast dataset of antibody sequences. These RF classifiers were designed to distinguish between human and non-human antibody variable domains (specifically, the VH, VL kappa, and VL lambda chains). Unlike some previous methods that focused solely on identifying human sequences, these RF models were trained on both positive (human) and negative (non-human) examples, leading to more robust discrimination.

These trained classifiers were then integrated into a computational tool called Hu-mAb. Hu-mAb takes an antibody sequence as input and systematically suggests mutations within the framework regions (while preserving the critical complementarity-determining regions, or CDRs, responsible for antigen binding) to increase its "humanness" score. The tool operates greedily, selecting mutations that yield the greatest increase in humanness score, aiming to minimize the total number of changes and thus preserve therapeutic efficacy.

Datasets

The primary dataset used was the Observed Antibody Space (OAS) database, containing over 500 million antibody sequences. This massive dataset, collected from diverse immune states and organisms, provided a rich source for training the RF models. For evaluation and comparison, the researchers utilized a benchmark dataset of 481 therapeutic antibodies from the Thera-SAbDab database, categorized by their origin (human, humanized, chimeric, mouse). Additionally, a subset of 25 experimentally humanized therapeutic sequences with known precursor sequences was used to directly compare Hu-mAb's suggested mutations against those determined experimentally.

Key Insights and Ideas

  • Large-scale repertoire data is key: The sheer volume of available antibody sequence data from OAS was crucial for building highly accurate classifiers.
  • RF outperforms LSTM: The study found that their RF models consistently outperformed previous state-of-the-art LSTM models in discriminating human from non-human sequences, likely due to the RF's ability to learn from both positive and negative examples.
  • Humanness correlates with immunogenicity: A strong negative correlation was observed between the "humanness" score predicted by the models and the experimental immunogenicity of antibody therapeutics. Higher humanness scores are associated with lower immunogenicity.
  • Framework regions are critical for humanness: Feature importance analysis revealed that residues within the framework regions, particularly framework region 4, play a more significant role in determining an antibody's humanness than CDR residues.
  • Hu-mAb mimics experimental strategies: The mutations suggested by Hu-mAb showed a high degree of overlap with mutations made in experimental humanization studies, indicating that the tool captures the underlying principles of effective humanization.

Key Results

  • High Classification Accuracy: The RF models achieved near-perfect classification performance, with AUC scores close to 1 and YJS scores ≥0.999 for distinguishing human from non-human sequences.
  • Reduced Immunogenicity: Therapeutics with high humanness scores (above the YJS threshold) exhibited significantly lower observed immunogenicity, with over 90% showing low levels of anti-drug antibodies (ADAs).
  • Efficient Humanization: Hu-mAb suggested, on average, 59% of the mutations made experimentally for VH and VL sequences.
  • High Mutation Overlap: An average adjusted overlap ratio (AOR) of 77% for VH and 85% for VL was achieved, meaning a large proportion of Hu-mAb's suggested mutations were also found in experimentally humanized antibodies.
  • Faster Process: Hu-mAb offers a systematic and efficient approach, achieving similar results to experimental methods in a fraction of the time.

Limitations and Future Work

  • Species Bias: The models are primarily trained on murine sequences and may perform less accurately on non-human sequences from other species not well-represented in the OAS database (e.g., rabbit sequences in the case of Clazakizumab). The authors acknowledge that Hu-mAb is mainly intended for murine precursor sequences.
  • Data Limitations for Immunogenicity Correlation: The correlation between model scores and immunogenicity was found to be weak (R² of 0.35), partly due to limitations in the available immunogenicity data (e.g., differences in patient demographics, dosage, and co-treatments).
  • Future Work: The authors plan to continuously train and update the RF models as new non-human species data becomes available in OAS, potentially expanding the tool's applicability. They also highlight the importance of the framework regions for humanness and the interactions between residues, suggesting that future models could further refine these aspects.

Summary updated

Abstract

Monoclonal antibody therapeutics are often produced from non-human sources (typically murine), and can therefore generate immunogenic responses in humans. Humanization procedures aim to produce antibody therapeutics that do not elicit an immune response and are safe for human use, without impacting efficacy. Humanization is normally carried out in a largely trial-and-error experimental process. We have built machine learning classifiers that can discriminate between human and non-human antibody variable domain sequences using the large amount of repertoire data now available. Our classifiers consistently outperform existing best-in-class models, and our output scores exhibit a negative relationship with the experimental immunogenicity of existing antibody therapeutics. We used our classifiers to develop a novel, computational humanization tool, Hu-mAb, that suggests mutations to an input sequence to reduce its immunogenicity. For a set of existing therapeutics with known precursor sequences, the mutations suggested by Hu-mAb show significant overlap with those deduced experimentally. Hu-mAb is therefore an effective replacement for trial- and-error humanization experiments, producing similar results in a fraction of the time. Hu-mAb is freely available to use at opig.stats.ox.ac.uk/webapps/humab.
BibTeX
@misc{chin2021humanization,
  author = {Mark Chin and Claire Marks and Charlotte M. Deane},
  title = {{Humanization of antibodies using a machine learning approach on large-scale repertoire data}},
  year = {2021},
  doi = {10.1101/2021.01.08.425894},
  howpublished = {bioRxiv},
}
Figure