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

Abstract

Motivation 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. Results Our classifiers consistently outperform the current best-in-class model for distinguishing human from murine sequences, 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 therapeutic antibodies 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. Availability Hu-mAb (humanness scoring and humanization) is freely available to use at opig.stats.ox.ac.uk/webapps/humab. Supplementary information Supplementary data are available at Bioinformatics online.

BioPhi: A platform for antibody design, humanization and humanness evaluation based on natural antibody repertoires and deep learning

“…Nevertheless, antibody germlines provide orders of magnitude smaller sequence space than antibody repertoires, making them more restrictive for antibody engineering applications. Although Hu-mAb [18] has outperformed all other methods including ours by a narrow margin on humanness classification (97.7% and 96.6% ROC AUC respectively for Hu-mAb and OASis medium identity) and on immunogenicity prediction (0.34 and 0.28 R 2 respectively), we identified two drawbacks. Firstly, Hu-mAb produces only a single score per chain.…”

“…Therefore, we devised a separate humanization mechanism that can capture humanness including long-range dependencies. Moreover, such separation of the humanization method from the humanness score ensures independent humanness evaluation, in contrast to optimizing and evaluating using the same score as employed by previous approaches [11][13] [18].…”

“…Ultimately, the goal of humanness evaluation is to capture and reduce the immunogenicity risk. Thus, we evaluated the ability of OASis and other humanness metrics to predict reported anti-drug antibody response of 217 therapeutics curated in [18]. We measured performance using Pearson correlation coefficient (R) and explained variance (R 2 ).…”

“…Correlation with clinical immunogenicity was evaluated on a dataset curated in Hu-mAb study [18], sequences were obtained from IMGT mAb DB [35]. Sequence for catumaxomab was not available, therefore only 217/218 therapeutics were included.…”

“…Both methods were released in a new open-source package AbNumber 6 . Humanization methods evaluation Validation set of 25 humanized sequences paired with their known parental sequences and Hu-mAb predictions was acquired from the Hu-mAb study [18]. Hu-mAb predictions from the 152 putative parental sequences were generated using the Hu-mAb web server (March 2021).…”