Human-likeness of antibody biologics determined by back-translation and comparison with large antibody variable gene repertoires

Abstract

The antibody (Ab) germline gene rearrangement of variable (V), diversity (D), and joining (J) gene segments, as well as somatic hypermutation, give rise to the human Ab variable gene sequence repertoire. It is common to characterize single nucleotide frequencies of the variable region by alignment to species-specific wildtype germline genes. The increasing application of next-generation sequencing to immune repertoire studies has led to the compilation of increasing large adaptive immunome receptor repertoire datasets. We have developed a method that maps the sequence of a target Ab onto an immunome dataset of 326 million human Ab sequences. For this purpose, we created a position-and gene-specific scoring matrix (PGSSM) and its corresponding antibody similarity score. We characterized our PGSSM score and found that it strongly correlated with the phylogenetic distance of 181,355 Ab sequences from GenBank across 20 species. The most likely human nucleotide back-translation was obtained given only PGSSMs and the amino acid sequence of an Ab achieving a nucleotide sequence recovery of 95.9% and 97.2% for human heavy and light chains, respectively. In conclusion, the scoring of our back-translation is a valuable estimate for the similarity of an Ab sequence to the natural human repertoire. As expected, Ab therapeutic molecules developed from a human source showed a higher similarity to the repertoire than engineered Abs. Thus, the PGSSM metric introduced here can be used to engineer human-like Ab therapeutics.

Current advances in biopharmaceutical informatics: guidelines, impact and challenges in the computational developability assessment of antibody therapeutics

“…developed a computational method that maps the sequence of a given antibody onto human B-cell repertoires comprising 326 million sequences of human antibodies. 141 Chin et al . built a machine learning-based predictive model that distinguishes human antibody sequences from non-human ones, which was trained on large-scale repertoire dataset.…”

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

“…Existing methods based on homology [9][10] are interpretable but lack granularity since they only provide a single score for each chain. Moreover, these underperformed in our analysis compared to more recent approaches such as IgReconstruct [17], which could be attributed to the modest size of their reference sequence libraries. Interestingly, Germline content, a baseline method we implemented based only on percent sequence identity to nearest human germline performed comparably to recent methods while providing both an interpretable and granular score.…”

“…Furthermore, it was shown that developability properties of natural human antibodies are comparable with those of clinical mAbs [16]. In IgReconstruct [17], the OAS database was used to construct a back-translation method producing human-like DNA and evaluating humanness based on positional nucleotide frequency. Most recently, OAS was used in Hu-mAb [18] to train a random-forestbased humanness score used as an optimization criterion for iterative humanization.…”

Machine-designed biotherapeutics: opportunities, feasibility and advantages of deep learning in computational antibody discovery

“…The availability of NGS increased the antibody sequence samples from the order of magnitude thousands to millions. Improved positional frequencies were created based on such data [ 103 , 104 ].…”