ImmunoGeNN: Accelerating Early Immunogenicity Assessment for Generative Design of Biologics

Magnus Haraldson Høie, Birkir Reynisson, Paolo Marcatili, Jesper Ferkinghoff-Borg, Kasper Lamberth, Katharina L. Kopp, Morten Nielsen, Vanessa Isabell Jurtz

Keywords: peptide immunogenicity prediction, peptide MHC-II binding, model distillation, population immunogenicity assessment

TL;DR: Large scale immunogenicity screening and deimmunization of biological drugs

Abstract:

A critical step in the design of biological drugs is assessment of T cell immunogenicity risk, which may cause loss of therapeutic effect and costly, late-stage clinical trial failures. Computational tools such as NetMHCIIpan are widely used for assessing risk, but are computationally expensive, limiting common drug design tasks like screening large variant libraries and exhaustively screening for deimmunizing variants.

Here, we present ImmunoGeNN, a blazingly fast, distilled neural network predicting peptide-DRB1 risk scores in the North American population at up to 300,000 times the rate of NetMHCIIpan, while maintaining a >95% Spearman correlation in risk scores. ImmunoGeNN rapidly identifies dominant peptide binding cores and checks their presence in the human proteome, sharing a >99% agreement with NetMHCIIpan. Furthermore, it maintains NetMHCIIpan's performance in experimental validation against MHC-associated peptide proteomics (MAPPs) of a failed clinical phase II drug, vatreptacog alfa, successfully identifying newly introduced epitopes relative to its endogenous human protein counterpart.

With this speed-up ImmunoGeNN enables screening across millions of designs in a reasonable timeframe, removing a potential bottleneck in de novo design of biologics.

Availability: ImmunoGeNN's source code is freely available under the MIT license with no commercial restrictions. A web-server and downloadable package is made freely available for ease of use, at DTU Healthtech (https://services.healthtech.dtu.dk/services/ImmunoGeNN) and BioLib (https://biolib.com/DTU/ImmunoGeNN/)

Submission Number: 28