Humanness evaluation and humanization of antibodies are well established, ranging from CDR grafting to deep learning methods. Here we present our contributions: OASis, a granular antibody humanness score, and Sapiens, a language model for antibody humanization, integrated into the open-source BioPhi web platform. Finally, we explore whether similar principles can be applied to de novo designed proteins, highlighting challenges stemming from computational complexity and fundamental biological constraints.

Epitope presentation by MHC Class II (pMHC) is a necessary condition for an immunogenic response to antibody therapeutics. Thus, high-performance pMHC models are a cornerstone for immunogenicity risk assessment. We have developed Pep2Vec, a modular, interpretable, pMHC model that achieves state-of-the-art performance on a variety of presentation and immunogenicity datasets. We will show how to use Pep2Vec to distinguish between high and low immunogenicity (human and humanized) antibody drugs.

Using a library-based method, HLA-II allele / epitope peptide interactions are measured at scale and epitope landscapes are mapped to protein sequences. To engineer reduced immunogenicity, thousands of single substitutions were tested to comprehensively identify transitional mutations that switch an epitope from HLA-II binding to non-binding. While existing AI/ML algorithms are poor predicters of transitional epitopes, these data are expected to improve ML models for engineering proteins with reduced immunogenicity.

Inverse folding (IF) and protein large language models (pLLMs) have become useful tools for antibody variant generation, with generally good performance, but limited ability to find mutations that enhance the binding to the antigen. Here, we show how IF models can be used to predict B cell epitopes, how to extend this approach to estimate antibody-antigen interaction energy and find mutations that increase affinity, and to fine tune this model to increase its predictive power.

During early-stage discovery of antibody therapeutics it is important to consider not only target affinity but also developability attributes before selecting a lead candidate for further development. To facilitate this we have developed two tools ImmunoScreen and the InSiDe (in silico developability) pipeline for biologics discovery. These tools provide the ability to assess immunogenicity along with other interdependent risks such as aggregation and post-translation modification to guide lead selection.

Accurate in silico immunogenicity predictions and effective deimmunization strategies have the potential to significantly enhance the clinical success of therapeutic large molecules. This analysis presents an integrated in silico examination of the immunogenicity of two bispecific antibodies, alongside clinical data on the specificity of their respective anti-drug antibody responses. It offers insights into how these tools can be effectively utilized within drug discovery pipelines.

• Best practices in data preparation for machine learning of peptidomics datasets
• Novel deep learning approaches for predicting MHC antigen presentation and the modeling challenges
• Interpretability and explainability of the available deep learning models
• New screening strategies for predicting immunogenic hotspots in therapeutic antibodies

• Enhance AI/ML model performance for immunogenicity prediction by optimizing HLA-II selection and tolerance determination.
• What confidence level is needed for immunogenicity prediction algorithms to impact regulatory decisions
• How can in silico and in vitro approaches for immunogenicity evaluation be strategically integrated, especially when their data conflict?
• Developing AI/ML algorithms for B cell prediction - what does the future hold?
• Application of the in silico prediction tools for the immunogenicity of AAV​.

Our group has been at the forefront of studying adaptive immune responses associated with vaccination and microbial outbreaks of new and old pathogens. Examples include SARS, pertussis, Mpox, and avian influenza. Our work is also currently directed at preparation for potential new pandemics by combining computational and experimental methods to predict immune responses to viral and bacterial families of cancer. The technology developed in these contexts is of general applicability.

Novel modalities such as CAR T and gene therapy have progressed in to the clinic and have shown remarkable efficacy. Just like any other biologic, these therapies have the potential to be recognized by the human immune system, resulting in various immune responses that can compromise patient safety and the efficacy of the therapeutic. Here we will discuss the particular challenges associated with immunogenicity and these modalities and methods and strategies to assess and minimize potential immunogenic elements. Immunogenicity can be an obstacle to efficacy for novel modalities. Cellular immunogenicity poses a challenging immunogenicity risk to mitigate and measure.

Biotherapeutics activate the immune system by the interactions with antigen presenting cells (APC), T cells and B cells. Identifying the immune hotspot is a critical step to deimmunize the drug candidates. In this presentation, we will introduce the factors that impact the immune response pathways, summarize the in silico prediction tools in immunogenicity risk assessment, and discuss multiple approaches to predict T cell and B cell epitope in drug candidate screening.

I will cover our work on machine learning methods for antibody and protein design from both an academic and biotech perspective. Notably I will cover our methods for de novo design as well as for optimization. I will cover some use specific use cases, including optimizing developability characteristics using ML methods.

In silico prediction of CD4+ T cell epitopes frequently results in numerous candidates, yet ranking these epitopes by their potential to activate naive CD4+ T cells and induce anti-drug antibodies remains challenging. Our research suggests that epitopes resembling bacterial sequences increase immunogenicity risk in antibody therapeutics. By integrating molecular mimicry characteristics into predictive models, we can better identify immunogenic hotspots, thereby enhancing the safety and efficacy of biotherapeutic development.

The safety and efficacy of therapeutic proteins are undermined by immunogenicity driven by anti-drug antibodies (ADA). Proteins administered subcutaneously can suffer from enhanced immunogenic potential compared to intravenous administration. The talk will cover mechanistic insight into the subcutaneous immune response and our efforts to develop novel preclinical tools as well as a database to predict clinical immunogenicity as a human biology-based animal trial alternative.

Immunogenicity potential assessment should be started as early as possible in the biotherapeutic development process to inform de-immunization approaches and to avoid resources spending on candidates with a high inherent immunogenicity potential. Oftentimes, this is only possible using in silico tools. In my presentation, I would like to introduce the audience to the in silico-based workflow we implemented to analyze and decrease the immunogenicity potential of biotherapeutics in early development.