Despite the central role that antibodies play in modern medicine, there is currently no way to rationally design novel antibodies to bind a specific epitope on a target. I will discuss the development and experimental validation of RFantibody, a deep-learning pipeline capable of designing de novo antibodies that bind to user-specified epitopes.

The L1000 program systematically generated 1.3 million gene expression signatures in human cell lines with diverse genomic and pharmacological perturbations. Similar L1000 gene signatures offer an unbiased data-driven mechanism to identify compound-target pairs. Current methods rely on matching gene identities when comparing gene signatures and fail to utilize preexisting gene function knowledge. We developed FRoGS, an approach that represents gene signatures projected onto their biological functions, instead of their identities, which results in more effective compound-target predictions. FRoGS can help uncover new relationships among gene signatures acquired by large-scale OMICs studies on compounds, cell types, disease models, and patient cohorts.

Our discovery biology platform provides a deep understanding of the regulatory circuits underlying human tissue Treg behavior. A suite of disease-relevant phenotypic assays facilitates multi-dimensional functional interrogation of key genes in human Tregs. These complex in silico tools combined with translational in vitro assays identify novel regulatory nodes and form the foundation of our growing pipeline of a new class of tissue Treg-focused therapeutics for immune-mediated diseases.

This research integrates bioimaging, proteomics, and AI to study human cell biology to explore protein distribution in time and space, investigating how localization variations affect cell functions and disease. Our goal is to create a spatiotemporal proteome model of a human cell. Using ML, we interpret spatial data from image collections and combine it with other datatypes to build whole-proteome multi-scale cell models, with potential to enhance drug discovery processes.

Monoclonal antibodies are important biotherapeutics, but developing them is expensive and difficult since they must be specific and able to be produced at commercial-scale. We present a transformer-based antibody language model, trained on large amounts of antibody data. This model can be used to inform and de-risk early antibody discovery by increasing the quantity of viable antibodies across antibody discovery projects.

We develop and apply machine learning models to optimize in parallel multiple drug-like properties of the bacterial enzyme IdeS, to design a therapeutic for chronic autoantibody-mediated diseases, while minimizing its immunogenicity and other liabilities. The success of this approach is demonstrated via in vivo and in vitro assays, and we illustrate its generalizability by engineering non-immunogenic bacterial cysteine proteases with a variety of immunoglobulin isotype specificities.

Analytical and formulation data is foundational to CMCDevelopment. In this presentation we discuss AbbVie’s CMC digital journey towardsconnecting analytical and formulation data to their respective study and sampleID, to build a data-driven CMC organization where structured data brakes silos, empowers storytelling visualizations that enable comparison acrossprojects, reduce manual transcription to other systems and colleagues, andallow for trending and other business metrics.

• Addressing skills gaps
• Benchmarking progress compared with traditional structures
• Best practices for collaboration between experimentalists and data scientists
• Examples of successful transitions
• Implementing models/tools and workflows based on AI/ML approaches

We present RESP2, an enhanced version of our RESP pipeline, designed for the discovery of antibodies against diverse antigens with simultaneously optimized developability properties. We used the RBD of the COVID-19 spike protein as a case study, and discover a highly human antibody with broad binding to different variants, which demonstrated the power of this pipeline for antibody discovery against a challenging target.

Accurate modeling of immune recognition remains a major challenge in computational biology. To provide insights into the performance of deep learning for modeling immune recognition, we performed detailed benchmarking of multiple AlphaFold2 and AlphaFold3 approaches for modeling antibody and T cell receptor complexes. This revealed approaches with higher performance, overall limitations in success, as well the utility of confidence scores in the selection of accurate models.

We have developed an NGS-guided workflow for multi-dimensional optimization of biologics such as antibodies and TCRs. The workflow involves the rational design of targeted CDR combinatorial diversity guided by Deep Mutational Scanning, followed by yeast library selections. We will present antibody and TCR optimization case studies demonstrating substantial gain in affinity that is competitive with standard approaches, good developability, and minimal variant mutational load.

In antibody discovery, repertoire sequencing unfolds a broad antigen-specific sequence space and informs of the potential mutational space. Integrating this information and advances in experimental techniques with different machine learning approaches, we are now able to efficiently explore and extend the sequence space to identify alternative binders and improve existing binders.

Targeting GPCRs with agonists is a significant challenge in biologics design. Our approach employs molecular mimicry to optimize both developability and functional outcomes, including biased agonism. Leveraging machine learning algorithms backed by extensive empirical data acquisition and protein engineering, we thoroughly explore the pharmacophore space within live cells. This enables us to create designer molecules that address complex biological processes, ultimately leading to the development of novel GLP1R modulators.