With a growing majority of its pipeline composed of biologics, there is an increasing need at Sanofi to bring more molecules to development at a faster pace. Generative AI and in silico screening methods provide opportunities to improve probability of success and decrease discovery-to-lead timelines. Combining deep repertoire mining technologies and generative ML modeling, we are building a de novo protein design platform and a more targeted drug discovery approach.

Dengue virus is a threat to global health. However, no specific therapeutics are available so far. Broadly neutralizing antibodies recognizing the various serotypes could serve as potential treatment. High-throughput adaptive immune receptor repertoire high-throughput sequencing (AIRR-seq) and bioinformatic analyses enable in-depth understanding of the B cell immune response. We investigated the dengue antibody response with these technologies and machine learning to identify rare and underrepresented broadly neutralizing antibody sequences. Dengue immunization elicited following signatures on the antibody repertoire: (i) an increase of CDR3 and germline gene diversity; (ii) a change in the architecture by eliciting power-law network distributions and CDR3 enrichment in polar amino acids; and (iii) an increase in the expression of JNK/Fos transcription factors and ribosomal proteins. Our work shows the applicability of computational methods and machine learning to AIRR-seq datasets for identification of potential neutralizing antibody candidate sequences. Further investigation of antibody expression and functional binding assays validated the obtained results.

Generative models have shown promise in capturing the distribution of natural proteins. In this talk, we'll cover some recent advances in large-scale machine learning models for functional protein design with focus on practical engineering tasks and experimental characterization.

Biologics require high specificity for targets, but current affinity-selection-based discovery methods do not guarantee this property. We present a method, computational counterselection, that identifies nonspecific candidates using machine learning models of affinity trained on high-throughput data from single-target affinity selection experiments.

While generators by deep leaning are bringing about revolutions in various fields of science, technology, and business, there are also warnings about the dangers of such generators. Generation of antibody sequences is also occurring in various generative models, but to what extent is the humanness of these models guaranteed? Evaluation of humanness has also been described as improved by deep learning, but they are all sequence-based humanness scores, and there is a common limitation in the presence of V-D-J sequence faults. To overcome this problem, the humanness should also be evaluated from the structure of the antibody. However, at present, this approach is limited to non-deep machine learning, so it is necessary to extract features from the structure in advance, which is extremely difficult to do. Deep learning is revolutionary in that it enables end-to-end prediction without prior extraction of such features. Therefore, we developed a new end-to-end method to evaluate the humanness of antibodies from 2D pixel images of antibody structures using CNN-VAE, which is a technique used to detect outliers in factory-produced products. This is expected to ensure the humanness of antibody sequences automatically generated by AI.

Antibody structures inform and improve machine learning predictions. We devise a method for the parameter-based unconstrained generation of synthetic lattice-based three-dimensional antibody-antigen-binding structures. Our method provides ground-truth access to conformational paratope, epitope, and affinity. We showcase the utility of synthetic datasets to benchmark the real-world relevance of machine learning models for antibody binding prediction.

Current technologies for antibody discovery and optimization have been widely successful, but are still subject to limitations. Established screening procedures are laborious, and targeting predetermined epitopes and optimizing multiple biophysical traits simultaneously remains a challenge. In this presentation, I will discuss emerging computational antibody design methods, which enable the targeted design of antibodies for predetermined epitopes and the prediction and modulation of their developability potential through the co-optimization of multiple biophysical properties. Overall, it is increasingly possible to complement well-established in vivo (first-generation) and in vitro (second-generation) methods of antibody discovery with in silico (third-generation) approaches, with time- and cost-saving benefits. These approaches are becoming sufficiently mature to be highly competitive for some applications, thus offering novel opportunities to streamline antibody development.

scifAI is a comprehensive, open-source explainable machine learning framework for the analysis of imaging flow cytometry data. In this presentation, I will focus on alterations to the immunological synapse, analyzing class frequency- and morphological changes of the cell, as well as showcasing the prediction of T cell cytokine production under stimulation with different antibodies, linking morphological features with function and thus demonstrating the potential to significantly impact antibody design.

Deep learning has led to substantial advances across many disciplines. However, many scientific problems of practical interest lack sufficiently large datasets amenable to deep learning. Prediction of antibody viscosity is one such problem where these methods have not yet been explored due to the relative scarcity of relevant training data. We will describe how we have overcome this limitation using a biophysically meaningful representation to develop generalizable deep learning models.

Immune repertoires represent a diverse collection of B and T cell receptors which interact with a seemingly infinite number of molecular structures. Recent advancements in deep sequencing and microfluidics allow high-throughput recovery of paired heavy and light chain sequences, thereby linking computational features of immune repertoires to biophysical properties of antibodies at an unprecedented resolution. I will explore the intersection of repertoires, ML, and biophysical features like antigen-specificity, affinity, and epitope.