Search

The antibody drug field has continually sought improvements to methods for candidate discovery and engineering. Historically, most such methods have been laboratory-based, but informatics methods have recently started to make an impact. Deep learning, a subfield of machine learning, is rapidly gaining prominence in the biomedical research. Recent advances in microfluidics technologies and next-generation sequencing have not only revolutionized therapeutic antibody discovery, but also contributed to a vast amount of antibody repertoire sequencing data, providing opportunities for deep learning-based applications. Previously, we used microfluidics, yeast display, and deep sequencing to generate a panel of binder and non-binder antibody sequences to the cancer immunotherapy targets PD-1 and CTLA-4. Here we encoded the antibody light and heavy chain complementarity-determining regions (CDR3s) into antibody images, then built and trained convolutional neural network models to classify binders and non-binders. To improve model interpretability, we performed in silico mutagenesis to identify CDR3 residues that were important for binder classification. We further built generative deep learning models using generative adversarial network models to produce synthetic antibodies against PD-1 and CTLA-4. Our models generated variable length CDR3 sequences that resemble real sequences. Overall, our study demonstrates that deep learning methods can be leveraged to mine and learn patterns in antibody sequences, offering insights into antibody engineering, optimization, and discovery.

Abstract Ecological communities are being impacted by global change worldwide. Experiments are a powerful tool to understand how global change will impact communities by comparing control and treatment replicates. Communities consist of multiple species, and their associated abundances make multivariate methods an effective approach to study community compositional differences between control and treated replicates. Dissimilarity metrics are a commonly employed multivariate measure of compositional differences; however, while highly informative, dissimilarity metrics do not elucidate the specific ways in which communities differ. Integrating two multivariate methods, dissimilarity metrics and rank abundance curves (RACs), have the potential to detect complex differences based on dissimilarity metrics and detail the how these differences came about through differences in richness, evenness, species ranks, or species identity. Here we use a database of 106 global change experiments located in herbaceous ecosystems and explore how patterns of ordinations based on dissimilarity metrics relate to RAC‐based differences. We find that combining dissimilarity metrics alongside RAC‐based measures clarifies how global change treatments are altering communities. We find that when there is no difference in community composition (no distance between centroids of control and treated replicates), there are rarely differences in species ranks or species identities and more often differences in richness or evenness alone. In contrast, when there are differences between centroids of control and treated replicates, this is most often associated with differences in ranks either alone or co‐occurring with differences in richness, evenness, or species identities. We suggest that integrating these two multivariate measures of community composition results in a deeper understanding of how global change impacts communities.

Abstract Background The anti-tumor activity of anti-PD-1/PD-L1 therapies correlates with T cell infiltration in tumors. Thus, a major goal in oncology is to find strategies that enhance T cell infiltration and efficacy of anti-PD-1/PD-L1 therapy. TGF-β has been shown to contribute to T cell exclusion, and anti-TGF-β improves anti-PD-L1 efficacy in vivo. However, TGF-β inhibition has frequently been shown to induce toxicity in the clinic, and the clinical efficacy of combination PD-L1 and TGF-β blockade has not yet been proven. To identify strategies to overcome resistance to PD-L1 blockade, the transcriptional programs associated with PD-L1 and/or TGF-β blockade in the tumor microenvironment should be further elucidated. Results We used single-cell RNA sequencing in a mouse model to characterize the transcriptomic effects of PD-L1 and/or TGF-β blockade on nearly 30,000 single cells in the tumor and surrounding microenvironment. Combination treatment led to upregulation of immune response genes, including multiple chemokine genes such as CCL5, in macrophages, and downregulation of extracellular matrix genes in fibroblasts. Analysis of publicly available tumor transcriptome profiles showed that the chemokine CCL5 was strongly associated with immune cell infiltration in various human cancers. Further investigation with in vivo models showed that intratumorally administered CCL5 enhanced cytotoxic lymphocytes and the anti-tumor activity of anti-PD-L1. Conclusions Taken together, our data could be leveraged translationally to complement or find alternatives to anti-PD-L1 plus anti-TGF-β combination therapy, for example through companion biomarkers, and/or to identify novel targets that could be modulated to overcome resistance.

Conventionally, hyperimmune globulin drugs manufactured from pooled immunoglobulins from vaccinated or convalescent donors have been used in treating infections where no treatment is available. This is especially important where multi-epitope neutralization is required to prevent the development of immune-evading viral mutants that can emerge upon treatment with monoclonal antibodies. Using microfluidics, flow sorting, and a targeted integration cell line, a first-in-class recombinant hyperimmune globulin therapeutic against SARS-CoV-2 (GIGA-2050) was generated. Using processes similar to conventional monoclonal antibody manufacturing, GIGA-2050, comprising 12,500 antibodies, was scaled-up for clinical manufacturing and multiple development/tox lots were assessed for consistency. Antibody sequence diversity, cell growth, productivity, and product quality were assessed across different manufacturing sites and production scales. GIGA-2050 was purified and tested for good laboratory procedures (GLP) toxicology, pharmacokinetics, and in vivo efficacy against natural SARS-CoV-2 infection in mice. The GIGA-2050 master cell bank was highly stable, producing material at consistent yield and product quality up to >70 generations. Good manufacturing practices (GMP) and development batches of GIGA-2050 showed consistent product quality, impurity clearance, potency, and protection in an in vivo efficacy model. Nonhuman primate toxicology and pharmacokinetics studies suggest that GIGA-2050 is safe and has a half-life similar to other recombinant human IgG1 antibodies. These results supported a successful investigational new drug application for GIGA-2050. This study demonstrates that a new class of drugs, recombinant hyperimmune globulins, can be manufactured consistently at the clinical scale and presents a new approach to treating infectious diseases that targets multiple epitopes of a virus.

In vitro affinity maturation of therapeutic monoclonal antibodies is commonly applied to achieve desired properties, such as improved binding kinetics and affinity. Currently there are no universally accepted protocols for generation of variegated antibody libraries or selection thereof. Here, we performed affinity maturation using a yeast-based single-chain variable fragment (scFv) expression system to compare two mutagenesis methods: random mutagenesis across the entire V(D)J region by error-prone PCR, and a novel combinatorial mutagenesis process limited to the complementarity-determining regions (CDRs). We applied both methods of mutagenesis to four human antibodies against well-known immuno-oncology target proteins. Detailed sequence analysis showed an even mutational distribution across the entire length of the scFv for the error-prone PCR method and an almost exclusive targeting of the CDRs for the combinatorial method. Though there were distinct mutagenesis profiles for each target antibody and mutagenesis method, we found that both methods improved scFv affinity with similar efficiency. When a subset of the affinity-matured antibodies was expressed as full-length immunoglobulin, the measured affinity constants were mostly comparable to those of the respective scFv, but the full-length antibodies were inferior to their scFv counterparts for one of the targets. Furthermore, we found that improved affinity for the full-length antibody did not always translate into enhanced binding to cell-surface expressed antigen or improved immune checkpoint blocking ability, suggesting that screening with full-length antibody or antigen-binding fragment formats might be advantageous and the subject of a future study.

To discover therapeutically relevant antibody candidates, many groups use mouse immunization followed by hybridoma generation or B cell screening. One modern approach is to screen B cells by generating natively paired single chain variable fragment (scFv) display libraries in yeast. Such methods typically rely on soluble antigens for scFv library screening. However, many therapeutically relevant cell-surface targets are difficult to express in a soluble protein format, complicating discovery. In this study, we developed methods to screen humanized mouse-derived yeast scFv libraries using recombinant OX40 protein in cell lysate. We used deep sequencing to compare screening with cell lysate to screening with soluble OX40 protein, in the context of mouse immunizations using either soluble OX40 or OX40-expressing cells and OX40-encoding DNA vector. We found that all tested methods produce a unique diversity of scFv binders. However, when we reformatted forty-one of these scFv as full-length monoclonal antibodies (mAbs), we observed that mAbs identified using soluble antigen immunization with cell lysate sorting always bound cell surface OX40, whereas other methods had significant false positive rates. Antibodies identified using soluble antigen immunization and cell lysate sorting were also significantly more likely to activate OX40 in a cellular assay. Our data suggest that sorting with OX40 protein in cell lysate is more likely than other methods to retain the epitopes required for antibody-mediated OX40 agonism.