A high-level representation of protein structure, the molecular surface, displays patterns of chemical and geometric features that fingerprint a protein’s modes of interactions with other biomolecules. We present MaSIF (molecular surface interaction fingerprinting), a conceptual framework based on a geometric deep learning method to capture fingerprints that are important for specific biomolecular interactions. Next, we sought to use MaSIF for the de design of de novo PPIs, a longstanding problem in computational protein design. We show that the learned surface fingerprints hold exquisite information that enables us to understand functional features of proteins as well as to design novel ones. 

Introducing disulfide bonds into protein has shown to improve protein thermostability and enhance function. Here we use a combination of bioinformatics and structure-based descriptors to predict putative disulfide crosslinking, and use rigorous physics-based methods (FEP+) to interrogate the effect of disulfides on protein thermostability or affinity. We show the approach can predict the effect of disulfide bonds on physical properties and is a valuable tool for in silico protein engineering. 

PD-1 expressed on activated T cells inhibits T cell function and proliferation to prevent an excessive immune response, and disease can result if this delicate balance is shifted in either direction. Using a combination of computation and experiment, we designed a hyperstable 40-residue miniprotein, PD-MP1, that specifically binds PD-1 at the PD-L1 interface. The apo crystal structure shows that the binder folds as designed, and trimerization of PD-MP1 resulted in a PD-1 agonist that strongly inhibits T cell activation. PD-MP1 was computationally designed with an all-beta interface, and the trimeric agonist could contribute to treatments for autoimmune and inflammatory diseases.

Natural antibodies are optimized for general “fitness” by evolution and in vivo selections. Taking natural protein sequences as a language spoken by nature, learning the underlying “grammars” and “semantics” can help various engineering tasks. We have built protein language models (PLMs) trained on >2 billion natural human antibody sequences. The model showed promising results in predicting affinity, expression, and functional readout when trained and evaluated on retrospective data.

Selection of multi-parameter optimized antibody molecules, taking into consideration biological function, safety, and developability, allows for streamlined and successful development. We developed an efficient and practical high-throughput developability workflow, which identified novel patterns and correlations between biophysical assays. These patterns and correlations represent the basis for training deep neural networks and establishing machine learning algorithms for in silico interrogation and prediction of developability profiles.

Transformer-based language models are powerful machine learning algorithms that can digest and extract information from huge text-based datasets. By applying language models to proteins, we can generate meaningful representations of their structural and sequence landscape, which in turn can be exploited to improve downstream prediction tasks, such as immunogenicity prediction, mutagenesis, and biophysical characterization.

Therapeutic antibodies require optimization of binding affinity and other properties. Traditional engineering approaches are time-consuming and explore only a subset of the solution sequence space. To address these challenges, we assist antibody development with AI. Models trained with affinity measurements of sequence variants of trastuzumab could quantitatively predict the binding strength of unseen variants. Models can also score antibody sequences for naturalness by comparison with human antibody repertoires, mitigating downstream developability issues.

I will present a set of machine learning technologies for the de novo design of antibodies as high-affinity binders to given epitopes. Our methods encompass structure-based design of CDRs for optimal epitope recognition and sequence-based generative models ensuring favorable developability properties. We show that we obtain nanomolar Fab binders to a specified novel epitope.

Efficient exploration of sequence outputs from discovery campaigns is often hindered by inefficient sampling of the CDR diversity and technical hurdles involved with the handling of big data, particularly from multiplexed experiments. To overcome these limitations, we developed a cloud-based bioinformatics platform, AbXtract™, which utilizes population-based statistics and unsupervised clustering of next-generation sequencing (NGS) data to rapidly identify leads from distinct and/or overlapping populations. To validate the platform, we carried out a discovery campaign using our in vitro Generation 3 Antibody Library Platform to select antibodies against the SARS-CoV-2 Spike protein trimer and its RBD and S1 subunits. After demultiplexing the NGS datasets from antibody selections against different antigen populations at decreasing concentrations, we were able to effectively sort through the maze of sequencing data to identify and test many of these top candidates for kinetics, binning, and in vitro neutralization experiments. The correlation of these data with NGS metrics strengthened our ability to identify high quality leads directly from NGS datasets, with many predicted leads displaying favorable properties such as sub-nanomolar affinities (≥13pM), distinct binding profiles and potent neutralization (IC50’s <1ng/ml) of many SARS-CoV-2 variants of concern.

We developed a novel high-throughput computational pipeline capable of generating millions of signal peptide mutants and utilizes a deep learning model to predict which of these mutants can alleviate the N-terminal miscleavage in antibodies. The pipeline was optimized to screen a library of 296077 unique mutants for each input antibody. We applied it to five antibodies with various extent of miscleavage. In each case, multiple mutants were obtained, with miscleavage reduced to a non-detectable level and titers comparable with or better than that of the original signal peptides.

Machine learning has shown its power across all of biology and in this talk, I will describe some of the novel machine learning tools we are pioneering in the area of biotherapeutics from computational humanization to accurate rapid structure prediction and virtual high-throughput screening.

We improve the binding of antibodies to a desired target and eliminate non-specific binders using machine learning (ML) models that are trained on CDR H3 sequences from high-throughput phage display assays. We tested 77,599 novel ML designed sequences from 6 ML methods and found that ML could provide superior binders. We next used single-target ML models to eliminate non-specific antibodies and observed that ML methods outperformed conventional affinity competition assays.

The continual evolution of the severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2) and the emergence of variants that show resistance to vaccines and neutralizing antibodies threatens to prolong the coronavirus disease 2019 (COVID-19) pandemic. Selection and emergence of SARS-CoV-2 variants are driven in part by mutations within the viral spike protein and in particular, the receptor-binding domain (RBD), which binds to the human ACE2 receptor and is a primary target site for neutralizing antibodies. Comprehensive mapping of single-position substitutions in the RBD has revealed key mutations that enhance binding to ACE2 and provide an escape from neutralizing antibodies. However, several SARS-CoV-2 variants such as beta (B.1.351) gamma (P.1), and delta (B.1.617.2) possess multiple, combinatorial mutations in their RBD. Here, we develop deep mutational learning (DML) - a machine learning-guided protein engineering technology that enables the comprehensive interrogation of combinatorial mutations in the RBD and prediction of their impact on ACE2 binding and antibody escape. DML reveals a highly diverse sequence landscape of possible variants that maintain or enhance binding to ACE2 and escape from different classes of neutralizing antibodies. DML may be used in the future to comprehensively profile the breadth of candidate therapeutic antibodies against existing and prospective variants of SARS-CoV-2.

Next-generation sequencing of antibody repertoires has provided new insights into the single domain antibody repertoire.  The correlation between the B-cell receptor repertoire and the serological antibody repertoire has only been analyzed in a small number of studies.  In this talk, we will discuss the use of serum antibodies to guide antibody discovery in an immunized llama.  Further, we use an autoencoder to deeply mine the clonal lineage of serum-identified antibodies.

The prediction of antibody-antigen binding is a central question in biotechnology. A fundamental premise for the predictability of antibody-antigen binding is the existence of paratope-epitope interaction motifs universally shared among antibody-antigen structures. In a dataset of non-redundant antibody-antigen structures, we discovered a motif vocabulary of paratope-epitope interactions that govern antibody specificity providing the proof-of-principle that antibody-antigen binding is predictable with implications for de novo antibody and (neo-)epitope design.

De novo design methods promise a cheaper and faster route to antibody discovery, while enabling the targeting of predetermined epitopes and the screening of multiple biophysical properties. I will present recent advances to design antibodies targeting structured epitopes and to predict solubility and formulation condition. We show that we can rapidly obtain highly stable nanobodies binding predetermined epitopes with affinities in the nM range, and that we can accurately predict protein solubility at varying pH values.