Randall Ketchem


Randy received his PhD in Molecular Biophysics from The Florida State University, focusing on experimental protein structure determination, resulting in the first membrane-bound protein structure solved by Solid State NMR. As part of this effort he developed computational methods for the structure calculation and refinement of membrane-bound proteins.

More Info

Randy continued his training in structural biology in a postdoctoral fellowship at The Scripps Research Institute where he expanded his experimental and computational efforts into solution-based protein structure. Randy then joined Immunex in 1997 to apply his expertise in protein structure modelling and analysis to efforts in the development of biologic-based therapeutics. During his many years of experience at Immunex, Amgen, and Just Biotherapeutics he has led efforts to engineer protein therapeutics, invented novel therapeutic modalities and computational methods for therapeutic design, and developed computational and experimental approaches to understanding and controlling biophysical properties of biological macromolecules. Randy has made significant contributions in antibody design, epitope mapping, molecular assessment, stability engineering, construct design, protein structure analysis, and protein engineering. At Just – Evotec Biologics, Randy leads the Molecular Design team, integrating biologics discovery and molecular design strategies into the entire therapeutic pipeline.


How does optimization of antibodies improve development?

Just – Evotec Biologics has developed a molecular optimisation platform for monoclonal antibodies to increase productivity, eliminate potential degradative modifications, and decrease inherent instability that may manifest during processing or long-term storage as aggregation, precipitation, viscosity, and chemical instability. This optimisation platform has been proven to decrease process development time and significantly improve yields, while reducing the potential need to re-engineer a late-stage molecule due to sequence issues. Molecular optimization begins with evaluation of the antibody sequence and molecular structure using Just’s proprietary Abacus™ in silico design suite. Abacus™ evaluates stability, germline background and pairing, potential post-translational modifications (PTM), missing or inserted residue errors, potential immunogenicity, and can drive the engineering modifications necessary to repair or modify antibody sequences. Abacus™ also drives germline switching applicable to both improving germline pairings and to humanisation by understanding residue positions required for core fold stability as well as potential antigen interaction. Characterisation of recombinant material using high-throughput biochemical, biophysical analytical, and mass spec tools are also used to help guide an understanding of molecular behaviour and can be used to guide optimisation designs.


What is the limitation of in vivo-derived antibodies?

In vivo-derived antibodies are matured from germline through somatic hypermutation. While this process allows for the generation of a very large range of antibody sequence space, which in turn allows for tremendous levels of antibody specificity and activity, hypermutation can also introduce mutations that lead to poor stability. This stability decrease can manifest itself in multiple ways, such as poor folding stability, leading to poor expression, increased aggregation, increased particulation, decreased resistance to a large range of pH exposure or agitation, decreased thermal stability, and other issues. We have developed a machine-learned neural network method, RANDAb (Residual Artificial Network for the Design of Antibodies), to evaluate the sequences for fitness and potential engineering to optimise in vivo-derived antibodies. Another issue is epitope biasing in which the B-cell response is focused on a dominant epitope. This leads to a loss of therapeutic diversity that a library screen could overcome. RANDAb uses a deep learning approach to capture higher-order interactions between every residue in an antibody sequence to model pairwise interactions. Using a model architecture and training procedure inspired by the latest research in the field of natural language processing, the core model of RANDAb encodes information about the properties of amino acids and the structure of antibodies, learned from millions of curated, matured human antibody sequences. RANDAb uses this model to assign probabilities to each amino acid in an antibody sequence and provides suggested mutations where these probabilities are low. These low probability residues represent places in an antibody sequence where the current amino acid is predicted to be a bad fit with the surrounding residues, offering attractive targets for molecular optimisation. The RANDAb method is fully automated for detection and residue replacement suggestion. Optimisation positions from RANDAb, structure evaluation, surface properties, and  posttranslational modifications then involve expert evaluation and structure-based protein engineering design to produce a set of potential combinatorial variants in an effort to improve antibody properties. Once optimisation positions for  both stability and post-translational modifications are identified they are used within Abacus™ to build combinatorial variants to optimize the sequence for therapeutic application while simultaneously maintaining function. This often may be accomplished in a single round of engineering and subsequent testing.


What were the limitations of previous library approaches and what makes J.HALSM so special?

J.HALSM is by design representative of a human response, but with the added ability to bias toward specific efficacy and developability characteristics such as HC-CDR3 length, surface properties, stability, avoidance of determinantal posttranslational modifications, and so on. Historically, libraries were built using random mutagenesis, resulting in most of the library being unusable due to clipping, stop codons, and un-manufacturable antibodies. More modern libraries generally utilise germline frameworks with CDR diversity. While this approach improves the library quality, it does not represent a human response with proper diversity in full somatic hypermutation. Some libraries further use positional frequency analysis to generate artificial CDR diversity which does not represent a multi-dimensional human sequence response. By using the J.HALSM GAN methodology we are able to generate clean sequences with true human representation while being able to bias toward desired properties.


Tell us about your vision – what do you expect from therapeutic antibodies in the future?

The use of J.HALSM not only results in a robust discovery platform, but also gives the ability to explore antibody biology in a hypothesis-driven manner. For example, Fv-directed mechanisms which impact pharmacokinetics, tissue interactions, clearance rates, degradation, titre, aggregation, particulation, intracellularisation, blood-brain barrier passage, and more. Coupling an understanding and purposeful engineering of these properties with multi-specific capabilities, such our J.HALSM expansion into VHH domains, will allow for both target and epitope-specific therapeutic modality designs with engineered biological functionality.