Early developability screens are used to predict the downstream biophysical characteristics and manufacturability of candidate drug biologics. To realize the full potential of early developability screens, a fully automatable, predictive, and high-throughput developability screen is needed. We present our data-driven approach to increasing the throughput of the early developability phase to accommodate a growing pipeline and generate the data needed to construct in silico developability prediction models.

Predicting protein aggregation remains difficult, limiting their use for biotechnology and therapeutic applications. We aim to design aggregation-resistance by collecting and learning from large, experimentally validated datasets. I quantified aggregation after thermal and pH stress for thousands of small protein domains using mass spectrometry. I’m developing machine learning models to predict aggregation from protein features. Through iterative experiments and design, I will refine my model to achieve unprecedented aggregation-resistance. 

Selection of highly developable leads is crucial for clinical translation and requires accurate developability assessments benchmarked against the clinical landscape. Combining large in vitro datasets with in silico tools, we have developed integrated wet lab and dry lab workflows that enable rational selection of both assays and candidates. The resulting developability pipeline enables efficient identification of highly developable leads with consideration of specific downstream risks.

The increase in biologics administered subcutaneously has required higher protein concentrations and highlighted liabilities such as protein aggregation, precipitation, and high viscosity. Identifying optimal high-concentration formulations that limit these liabilities can be slow and costly, and often prevent therapeutics from moving rapidly into the clinic/market. Here, we present advances that have been made in understanding high-concentration protein behavior as well as interesting case studies.

Initial antibody discovery generates molecules with a wide range of biophysical characteristics that can be used to predict developability, presenting an opportunity to filter or improve their properties. We present machine learning models for developability predictions of properties such as hydrophobicity, chemical stability, and viscosity, and explain how they are deployed to obtain actionable information. We describe the generation and benchmarking of the models and associated experimental input training data.

Protein aggregation impacts industrial protein production and formulation. Aggrescan3D (A3D) was developed to aid in understanding and engineering aggregation in globular proteins. It has become one of the most popular structure-based predictors for aggregation studies and protein redesign. Here, we present Aggrescan4D (A4D), which largely extends A3D’s functionality by incorporating pH-dependent aggregation prediction and an evolutionarily informed automatic mutation protocol to engineer protein solubility.

Because membrane proteins can be structurally dynamic, they are often unstable or difficult to discern mechanism of their many conformations. The chemical intuition and tools for engineering/design of protein in lipid are still far from the advanced capabilities for water-soluble proteins. Our group uses de novo design of simple transmembrane (TM) proteins for rapid Build-Test-Learn cycles to develop software that reliably encodes protein stability and molecular recognitions for membrane protein engineering. A generative sequence design method focused on TM interactions in lipid was tested by design and in vitro folding of >20 synthetic TM protein assemblies, where most correctly folded. A few reached hyper-stability, folded in high SDS, temperature, etc.—providing feedback on what molecular features/patterns idealized sequence-structure principles between TM spans. With improved principles, we devised an in silico approach to design de novo TM proteins that bind and recognize target membrane proteins directly by their TM spans to modulate structure and function. Past work proved we can recognize and functionally regulate single-pass proteins by their TM spans, integrins and EPO cytokine receptor. We recently advance the technology to correct misfolding of an ion channel and target GPCRs. These tools and chemical “rules” have advanced targeting and stabilization tools for membrane proteins.

Biological relevance tools are essential to gain knowledge about attribute impact which can be used to inform clinically relevant specifications.  A new data science method called the Clinical Impact of Attributes (CIA) approach will be shared that uses clinical trial information to justify clinically safe specifications.  CIA analyzes clinical studies to determine if any correlations exists between attribute levels exposed in patients and the development of adverse events.  Several case studies will be shown.

Approved therapeutic antibodies provide valuable insights into which biophysical properties can be considered safe from a developability perspective, aiding the design of biotherapeutics with de-risked developability profiles. I will present our work on building a database of therapeutic antibodies (TheraSAbDab), using this to develop a predictive tool for developability risk (Therapeutic Antibody Profiler 2), and finally our generative machine learning model (p-IgGen) for creating developability-conditioned in silico screening libraries.

Effective machine learning (ML) relies on high-quality data and standardized analysis procedures. We will explore the critical need for a collaborative, community-driven approach to standardizing ML analytics and contemplate strategies for producing better data. By establishing best practices and shared resources, development groups across the industry will be empowered to efficiently integrate different data types and leverage the full toolbox of ML techniques, ensuring reproducibility, interpretability, and robust model performance.

Public databases of sequences and bioactivity data, including from patent extractions, are a crucial but overlooked resource for peptide researchers. Because natural endogenous or designed therapeutic peptides fall between the formal representations of small-molecule cheminformatics and protein-sequence bioinformatics, they have database representational challenges that make them difficult to find. This presentation will review various sources of peptide entries in PubChem and offer searching tips.