Transforming Therapeutic Protein Engineering Bram Estes1, Carolyn Shomin1, Marissa Mock1, Suzanne Edavettal1 1Department of Protein Therapeutics, Amgen Inc, Thousand Oaks, CA 91320, USA. To accelerate the design and success of therapeutic proteins, we have evolved from a traditional rational design and measurement approach to a predictive, AI/ML-driven platform. This next-generation platform combines predictive screening with generative biology to rapidly engineer proteins optimized for manufacturability, efficacy, and safety. Our AI/ML tools are tailored to protein therapeutic challenges—guiding engineering decisions to improve properties such as low viscosity for high-dose administration and eliminating chemical liabilities like oxidation or isomerization. Critically, our predictive platform is coupled with our Next Generation CHO expression system that mirrors our manufacturing process, using the same host cell and vector system. This system has demonstrated predictive production yields and product quality across a range of therapeutic modalities ensuring design decisions made early in discovery are highly translatable to scale-up and GMP production.

Multi-specific antibodies resemble an extreme diversity of formats and designs that facilitate a wide range of functionalities and modes of action. The search for suitable candidates with the desired properties is quite tedious and limits a timely identification of the best lead molecules. Already the rather simple combination of different binding arms in bispecific antibodies representing a variety of affinities, epitopes, distances and flexibility can exhaust the capacity of conventional methods quite quickly. Therefore, we established a platform based on trans-splicing that allows the seamless in vitro reconstitution of many individual antibody fragments into fully functional multi-specific antibodies, thus enabling a fast turnover of well-defined building blocks amenable to high throughput and automation. Furthermore, this universal platform process avoids the generation of product related impurities, is agnostic to a wide variety of binders and enables fast reshuffling into the final candidate. This approach resolves the bottleneck of molecule prototyping through combinatorial screening. The achieved quantity and purity is sufficient for a multitude of functional assays, speeding up the discovery process for multi-specific antibodies which are in high demand for modern therapies.

Advances in protein engineering are driving transformative applications in biocatalysis, therapeutics, enzyme design, and synthetic biology. A major challenge in experimental protein optimization lies in efficiently designing stable, soluble, and expressible variants prior to synthesis and characterization. To address this, we have developed SPASE (Soluble Protein Analog Selection Engine), a web-based tool that leverages deep learning and computational filtering to generate high-quality protein analogs, with favorable solubility, stability, and reduced aggregation propensity. SPASE takes an input protein structure (PDB format, monomeric) and generates 10,000 synthetic sequence variants using ProteinMPNN, retaining the overall 3D fold of the original template while enhancing key properties. The generated sequences are then filtered through a automated multi-step computational pipeline to ensure solubility and reduced aggregation propensity: 1) Protein-Sol assesses solubility, 2) ESMFold predicts the 3D folds of the top 100 candidates, 3) Aggrescan3D identifies and removes sequences with high aggregation propensity to further refine the selection. Finally, a selection of sequences with the lowest aggregation propensity is obtained, prioritizing candidates with the highest confidence folding scores. The final output is a curated set of ~25-35 high-quality protein analogs optimized for experimental expression and characterization. Importantly, SPASE allows users to preserve functionally critical residues—such as catalytic or binding sites—ensuring that designed proteins remain biologically relevant. By integrating these computational tools into an easy-to-use web interface, SPASE lowers the barrier for experimentalists to incorporate rational and combinatorial in silico library design into their workflows, accelerating the development of improved biocatalysts, therapeutics, enzymes, and other functional proteins. We hope that SPASE will help facilitate the transition from computational predictions to experimental validation, bridging the gap between in silico design and real-world applications. Our aim is to demonstrate how computational advancements can streamline protein engineering, making it more accessible and effective for experimentalists while democratizing in silico protein design across diverse fields, from green chemistry to industrial biotechnology and pharmacology.

A long-standing challenge in the biopharmaceutical industry has been the need for a rapid Cell Line Development (CLD) process capable of screening a significant number of clones for the ability to produce high amounts of drug substance with the desired product of interest. To achieve high expression of monoclonal antibodies (mAbs), we have employed a flow-based cytometric reporter system to screen for productivity. FLARE (FLow-cytometric Attenuated Reporter Expression) uses an alternate start cell surface reporter which serves as a specific productivity surrogate. FLARE can be used analytically to identify productivity differences in both pools and clones. It can also be manipulated via sorting to isolate individual cells (clone generation) and pooled populations (bulk sorting) to yield either pools or clones with enriched productivities. It has proven to be a powerful tool for both enriching populations and yielding high producing clonal mAb cell lines. In recent years, multispecific antibodies have emerged as promising candidates for new therapeutics. Although these are mAb-like, they typically are of greater complexity, in part due to being comprised of > 2 peptide chains. These molecules can pose a challenge for FLARE, as enrichment of just one peptide chain coding transcript could lead to a transgene expression imbalance. To this end, we developed a second reporter system (FLARE 2.0), that allows for the ability to simultaneously enrich for cells expressing two different transgenes, at various levels. This presentation will highlight the utility of the FLARE 2.0 system in the establishment of cell lines expressing a 3-chain multispecific molecule with optimal product quality.