Leveraging Data Science for Enhanced Protein Expression

Teaching AI models’ protein prediction can be slow and cumbersome, where design cycles (design, make, test, and analyse) often take months and require specialist expertise across multiple teams. The data that research teams have is often not suitable for training models because it is generated to support a specific project and is not standardised. Using IVTT (cell-free protein expression), integrated liquid handlers, and automated data analysis, we streamline the process and generate fit-for-purpose data. In 7 days, we go from DNA to analysed data.

AstraZeneca’s Protein Sciences & Analytics is building the lab of the future based on an automation-first operating model and AI-ready data capture. The enterprise-scale blueprint is grounded in harmonized ways of working, small- and large-scale HT expression platforms, and integrated data pipelines. The ongoing transformation has already shown measurable impact on DMTA timelines and first-pass success and aims to be transferable across modalities and sites—offering a clear path to enable faster, data-driven decisions.

High-throughput biophysical data generation is emerging as the missing link in AI-driven protein design. We present a platform for scalable production of diverse drug target proteins combined with multi-modal characterisation, including nano-DSF, BLI, SPR, and FIDA. Our approach captures protein stability, aggregation, and binding interactions—particularly for VHHs—at high granularity. By generating paired datasets that link sequence, expression, and biophysical behaviour, we expand both the diversity and resolution of training data. This data-centric strategy enables more robust, generalisable, and predictive AI models, helping to unlock design capabilities in data-sparse and challenging regions of protein space.

Improving protein production is critical in today’s rapidly advancing protein technology landscape. Using experimental data, we developed a codon-efficiency metric that correlates with the levels of native and recombinant proteins in Escherichia coli. Codon content influences protein expression more than mRNA folding, except in the first few codons, by modulating translation and mRNA degradation. An A-rich, G-poor base composition in the first six codons enhances expression and mRNA stability. We integrated these findings into a sequence optimisation algorithm that predicts the expression of a given DNA sequence in E. coli and proposes strategies for high-yield protein production.

Protein engineering increasingly demands methods that can efficiently navigate vast sequence spaces to identify variants with desired properties. We developed an integrated deep mutational learning pipeline combining high-throughput experimental fitness landscapes with protein language models to engineer proteins across two distinct applications: enhancing myoglobin peroxidase activity through electron-hole hopping pathways, and tuning the sensitivity and dynamic range of HucR-based biosensors. In both systems, neural network models trained on sort-seq and EP-Seq data accurately predicted improved variants, achieving up to 100-fold enrichment over random mutagenesis. Our results establish a generalisable framework for accelerating precision protein engineering.

We present our experience expressing over 80,000 AI-designed proteins using cell-free systems, systematically evaluating expression tags, codon optimisation strategies, and system variants. Applying our automated cloud laboratory for high-throughput protein validation we describe integrating these workflows into this fully automated platform capable of expressing arbitrary customer proteins at scales from sub-microgram to milligram quantities.