Protein Sciences In Drug Discovery 101

Protein science underpins every drug discovery project, and in this episode of the Drug Discovery 101 webinar series, we invited experts from GSK and AstraZeneca to show how they make proteins at industry scale. The session explores how high-throughput expression platforms, AI prediction models, lab automation and FAIR data help pharma scientists deliver the right protein, at the right quality, at the right time.

Speakers:

Maja Firczuk, Protein Expression Technologies Group Leader, Protein and Cellular Sciences, Large Molecule Research, GSK
Bradley Peter, Associate Principal Scientist, Protein Science, AstraZeneca
Webinar host: Siddique Amin, Senior Scientist, Peptide Chemistry, AstraZeneca

Watch the webinar recording now

*If the webinar does not appear, please ensure you have accepted the “Experience” cookies for this website. Alternatively, you can view it directly on YouTube here.

Key takeaways

Protein reagents support assays, structural biology and hit finding, so slow protein supply slows whole projects.
Not all proteins are equally hard to make, so GSK builds modular platforms that match the screening approach to protein complexity.
Screening hundreds of constructs at once, including likely failures, supports data-driven decisions and produces the negative data predictive models need.
AstraZeneca uses AI tools to cut target scoping from days to under a minute, and automation to speed up molecular biology, screening and purification.
Computationally designed proteins are hypotheses: they reduce experimental iterations but still need validation for structure, activity and fitness for purpose.
Consistent, FAIR data reporting decides whether AI models for protein science work at all.

Why protein reagents matter in drug discovery

Protein reagents are used in biochemical assays, in selections for small and large molecule binders, in structural biology, in mechanism-of-action studies and in quality control of drug molecules. Early projects need many different proteins in small quantities, including orthologs and paralogs to check for desired and undesired cross-activity. Later stages usually need the human protein in much larger amounts.

Historically, getting the first protein batch has taken two to four months. The process relies on expert knowledge, manual handling, multiple handoffs and information spread across disconnected systems, which is what both GSK and AstraZeneca are working to change.

GSK: matching platforms to protein complexity

Maja Firczuk, Protein Expression Technologies Group Leader at GSK, argues that “not all proteins are made equal”, so GSK builds platforms that match the experimental approach to the complexity of each protein reagent. Every request starts with three questions: where to express the protein, what to express (domain boundaries, signal peptides, variants) and how to purify it.

Screening hundreds of constructs at once

When the answers are uncertain, Maja’s team screens many constructs in one go so each design only needs testing once. Their high-throughput mammalian expression platform handles soluble and membrane proteins differently:

Soluble (secreted) proteins: up to 320 constructs per run, with controls on every plate, replicates, small-scale purification and quality control before scale-up.
Membrane proteins: fluorescence-based screening, microscopy and a small-scale extractability screen to choose the best constructs and detergent conditions.

Two worked examples

Cytokines: 14 recombinant cytokines were optimised at the same time, with 15 constructs each (five signal peptides combined with three tags).
Domain boundaries: for one target, a full 320-construct library explored domain boundaries for a reagent used in SPR quality control. The best constructs bound better, expressed higher and needed far less protein per assay than the full-length reagent, which would have required up to a gram. Shifting a boundary by a single amino acid sometimes changed yield dramatically.

Capturing the data that models need

Because the platforms record everything on the plate, they also capture negative data, which Maja named as the biggest limitation for predictive models. Automated purification runs five two-step purifications overnight, and in-house apps turn quality control results into structured data and certificates of analysis.

AstraZeneca: AI for protein

Bradley Peter, Associate Principal Scientist in Protein Science at AstraZeneca, presents the company’s “AI for protein” strategy. It combines in silico decision support, automated lab execution and connected data systems to deliver the right protein, at the right quality, with fewer experimental iterations.

AI tools for scoping and design

Protein Scoper: enter a UniProt accession and the tool returns ranked expression systems, construct options, risks and suggested next steps in under a minute, compared with a couple of days manually. It was built with AI coding assistance and little coding experience, and Bradley noted that other teams could replicate it.
Protein Hub: one workflow from sequence analysis and expression prediction through construct design to internal ordering, linked to the molecular biology platform and protein database.
Expression prediction: AstraZeneca makes well over 2,000 constructs a year, and only about a quarter reach scale-up. A model developed with EMBL-EBI uses protein language models to predict whether a sequence will express in E. coli, and a second model looks at codon optimisation.
In silico protein design: for a hydrophobic target domain that could not be produced, redesigning the surrounding residues gave four soluble expressers out of five, and one construct went on to a solved crystal structure.

Automation across the workflow

AstraZeneca has automated plasmid assembly, transformation and plasmid preparation. Plasmid delivery currently takes about three weeks because DNA synthesis is outsourced, so an in-house pipeline aims to go from digital sequence to plasmid in five days. A gene-to-screen platform using capillary electrophoresis replaces manual gel densitometry, and semi-automated parallel purification frees scientists from repetitive processing.

Why data quality decides everything

Increasing lab throughput without better data infrastructure simply moves the bottleneck to data entry. Automated QC reports and database-to-ELN integration reduce duplicate reporting. Bradley’s message on models: garbage in, garbage out. Expression results are recorded as mg/L by some groups and as low, medium or high by others, negative data is rarely reported, and public and private datasets differ in size, quality and overlap. A recent publication from pharma and academic collaborators proposes a reporting framework for protein expression and purification data.

Career skills and Q&A highlights

Skills for the future: Maja highlighted data analysis, including understanding data limitations, plus automation and high-throughput approaches, as the “lab in the loop” vision gets closer.
Small versus large molecules: small molecule projects care about activity, so truncated or bacterially made proteins can be acceptable. Biologics projects care about shape, so they usually need full-length protein made in human cells with human modifications.
Models and experiments: AstraZeneca still tests a few constructs that its model predicts will not express. The model is trained on well-behaved E. coli proteins and is weaker for targets such as intrinsically disordered proteins, and failed constructs add negative data.
Looking ahead: Bradley hopes structured data capture will be built into instruments and models within five years, and mentioned the SGC’s goal of a small molecule binder for every human protein by 2030.

Frequently asked questions

What is protein science in drug discovery?

Protein science covers designing, producing, purifying and characterising the proteins used as reagents in drug discovery. These proteins support biochemical assays, binder selection, structural biology, mechanism-of-action studies and quality control.

Why are protein reagents so important?

Without protein there is no drug discovery project. Delays in producing the first protein batch, which has historically taken two to four months, hold up assays, structural work and hit finding.

How is AI used to make proteins at AstraZeneca?

AstraZeneca uses AI for target scoping (Protein Scoper), construct design (Protein Hub), expression prediction and in silico protein design. These tools generate hypotheses and reduce experimental iterations, but results still need experimental validation.

Why does negative data matter?

Predictive models learn from failures as well as successes. Both speakers said high-confidence negative data, such as constructs that did not express, is scarce because it is rarely recorded or reported.

What does FAIR data mean?

FAIR data is findable, accessible, interoperable and reusable. In protein science it means recording expression and purification results consistently so teams and models can reuse them.