Improving Staphylococcus aureus vaccine thermostability by 2.5 °C in a single round, 7× faster than rational design

Context

Staphylococcus aureus is a pathogenic bacterium responsible for a wide range of infections, from minor skin infections to severe invasive disease. The bacterium was associated with over 1 million deaths in 2019, making it a significant global health burden. Vaccine development for S. aureus remains challenging due to the bacterium's immune evasion mechanisms and antigenic diversity.

Thermostability is a critical property for vaccine antigens, particularly for distribution in regions with limited cold chain infrastructure. Unstable antigens can denature during storage or transport, losing immunogenicity and rendering the vaccine ineffective.

Challenge

The commercial partner had a chimeric vaccine candidate requiring improved thermostability for practical deployment. Based on historical rational design campaigns for similar proteins, the partner estimated that achieving a 2.5 °C improvement would require approximately seven rounds of wet lab optimization—representing several months of calendar time and substantial experimental cost.

The single-property optimization (thermostability only) simplified the problem relative to multi-property campaigns, but the need for a specific magnitude of improvement (2.5 °C) within manufacturing and regulatory timelines created pressure to minimize the number of experimental rounds.

Approach

We performed a single round of Cradle optimization in zero-shot mode, generating 96 candidates. No prior sequence-function data was available beyond the template sequence itself. The optimization relied entirely on evotuning using the evolutionary context of the antigen.

The partner specified thermostability as the sole optimization objective, without additional constraints on other properties (e.g., immunogenicity, expression). This focus allowed Cradle to explore the full sequence space of stability-enhancing mutations without balancing trade-offs.

Results

The best candidate achieved a 2.50 °C increase in melting temperature in a single experimental round:

The 7× speed improvement translates directly to reduced development timelines and costs. Where rational design would have required seven sequential design-build-test cycles (with typical turnaround of 2–4 weeks per cycle, totaling 3.5–7 months), Cradle delivered the target improvement in a single cycle (2–4 weeks).

The distribution of thermostability values across the 96-candidate library demonstrated that Cradle did not simply generate one lucky hit but rather a population of candidates with improved stability. Multiple variants achieved ΔT_m improvements between +1.5 °C and +2.5 °C, providing the partner with backup options should the best candidate encounter issues in downstream development.

What this means

This result demonstrates Cradle's effectiveness for accelerating time-sensitive development programs where every round of wet lab experimentation represents schedule risk. For vaccine development targeting infectious diseases with epidemic potential, reducing timelines from months to weeks can have substantial public health impact.

The 7× speed improvement metric, estimated by the partner based on their historical rational design experience, provides a concrete benchmark for Cradle's efficiency gains in single-property optimization. While multi-property campaigns showcase the system's capability to navigate complex fitness landscapes, this simpler case validates that even straightforward optimization problems benefit from machine learning guidance.

The success in zero-shot mode (no sequence-function training data) suggests that pre-trained protein language models have internalized sufficient knowledge about thermostability determinants to enable substantial improvements without iterative feedback. This is particularly valuable for early-stage programs where no historical data exists.

Methods note

Cradle performed evotuning on the evolutionary alignment of the chimeric antigen, using a pretrained transformer model using an MLM loss as the foundation. Generation proceeded via masked sequence infilling with diversity-aware ranking to ensure broad sampling of predicted high-stability sequence space. Thermostability was measured via differential scanning fluorimetry (DSF). No immunogenicity testing or other functional assays were performed for this single-property campaign. Full details remain confidential to the commercial partner.