AI-Designed Antigen Clears First-in-Human Safety Hurdle
University of Cambridge scientists used a generative protein-design model, trained on 120 million sequences, to create a stabilized pre-fusion SARS-CoV-2 spike fragment in 19 days. The resulting vaccine candidate, named CamVac-1, completed a 40-participant Phase I trial with zero serious adverse events and induced neutralizing titers 4-fold above convalescent levels.
The case demonstrates that an end-to-end AI protein-design loop can compress antigen discovery from years to weeks. Practitioners should add a generative-model stage to their biologics pipeline and validate candidates first in silico before committing to wet-lab synthesis.
The Cambridge Centre for AI in Medicine, working with GSK, moved CamVac-1 from sequence output to GMP-grade material in 11 weeks and secured a $4.2 million Wellcome Trust award for a 300-person Phase II expansion.
Step 1: Open the open-access Colab notebook at https://colab.research.google.com/github/cam-ai-design/vax-design and load the pre-trained ESM-2 650M weights. Step 2: Input the target viral protein FASTA and request 500 candidate sequences with temperature set to 0.6. Step 3: Export the top five designs to Rosetta for stability scoring; retain any candidate whose predicted ΔΔG is less than -1.5 kcal/mol for downstream expression.