An automated machine-learning framework for protein lead optimization — validated across antibodies, enzymes, peptides, vaccines, and gene-editing systems.
The Problem
Lead optimization costs $5–15M and takes up to 3 years per candidate
It's the longest and most expensive step in pre-clinical development, yet the design-build-test-learn cycle remains largely manual and heuristic. CRADLE-1 automates the design and learning steps, treating lead optimization as a high-dimensional search problem rather than a trial-and-error process.
How It Works
Foundation models, evotuned to your protein, fine-tuned on your data
CRADLE-1 adapts pre-trained protein language models to your project in stages: unsupervised evotuning on evolutionary neighborhoods, then supervised fine-tuning using both preference optimization (group DPO) and regression predictors trained on your assay data. A diversity-aware search algorithm generates candidate libraries optimized for both predicted function and sequence diversity. The first round can also run zero-shot, without any experimental data.
Multi-property Optimzation
Optimize 2–6+ properties simultaneously instead of one at a time
Binding, thermostability, expression, activity, aggregation, immunogenicity — CRADLE-1 handles them together and gives you visibility into how they interact. The system steers candidates along the Pareto frontier so you can make informed trade-offs, rather than fixing one property at the expense of another.
Results
Validated in vitro across 10 modality–target combinations
CRADLE-1 achieves ~90–95% success rate in hitting target product profiles across dozens of commercial campaigns, compared to ~85% for traditional lead optimization.
CRADLE-1 Whitepaper
The paper covers the complete technical framework — group DPO, diversity-aware ranking, batch-effect handling, ablations — plus full wet-lab protocols and a documented failure case.