A Frontier-Scale Protein Language Foundation Model for Conditional De Novo Enzyme Design

CureCraft proposes to train a 273B-parameter industrial-enzyme protein language model (PLM) on a ~350B-sequence corpus using 600k GPU-hours on MareNostrum (12 months) to enable de novo enzyme generation and conditional “feature installment” (property-specific manipulation) without task-specific fine-tuning. Enzyme engineering is fundamentally constrained by sparse and inconsistent functional labels across natural catalogues and by the impracticality of human-led, multi-objective optimisation across coupled properties (activity, stability, folding/solubility, expression, and operating windows such as pH and salt). 

Our project addresses this bottleneck by combining frontier-scale pretraining with instruction fine-tuning and in-context learning to make the model controllable by explicit design specifications, and reinforcement-learning (RL) post-training to improve developability (solubility, expression, and folding/structural viability) alongside catalytic performance. The approach is de-risked by CureCraft’s prior wet-lab validation of a highly novel IsPETase-class depolymeriser with ~50% sequence identity to wild-type while matching activity (enabled by our previous EuroHPC allocation), demonstrating true novelty-to-function generalisation rather than near-neighbour memorisation. A closed-loop wet-lab–model active-learning workflow will further anchor improvements to experimental readouts and continuously raise hit-rate. 

The outcome will be a European industrial enzyme foundation model and engineering pipeline capable of accelerating innovation across sustainable materials, green chemistry, biomanufacturing and food systems—illustrated by ongoing efforts spanning PETase validation, nylonase development, and horizontal expansion opportunities (e.g., polyurethanase-like deconstruction). To amplify ecosystem impact, the project plans to release an open ~10B-parameter model with responsible safeguards, enabling European researchers and SMEs to build downstream tools and applications on a high-performance public foundation.