Skip to content

EnzGFM

Summary

Wang, Li, Geng, et al. (2026, in press at time of ingestion) present EnzGFM, an enzyme-specific protein language model (PLM) built on a Mamba-Transformer hybrid architecture with hierarchical pre-training, addressing the computational inefficiency and lack of enzyme-domain adaptation of general-purpose PLMs (e.g., ESM, ProtTrans) when applied to catalytic property prediction. EnzGFM achieves 2–5-fold inference acceleration over Transformer-based PLMs while improving accuracy across four enzyme-specific benchmarks; a companion agentic system, EnzGFM-Agent, translates its predictions into actionable enzyme-engineering workflows and was validated experimentally on secreted alkaline phosphatase (SEAP) mutagenesis.

Method

  • Architecture: a hybrid combining Mamba structured state-space model (SSM) layers, which process long-range sequence dependencies with linear-time complexity O(n), with Transformer self-attention layers for local catalytic-motif learning.
  • Hierarchical pre-training: stage 1 pre-trains on 48.28 million general protein sequences (16.67 billion tokens) from UniRef50 via masked language modeling; stage 2 continues pre-training on 5.09 million enzyme-specific sequences (1.96 billion tokens) curated from CAZy, BRENDA, and SABIO-RK, spanning all seven EC classes and sequence lengths from 50–5,000 amino acids.
  • Scaling: evaluated from 30M to 1.5B parameters; training loss decreased from 2.0881 (EnzGFM-30M) to 0.9230 (EnzGFM-1.5B), with perplexity following a power-law relationship with model size (scaling exponents α=0.3682 general, α=0.2935 after enzyme-specific pre-training). Enzyme-specific pre-training reduced perplexity by 5–35% at matched parameter counts relative to general-only pre-training.
  • EnzGFM-Agent: a multi-agent system built on EnzGFM that provides automated code generation, analysis, and decision support for enzyme mutagenesis, retrieval, and discovery through an interactive interface.

Performance

  • Benchmark improvements over Transformer-based PLMs: 16.67% relative improvement in kinetic parameter prediction, 15.69% in enzyme-reaction mapping, 13.19% in EC number classification, and 20.04% in mutation effect assessment, alongside 2–5-fold inference acceleration.
  • Experimental validation (EnzGFM-Agent): in a secreted alkaline phosphatase (SEAP) mutagenesis workflow, all EnzGFM-Agent-generated single-point mutants showed statistically significant activity improvements over the original sequence, with the best variant reaching a 7.13-fold increase in activity.

Availability

Code: github.com/DeepBxM/EnzGFM. Data and model weights: Zenodo 10.5281/zenodo.19412283.

See Also

Citations

[1] Wang, C., Li, M., Geng, S., Li, W., Zhou, X., Wang, Y.G., Yu, Y., Wang, T., & Shen, Y. (2026). An enzyme-specific protein language model for catalytic property prediction. Nature Communications (in press).