ProtGPT2

ProtGPT2 generates de novo protein sequences using an unsupervised Transformer-based language model trained on large corpora of natural protein sequences to explore and expand protein sequence and structural space for protein design.

Key Features:

• Transformer-Based Architecture: ProtGPT2 uses an unsupervised Transformer-based language model adapted to protein sequence generation.
• Natural Sequence Propensities: The model generates sequences that exhibit natural amino acid propensities consistent with biologically plausible proteins.
• Globular Protein Generation: Disorder predictions indicate that 88% of ProtGPT2-generated proteins are predicted to be globular.
• Exploration of Uncharted Protein Space: Sensitive sequence searches and similarity-network analyses show ProtGPT2 sequences are distantly related to natural sequences, enabling exploration of previously uncharted regions of protein space.
• Structural Prediction and Novelty: AlphaFold predictions of ProtGPT2 sequences yield well-folded structures with features such as embodiments and large loops and reveal topologies not currently represented in existing structure databases.
• Computational Efficiency: The model generates protein sequences within seconds.

Scientific Applications:

• Drug discovery: Generation of novel protein sequences to support identification and design of therapeutic candidates.
• Enzyme engineering: Design of de novo enzymes with tailored sequence properties for catalytic applications.
• Synthetic biology: Creation of customized protein sequences for incorporation into synthetic biological systems.
• Environmental and biomedical applications: Exploration of novel protein structures and functions relevant to environmental biotechnology and biomedical challenges.

Methodology:

Unsupervised training of a Transformer-based language model on large protein-sequence datasets; disorder prediction for globularity assessment; sensitive sequence searches and similarity-network analyses for relationship to natural sequences; structural prediction using AlphaFold.