SeqVec

SeqVec generates continuous vector embeddings of protein sequences using an ELMo language model to enable prediction of protein structure and function.

Key Features:

• Embedding Generation: Transforms amino acid sequences into continuous SeqVec (Sequence-to-Vector) embeddings using an ELMo language model.
• Predictive Performance: At the per-residue level, outperforms one-hot and Word2vec-like encodings for secondary structure prediction (Q3 = 79% ± 1, Q8 = 68% ± 1) and intrinsic disorder (MCC = 0.59 ± 0.03); at the per-protein level, predicts subcellular localization across ten classes (Q10 = 68% ± 1) and discriminates membrane-bound versus water-soluble proteins (Q2 = 87% ± 1).
• Speed and Scalability: Generates embeddings in approximately 0.03 seconds per protein compared to roughly 2 minutes for evolutionary information generation with HHblits, enabling high-throughput analysis.
• Transfer Learning Capabilities: Leverages large unlabeled sequence databases to extract information via transfer learning for diverse protein prediction tasks.
• Complementary to Evolutionary Information: Provides competitive performance and is applicable where evolutionary information is unavailable or impractical, though it does not surpass the best methods that use evolutionary profiles.

Scientific Applications:

• Protein structure and function prediction: Uses SeqVec embeddings as input features for predicting secondary structure and intrinsic disorder and for classifying subcellular localization and membrane association.
• Large-scale proteomic and metaproteome analyses: Enables rapid analysis of large proteomic datasets, including microbiome and metaproteome studies, by providing fast, scalable sequence representations.

Methodology:

Uses an ELMo language model trained on large unlabeled sequence data such as UniRef50 to produce continuous SeqVec embeddings from amino acid sequences and applies transfer learning from unlabeled sequence databases.