embed

The biocentral embed module allows to run protein language models and other baseline embedders. These contain meaningful representations for protein sequences - the ideal basis for downstream analysis!

from biocentral import Biocentral, CommonEmbedder

biocentral = Biocentral()

embedder_name = CommonEmbedder.ONE_HOT_ENCODING
sequence_data = {
    "Seq1": "MMALSLALMM",
    "Seq2": "PRTEINMMALM",
    "Seq3": "PRTSSSLAMAM",
    "Seq4": "SEQWENCEAWMMWW",
}
result = biocentral.embed(embedder_name=embedder_name, sequence_data=sequence_data)
print(result)

biocentral embed sequences.fasta

Common Embedder Models

The following embedder models and baselines are primarily supported in biocentral. You can still provide any huggingface model string as an input to the API as well, as long as it is supported by biotrainer.

Protein Language Models

• ProtT5 (Rostlab/prot_t5_xl_uniref50): T5-based model pre-trained on UniRef50, excelling at capturing long-range dependencies in protein sequences.
• ProstT5 (Rostlab/ProstT5): Structure-aware T5 model that incorporates 3D structural information alongside sequence data.
• ESM2-3B (facebook/esm2_t36_3B_UR50D): Large-scale evolutionary scale modeling transformer with 3 billion parameters, trained on UniRef50.
• ESM2-650M (facebook/esm2_t33_650M_UR50D): Medium-sized ESM2 model with 650 million parameters, balancing performance and computational efficiency.
• ESM2-8M (facebook/esm2_t6_8M_UR50D): Compact ESM2 model with 8 million parameters, ideal for quick sanity checks and fast analysis.

Baseline Embedders

• One-Hot Encoding (one_hot_encoding): Simple binary encoding representing each amino acid as a unique vector.
• Random Embedder (random_embedder): Generates random embeddings, useful as a baseline for comparison.
• AAOntology (AAOntology): Embeddings based on amino acid properties and biochemical characteristics.
• BLOSUM62 (blosum62): Substitution matrix-based embeddings capturing evolutionary amino acid relationships.
• Length Embedder (length_embedder): Encodes sequence information based on protein length.