FANTASIA V3 - LookUp Table - UniProt July 2025 - Experimental Evidence code

📘 FANTASIA V3 - LookUp Table (UniProt July 2025)

Experimental Evidence Code Only

Overview

This is a PostgreSQL database backup using the pgvector extension to store high-dimensional protein embeddings. It contains precomputed embeddings and functional annotations from the UniProt July 2025 release, including only entries supported by experimental evidence.

This lookup table was generated using version v2.0.0 of the Protein Information System (PIS), an integrated biological information system designed for the automated extraction, processing, and management of protein-related data. PIS consolidates information from UniProt, PDB, and GOA, allowing efficient retrieval and organization of sequences, structures, and annotations.

The resulting database is designed for compatibility with FANTASIA V3, an advanced pipeline for large-scale functional annotation of proteins using state-of-the-art Protein Language Models (PLMs). While the lookup table is stored in a vector database for persistence, FANTASIA loads the relevant data into memory at runtime to enable high-speed annotation.

FANTASIA uses precomputed deep learning embeddings to perform nearest-neighbor searches in embedding space and transfer Gene Ontology (GO) terms from experimentally annotated proteins to query sequences.

Dataset Details

• Total proteins: 127,546

• Total sequences: 124,397

• Total embeddings: 621,849

• Total GO annotations: 627,932

• Included evidence codes (Gene Ontology, experimental only):

  • EXP – Inferred from Experiment

  • IDA – Inferred from Direct Assay

  • IPI – Inferred from Physical Interaction

  • IMP – Inferred from Mutant Phenotype

  • IGI – Inferred from Genetic Interaction

  • IEP – Inferred from Expression Pattern

  • TAS – Traceable Author Statement

  • IC – Inferred by Curator

Included Embedding Models

• ESM-2 (650M parameters)
  A transformer-based protein language model trained on UniRef50 using masked language modeling. It captures structural and functional features directly from raw sequences without requiring MSAs. ESM-2 is widely used for contact map prediction, unsupervised learning, and representation extraction.

• ProtT5-XL-UniRef50 (~1.2B parameters)
  A large-scale encoder-decoder model using the T5 architecture, trained on UniRef50 via masked span prediction. It generates high-dimensional sequence representations that perform well across structure and function prediction tasks.

• ProstT5 (~1.2B parameters)
  A multi-modal extension of ProtT5, trained to predict both sequence and coarse-grained 3Di structural states. Useful for downstream applications like contact prediction, functional annotation, and classification.

• Ankh3-Large (620M parameters)
  An encoder-only T5-style model trained with masked span prediction. Optimized for fast inference, it encodes both semantic and structural protein information and can replace ProtT5 in many ML pipelines.

• ESM3c (Cambrian 600M)
  Part of the new ESM C model family, trained on UniRef, MGnify, and JGI datasets. With rotary embeddings and 36 layers, it offers enhanced performance for masked language modeling, producing high-quality structural and functional embeddings without alignments.

Missing Proteins

A small subset of proteins could not be processed on the Finisterrae III (CESGA) supercomputer due to memory limitations with 40 GB A100 GPUs.

The file missing_proteins.csv lists all affected UniProt identifiers. These entries are excluded from the final lookup table.