Gaia

A protein of unknown function from Pseudomonas vancouverensis LMG 20222 returns nothing but "hypothetical protein" from a BLASTp search of NCBI nr, and a structural search of its predicted monomer returns kinesin, an RNA polymerase subunit and a snail agglutinin — hits that say nothing about what it does. Its genomic neighborhood settles the question at a glance: capsid, baseplate and spike genes on an integrated prophage. Gaia is built on the premise that such context is itself searchable, and that one query should retrieve proteins in similar genomic company, not only proteins with similar residues or folds.

Gaia (Genomic AI Annotator) is a sequence search and annotation platform from Tatta Bio, with collaborators at the DOE Joint Genome Institute, described in a November 2024 bioRxiv preprint. Its retrieval engine is gLM2_embed, a fine-tuned variant of the gLM2 mixed-modality genomic language model. Because gLM2 is pretrained on whole metagenomic contigs — coding sequences tokenized as amino acids, intergenic regions as nucleotides — its per-protein representation already carries information about the genes around it. Gaia distills that representation into a compact vector optimized for nearest-neighbor search, then runs it against a precomputed index of microbial protein space.

The result complements rather than replaces alignment: Gaia sorts candidates by combined sequence, structure and context similarity, which makes conserved multi-gene systems such as defense islands and biosynthetic gene clusters findable from a single member. It is a sibling of FlashPPI, the same group's gLM2-derived interaction predictor, and both are served through the SeqHub platform that succeeded the original Gaia web app.

#Key Features

• Context-aware retrieval: Query embeddings inherit the genomic-neighborhood signal learned during gLM2 pretraining, so proteins in conserved gene arrangements retrieve one another even when pairwise identity is low.
• Structure-aligned contrastive fine-tune: A projection layer trained against AlphaFold Database structural clusters pulls the embedding space toward fold similarity without requiring any structure at query time.
• Compact search vectors: The projection outputs 512 dimensions against gLM2's 1280-wide hidden state, which is what makes approximate nearest-neighbor search across tens of millions of proteins fast enough to be interactive.
• Interpretable result pages: Each hit arrives with its genome context diagram, co-occurring genes, Pfam HMM annotations, alignment statistics and a predicted structure, so a retrieval can be judged rather than merely trusted.
• Open, ungated checkpoint: The fine-tuned encoder is published under Apache-2.0 and runs zero-shot on new sequences with no per-dataset retraining.

#Technical Details

gLM2_embed is trained in two stages. The first fine-tunes gLM2-650M for one epoch on UniRef50 at learning rate 1e-4 with batch size 256, reducing the mismatch between a model pretrained on multi-gene contigs and a task that embeds single proteins. The second freezes that model and trains one linear projection on mean-pooled representations from the 16th layer, chosen over the final layer for better transfer, under an InfoNCE contrastive loss over 2.3M AlphaFold Database structural clusters — same-cluster sequence pairs, batch size 32,768, 30,000 steps. The released checkpoint holds 325.8M parameters, roughly half of gLM2-650M, since inference stops at the layer the projection reads.

The search index, OG_prot_90, contains 85,007,726 centroid sequences clustered at 90% identity with MMseqs2 from the 131,744 prokaryotic and viral genomes of the OpenGenome dataset, queried through a Qdrant vector database using HNSW approximate nearest neighbors under cosine similarity. On the paper's benchmarks, Gaia retrieves sequence-similar proteins (n=666) more sensitively than ESM2 embeddings while running one to three orders of magnitude faster than alignment, and tops ESM2 embeddings, BLASTp and MMseqs2 on genomic context retrieval (n=3,000). On SCOPe-40 2.01 remote homology (n=2,207) it trails ESM2 and Foldseek — the cost of representations that encode sequence and context alongside structure.

#Applications

Gaia is aimed at microbial and viral genomics, where many predicted proteins carry no usable annotation and the informative signal is the operon rather than the gene. In the paper's case studies it assigns a phage major tail fiber function to an uncharacterized prophage protein, corroborated by a diversity-generating retroelement among the most frequent co-occurring genes across retrieved contexts, and recovers putative siderophore loci in Pseudomonas flavescens and Nocardia amamiensis from a quinolobactin CoA transferase-ligase query — loci that rule-based detection misses for lack of a condensation-domain signature.

#Impact

Gaia demonstrates that genomic context can be made a first-class axis of similarity search, retrievable across tens of millions of sequences in the time budget of a BLASTp query. Its released encoder is reusable well beyond the hosted platform: any workflow needing context-aware protein embeddings can load it directly. The tradeoffs are explicit — structural sensitivity is lower than dedicated structure search, only the benchmarking code is open rather than the indexing and serving stack, and the preprint remains unreviewed. The hosted service has since been folded into SeqHub, but the checkpoint and the OG_prot_90 database remain independently available.