You provide a single-chain protein sequence — pasted directly, fetched by UniProt ID, or imported from a previous job. The parameter that most affects your result is model size: larger variants produce higher-quality embeddings at the cost of speed, so pick the variant that matches whether you value throughput or embedding fidelity.

Webapp input methods

The .bloom* webapp provides three input methods for ESM-3 protein embeddings.

Input type selection

Input Type	Description	Example
Protein Sequence	Direct amino acid sequence entry	`MKTAYIAKQRQISFVKSHFSRQDILDLWIYHTQGYFP...`
UniProt ID	Automatic sequence fetch from UniProt	`P68871`, `Q9Y261`
From Previous Job	Import from BLAST, ESMFold, or other tools	Select from job outputs

Protein sequence input

Format: Raw amino acid sequence using single-letter codes
Validation: Real-time checking for invalid characters
Length range: 10-2,048 residues (optimal: 50-1,000)
Case sensitivity: Automatically converted to uppercase
Multiline: Supported for longer sequences
FASTA format: Header lines starting with > are supported; multiple sequences allowed

UniProt ID input

Format: UniProt accession codes (e.g., P00533, P68871)
Auto-fetch: Automatic sequence retrieval from UniProt database
Validation: Real-time format checking
Isoforms: Selection of alternative splice variants when available
Metadata: Protein names, organism, and function annotations loaded automatically

Previous job integration

Compatible tools

BLAST: Sequence search results with hit sequences
ESMFold: Predicted structures (sequence extraction)
Sequence analysis: Processed or extracted sequences
Other ESM-3 runs: Re-embed with different model sizes or parameters

Data handling

Multi-sequence files: Automatic parsing and selection
Format conversion: Handling various sequence file formats
Provenance tracking: Maintains connection to original data source

Sequence requirements

ESM-3 accepts single-chain protein sequences containing standard amino acids.

Character validation

Standard amino acids only: ACDEFGHIKLMNPQRSTVWY
Excluded characters: B, J, O, U, X, Z are not supported
Special characters: Whitespace, numbers, and punctuation are automatically removed
Case normalization: Converted to uppercase for consistency

Length considerations

Minimum: 10 residues (shorter sequences lack sufficient context)
Optimal range: 50-1,000 residues for high-quality embeddings
Maximum: 2,048 tokens (model context window limit)
Long sequences: Sequences exceeding the context window are truncated with a warning

Batch processing

Multi-FASTA: Submit multiple sequences in a single FASTA file
Unique headers: Each sequence must have a unique FASTA header for result mapping
Batch size: Up to 100 sequences per batch job
Memory scaling: Total memory increases with batch size and maximum sequence length

Model parameters

Model size selection

Model	Parameters	Embedding Dim	Speed	Quality
ESM-3 Small	1.4B	1,536	Fast	Good
ESM-3 Medium	7B	2,560	Moderate	Better
ESM-3 Large	98B	6,144	Slow	Best

Choosing a model size

Small: Best for batch processing, initial exploration, and when speed matters more than embedding quality
Medium: Recommended default; good balance of quality and speed for most use cases
Large: Best for critical analyses where embedding quality is paramount; significantly slower

Embedding parameters

Parameter	Type	Default	Description
model_size	string	`esm3_sm`	Model variant (esm3_sm, esm3_md, esm3_lg)
task	string	`embeddings`	Analysis task: embeddings, structure_prediction, function_prediction, sequence_generation, or masked_prediction
return_attention	boolean	false	Return attention maps for contact prediction
pooling	string	`mean`	Pooling strategy for per-sequence embedding (mean, cls, none)

Output type details

Per-residue embeddings (L x D):

One embedding vector per amino acid position
Dimension D depends on model size (1,536 to 6,144)
Captures local structural and functional context at each position
Best for: residue-level property prediction, contact maps, variant effect estimation

Per-sequence embeddings (1 x D):

Single embedding vector representing the entire protein
Computed by pooling per-residue embeddings
Best for: protein-level classification, similarity search, clustering

Pooling strategies

Strategy	Description	Best For
mean (default)	Average across all residue positions	General-purpose protein comparison
cls	Use the CLS token embedding	Tasks where global context is critical
none	Return unpooled per-residue embeddings	Residue-level ML tasks; skips sequence-level pooling

Sequence preprocessing

Automatic steps

Whitespace and special character removal
Case normalization to uppercase
Length validation against model context window
Character validation against standard amino acid alphabet

Manual recommendations

Signal peptides: Consider removing signal peptides if you want embeddings of the mature protein only
Propeptides: Remove propeptide regions if not part of the functional protein
Domain extraction: For multi-domain proteins, consider embedding individual domains separately for higher-resolution analysis
Isoform selection: Choose the most biologically relevant isoform

ESM-3 inputs and parameters