Protein sequence representations

Starplast includes 320 ESM-2 features for 8,064 Toxoplasma genes. These are computed from the bundled coding sequences using the frozen 8-million-parameter ESM-2 model. They give prediction methods access to patterns in protein sequence without requiring a known function or localization label for each protein.

An ESM coordinate is a learned representation, not a named biological trait. It should enter a prediction model as part of a feature block. Its value alone does not mean that a gene has a particular function.

The model is facebook/esm2_t6_8M_UR50D, pinned to revision c731040fcd8d73dceaa04b0a8e6329b345b0f5df. Features are in starplast/data/esm_features.parquet; the accompanying manifest records the sequence hash, sequence length, model revision, and pooling recipe for each gene. No Starplast target labels were used to train or fine-tune this model.

Long proteins are split into windows of up to 1,000 residues, overlapping by 128. Each residue's representations are averaged across the windows that contain it, then pooled across the full protein. Every residue contributes a total weight of one. This preserves coverage, though a window cannot capture interactions with residues outside its context. These features are not a full-length structure model.

Invalid coding sequences, internal stop codons and conflicting translations are withheld. The 76 genes without a valid bundled translation retain missing sequence features; they do not receive a zero vector.

Regenerate or use the features

The bundled feature table needs only pandas to read. Generating new representations requires optional dependencies and downloads the pinned model:

pip install 'starplast[sequence]'
starplast-encode-proteins --output /path/to/esm_features.parquet --device cpu

Use --device cuda for a compatible NVIDIA installation. Encoding runs in evaluation mode with gradients disabled. Per-sequence caches under ~/.cache/starplast/esm allow interrupted runs to resume. A different sequence, model revision or pooling recipe produces a different cache key.

from pathlib import Path
import pandas as pd
import starplast

path = Path(starplast.__file__).parent / 'data'
features = pd.read_parquet(path / 'esm_features.parquet')

The small model keeps regeneration practical. Larger models or structure-derived representations are candidates for further benchmarks, not assumed improvements. See the ESM model card and official ESM repository.