nshuaib01/ESM-2_Data

Soda Lake CAZyme Discovery — Data and Model Artefacts

Supporting data and trained models for Uncovering Novel CAZymes in Soda Lake Metagenomes using a Protein Language Model (MSc Data Science, Bioinformatics and Biological Modelling, Durham University).

---

Contents

File	Description	Rows
cazyme_counts_with_sodalake_metadata.csv	Per-MAG CAZyme counts joined to soda lake sample metadata. Source table for the taxonomic and CAZyme-family analyses.	36697
soda_lake_cazymes_metadata.csv	dbCAN3-annotated CAZymes with family, class, substrate and source MAG.	78650
all_candidates.csv	The 1,407 classifier-identified candidate proteins (score ≥ 0.5, no dbCAN hit, absent from training data), with CD-Search and downstream validation results.	1,407
Metagenomic_Dataset.csv	Labelled training set used to fit the classifiers: dbCAN-positive (≥ 2 tools) and downsampled negative proteins, with CD-HIT cluster assignments.	2987336
metrics.json	Cross-validation and family-holdout performance metrics.	—
model.joblib	Trained MLP classifier (scikit-learn).	—
scaler.joblib	Feature scaler fitted on training folds only (see Leakage control).	—
'Features.csv'	Physicochemical Features from Protr and Peptides	6571

Generating embeddings

The models take 640-dimensional mean-pooled ESM-2 embeddings as input. New sequences must be embedded identically or scores will not be comparable.

• Model: esm2_t30_150M_UR50D (30 layers, 150M parameters, 640-dim representations)
• Representation layer: <30>
• Maximum input length: 1022 amino acids
• Pooling: mean across all residue positions, giving one 640-dim vector per protein

import torch, esm, joblib, numpy as np

model, alphabet = esm.pretrained.esm2_t30_150M_UR50D()
batch_converter = alphabet.get_batch_converter()
model.eval()

LAYER = 30

def embed(seqs):
    """seqs: list of (label, sequence) tuples. Returns (n, 640) array."""
    _, _, toks = batch_converter(seqs)
    with torch.no_grad():
        out = model(toks, repr_layers=[LAYER])
    reps = out["representations"][LAYER]
    # mean-pool over real residues only, excluding BOS/EOS
    vecs = []
    for i, (_, seq) in enumerate(seqs):
        vecs.append(reps[i, 1:len(seq) + 1].mean(0).numpy())
    return np.vstack(vecs)

scaler = joblib.load("scaler.joblib")
model_clf = joblib.load("model.joblib")

X = embed(my_sequences)        
X = scaler.transform(X)         
p = model_clf.predict_proba(X)[:, 1]
calls = p >= 0.5

Threshold. A fixed decision threshold of 0.5 was used throughout. Scores are the probability that a protein resembles a dbCAN-annotated CAZyme in embedding space, not the probability that it is a CAZyme.