ESM Atlas - Metagenomic Protein Universe

50,000 proteinsbiologyproteinsstructural3D

A TMAP of 50,000 metagenomic proteins sampled from the ESM Atlas. Each protein is embedded with ESMC-600M (1,152-d) and folded with ESMFold2, and the viewer offers two layouts of the same set: Map A from the raw ESMC embeddings, and Map B from sparse-autoencoder features (16,384 interpretable directions, k64 codebook, pooled per protein). Because TMAP connects each point to its true nearest neighbor in the full embedding space, the tree edges are meaningful - neighboring proteins are genuinely similar. Click any point to load its ESMFold2-predicted structure, then color, search, filter, or lasso-select from there. Models: ESMC-600M and ESMFold2-Fast (Biohub).

This demo is hosted externally

The dataset is too large to embed inline. Opens in a new tab.

Open Interactive Demo

How it was made

generate.pypython

from tmap import TMAP

# 1,152-d ESMC-600M embeddings -> Map A
# (16,384-d sparse-autoencoder features -> Map B)
X = esmc_embeddings  # shape (N, 1152)

viz = (
    TMAP(metric="cosine", n_neighbors=20, seed=42)
    .fit(X)
    .to_tmapviz()
)
viz.add_3d_structures(cif_urls, fmt="cif")  # ESMFold2 folds, click to view
viz.write_static("esm_atlas_out/")  # one self-contained page