genesisml/zPocket-human-proteome

zPocket human proteome scan

Ligand-binding pocket predictions over the reviewed human proteome, plus cavity proposals from three geometric detectors scored by the same head.

Two independent datasets:

directory	contents	size
zpocket_scan/	per-residue pocket logits, Cα coordinates, and decoded pockets from the zPocket pairformer2 head	326 MB
detector_proposals/	1,794,898 cavity proposals from fpocket / P2Rank / Lacuna over the same proteins, ranked and scored by the zPocket head, not by the detectors	47 MB

Everything is Parquet, so the dataset viewer works and load_dataset reads it directly.

---

zpocket_scan/

A 330k-parameter pairformer2 head reads out over a frozen co-folding trunk's pair representation, producing one logit per residue. Inference used 10 recycles, bf16, seed 101, MSA on and no templates, with a 2,560-residue length cap.

index.parquet — 20,092 rows, one per protein

column	meaning
uniprot_id	join key for every other file
sequence, sequence_length, sequence_sha256	the exact sequence folded; the hash is the identity key
head_split	train / validation / test / unseen — see "Leakage" below
plddt_0_100, ptm	whole-chain confidence of the predicted structure (pLDDT on a 0–100 scale)
n_noncanonical, noncanonical_positions	residues the tokenizer maps to UNK (BJOUXZ). A selenocysteine folded as a generic stub means the model never saw its chemistry; a null pocket call at such a site is not evidence.
model	the trunk family
n_samples	diffusion samples per seed

residue_scores.parquet — per-residue logits

One row per protein: uniprot_id and logits, a list of float32 of length sequence_length. These are logits, not probabilities — apply a sigmoid:

import pandas as pd, numpy as np
s = pd.read_parquet("zpocket_scan/residue_scores.parquet").set_index("uniprot_id")
p = 1 / (1 + np.exp(-np.asarray(s.loc["P35222", "logits"], dtype=np.float64)))

ca_coords.parquet — Cα coordinates

One row per protein: uniprot_id and ca_coord, a flat list of float32 to be reshaped (-1, 3). Enough geometry to re-decode pockets without shipping structures, which is why the decoder settings below are provisional — the whole proteome re-decodes on CPU in minutes. Atom-level coordinates are not included; they are needed only to recompute burial, which is not among the released fields.

pockets.parquet — 21,646 decoded pockets over 17,527 proteins

Residues scoring ≥ 0.5 are clustered by single linkage on Cα within 8 Å, keeping clusters of at least three — recorded in the decoder column as linkage(t=0.5,d=8.0,m=3). positions are 1-based canonical UniProt positions; mean_score is the mean probability over the lining; center_* is the pocket centroid in the fold's frame.

2,565 proteins decode no pocket at all. That is a real result rather than a failure — nothing clears the threshold in a group of three. Re-decode lower to see sub-threshold signal, and treat what appears as a hypothesis rather than a prediction.

Re-decoding

The decoder settings are a starting point, not a tuned optimum. They were chosen as the best detector-free result measured at the time; nothing upstream has to re-run to change them.

# cluster residues above a threshold by single linkage on CA within 8 A
import numpy as np
from scipy.cluster.hierarchy import linkage, fcluster
from scipy.spatial.distance import pdist

import pandas as pd
acc, thr = "P35222", 0.30
s  = pd.read_parquet("zpocket_scan/residue_scores.parquet").set_index("uniprot_id")
c  = pd.read_parquet("zpocket_scan/ca_coords.parquet").set_index("uniprot_id")
p  = 1 / (1 + np.exp(-np.asarray(s.loc[acc, "logits"], dtype=np.float64)))
ca = np.asarray(c.loc[acc, "ca_coord"], dtype=np.float32).reshape(-1, 3)
idx = np.where(p >= thr)[0]
lab = fcluster(linkage(pdist(ca[idx]), method="single"), t=8.0, criterion="distance")
pockets = [sorted(int(i) + 1 for i in idx[lab == c])
           for c in np.unique(lab) if (lab == c).sum() >= 3]

---

detector_proposals/

proposals.parquet — 1,794,898 rows.

This is not a baseline and must not be used as one. fpocket, P2Rank and Lacuna supplied the proposals — the residue sets are genuinely theirs — but every proposal was then re-ranked and re-scored by the zPocket head, and each detector's own score was discarded in the process: fpocket's druggability, P2Rank's probability and Lacuna's rank are not in this file and cannot be recovered from it. Comparing head_score across detectors measures the head's opinion of three different candidate generators. It says nothing about how those detectors rank their own output, and any method-vs-method claim drawn from this file would be wrong.

The honest description is: commodity detector proposals, ranked by our head. To benchmark the detectors, re-run them and use their native scores.

column	meaning
uniprot_id	join key
detector	which tool generated the proposal: fpocket, p2rank or lacuna
head_rank	rank by the zPocket head over that detector's proposals for that protein — not the detector's own rank
head_score	the head's aggregate over the proposal's lining (topk_mean, k=8), on a probability scale
n_residues, positions	the proposal's lining, 1-based canonical UniProt positions
zpocket_found_pocket	whether the head's own decoder emitted any pocket for this protein

What the file is good for: a large set of candidate pocket lining sets over a whole proteome from three independent generators, and a record of how one head scores them. On one worked example a site the head ranked 1st was fpocket's own rank 39 of 40 — which is the gap between the two orderings, and the reason the native order cannot be inferred from head_rank.

2,564 Lacuna proposals with zero residues were dropped.

Known biases, measured

• Lacuna proposes large regions. Its proposals that contain a zPocket pocket are a median 3.0x larger than that pocket. Treat its size as a prior, not a boundary.
• All three place proposals in disordered regions, where a pocket is not meaningful. Filter on plddt_0_100 before using them as ground truth.
• Adding these proposals surfaced no new sites. Of the proposals clearing the head's threshold on proteins that already had a zPocket pocket, every one overlapped a pocket the head had already proposed. The proposal stage is not what limits this pipeline.

---

Leakage

head_split records each protein's membership in the head's own training split. 1,570 proteins are train, 276 validation, 315 test, 17,931 unseen.

Report anything you measure stratified by this column, and lead with unseen. A figure pooled across splits is not a generalisation result. Note also that the split is a property of the protein, while leakage is really a property of the site: a pocket type seen on a homologue is leaked even when this protein is unseen.

Reproducibility

These predictions come from a frozen co-folding trunk. The publicly released zPocket code uses a different trunk, so running it will not reproduce these numbers exactly; the released code reproduces the published benchmark results, not this scan. No trunk weights or trunk artifacts are included or redistributed here — this release contains predictions only.

Caveats worth reading before use

• Long shallow grooves score poorly. Armadillo and similar superhelical surfaces are the geometry every method here handles worst, the head included.
• A null call is not evidence of absence, particularly on proteins with UNK-tokenized residues (n_noncanonical > 0) or low plddt_0_100.
• Pocket boundaries are approximate. The decoder emits residue sets, not cavity volumes, so residue counts are not commensurable with a geometric detector's.

Provenance

Sequences: UniProt reviewed human proteome UP000005640, release 2026_02. Structures: predicted, not experimental. Detector versions: fpocket 4.2.3, P2Rank 2.5.1 (alphafold config), Lacuna 1.2.0.

Citation

Pre-print: https://arxiv.org/abs/2610.17696