kfold-benchmarks

Evaluation benchmarks for protein-complex representations. Mirror of a shared benchmark collection, uploaded so the same evaluations can run on a server that has no access to the original storage.

Every benchmark carries its own README.md, and that file is the contract. Read it before quoting a number: it states the task, where the labels came from, what the split controls for, and which trivial baseline has to be cleared. Protocols here have changed more than once, and a result measured under a superseded protocol is not comparable to one measured under the current one.

Contents

directory	size	task
qs_topology/	9.6 G	quaternary structure topology, 399 classes, Ahnert et al. 2015 definition
qs_topology_v2/	12.8 G	the revised build of the same task
real_vs_fake_complex/	7.8 G	is this two-chain complex assembled correctly, AUC
go/	734 M	Gene Ontology term prediction
ec/	269 M	Enzyme Commission number prediction
repsp_homodimer/	574 M	per-residue apo-to-holo labels on homodimers
binding_affinity/	6 K	README and baselines only, data is elsewhere
thermostability/	5 K	README and baselines only, data is elsewhere

Each follows a fixed layout: data/ for the records, splits/ for the assignment and whatever defines the grouping, baselines/ for measured baseline scores, build/ for what is needed to rebuild it.

Two things that are easy to get wrong

The floor depends on the probe. On real_vs_fake_complex the cheap-coordinate-descriptor floor is 0.647 under logistic regression, 0.712 under a linear probe and 0.750 under an MLP. A result must be compared against the floor measured with the same probe. Comparing an MLP-probed model against the logistic floor is the specific error the benchmark README calls out, and it is easy to make.

Probe linearly by default. The collection's convention is a linear probe on a frozen encoder, with the scaler fitted on train only. An MLP measures the head as much as the representation, and two encoders can swap places when the head does the work.

Reading the LMDBs

Records are LMDB directories. lock.mdb is a runtime lock and is deliberately not included.

import lmdb, pickle
env = lmdb.open("qs_topology/data/assemblies.lmdb", readonly=True, lock=False)
with env.begin() as txn:
    rec = pickle.loads(txn.get(b"1a02"))

Provenance

Labels are not experimental ground truth where they are computed. qs_topology reimplements a published definition and reproduces the authors' own table to 99.4%; that is the strongest claim available for it, and its README says so. Splits are homology-aware, and each README states the residual leakage rather than asserting it away.