For researchers

Independent K-Fold benchmark record

Measurements from running the public complex-structure model K-Fold: confidence, runtime, and interpretation limits for five protein targets and one ternary complex.

K-Fold predicts structures of complexes made from proteins and other molecules. This record describes confidence, runtimes, and experimental-structure comparisons from running kaist-ai-bio/kfold 1.0.0 on 2026-10-22. The code and weights were checked under Apache-2.0 at that time. These limited inputs and hardware do not establish an overall model ranking.

What was measured

We installed the public PyPI package unchanged and predicted structures for five public single-chain protein sequences, using 3 seeds × 5 samples. Hardware was one NVIDIA RTX PRO 4000 Blackwell 24GB. We read the model's confidence outputs (pLDDT and pTM) without modification. The comparison is earlier ColabFold (AlphaFold2-family) measurements for the same sequences on the same server.

Results for five single-chain targets

Target (PDB)Residue countBest K-Fold pLDDTAF2 pLDDTK-Fold time per seed
ubiquitin (1ubq)7688.695.94–12 s
PD-L1 (5jds)24587.989.511–18 s
PD-1 (7acl)35583.586.920–25 s
TEV protease (1lvm)46777.084.834–36 s
Antibody Fab (8tfl)94084.386.494–97 s*

*The 940aa input ran out of memory with standard settings on the 24GB GPU. It completed with the developer-recommended cpu-offload setting, using approximately 70GB of system memory.

Ternary complex (PROTAC) — 8BDS

The query contained VCB + BRD4-BD1 + PROTAC(QIY): four chains, 490 residues, and one ligand. Using the original sequences from the experimental structure, all 3 seeds × 5 samples completed.

  • Best sample: pLDDT 92.2, ipTM 0.829 (ipTM 0.799–0.829 across 15 samples).
  • 39–43 seconds per seed (standard settings, one GPU, no out-of-memory failure).

How to interpret the results

  • Single-chain prediction is not K-Fold's claimed focus. It targets complex assembly and binding-induced structural changes. Single-chain numbers are observations, not a ranking of complex-prediction performance.
  • pLDDT, pTM, and ipTM are model confidence measures, not accuracy against an experimental structure or binding affinity.
  • The ternary-complex task reassembles components using the original protein sequences from the experimental structure. Reproduction of that structure requires a separate comparison, such as DockQ.

DockQ structure comparison — added 2026-10-22

We compared the best predicted ternary-complex sample with experimental structure 8BDS using DockQ (MIT, wallnerlab).

InterfaceDockQiRMSD(Å)fnat
ElonginB–ElonginC (A–B)0.9350.450.93
ElonginB–VHL (A–C)0.9590.511.00
ElonginC–VHL (B–C)0.8920.510.84
VHL–BRD4 (C–D, PROTAC interface)0.5851.490.60
Mean across all four0.843——
  • The PROTAC-induced VHL–BRD4 interface scored DockQ 0.585, iRMSD 1.49Å, and fnat 0.60, showing local contact reproduction. The global relative arrangement still differed by LRMSD 6.18Å. This comparison alone cannot attribute the difference to molecular motion or crystallization conditions.
  • Separately from model confidence, DockQ shows how similar this prediction is to the reference structure for this input. It does not validate actual binding or function.
  • Runtimes apply to our hardware. Measurements on other devices, such as L40S, use different conditions and cannot be compared directly.
  • Across the best samples from three seeds, the pLDDT standard deviation was about 0.2–0.3. This describes only the three measured seeds; future runs are not guaranteed to stay within that range.

Reproduction

The query YAML (public PDB sequences), seed and sample settings, commands, and summary CSV are retained in the evidence folder. See the tool reference for kfold inputs and outputs (summary CSV and structure files). Current execution options and plan review determine whether a project can submit.

On this page