KinConfBench: A curated benchmark for cofolding models on kinase conformational states

---

Overview

KinConfBench is a reproducible workflow for benchmarking protein–ligand cofolding on kinases. It wires together:

Inputs	MSAs and YAML configs for Boltz, Chai, and Protenix
Compute	Batched inference, optional Boltz affinity, structured output under predictions/
Labels	Kinase-centric, KinCoRe-style annotation CSVs
Paper metrics	Optional analysis package: success filters, diversity, apo/holo splits, plots

Everything assumes a shared directory layout (predictions/, analysis_csvs/, etc.) so steps compose cleanly from MSA generation through figures.

Released inference archive: Step-2 cofolding outputs (Boltz, Chai, Protenix) for the benchmark are on Figshare as 10.6084/m9.figshare.31986663 (download and unpack to match the kinconfbench_inference_data layout used by the analysis scripts).

---

Contents

• Installation
• Pipeline
• Inference outputs (Figshare)
• Quick start
• Repository layout
• License
• Acknowledgments
• Citation

---

Installation

Prerequisites

• conda or Mamba for the main kinconfbench environment.
• Separate conda environments for each cofolding backend (Boltz, Chai, Protenix) used in Steps 1–2. Paths are centralized in cofolding_inference/envs.py; set KINCONFBENCH_PREFIX / KINCONFBENCH_ENV if you use a prefix install.
• For kinase annotation (Step 3), install dependencies from env.yml (e.g. Biopython, HMMER, pandas, RDKit where needed).

Create the environment

From the directory that contains this KinConfBench folder (typical clone layout):

# Named env (recommended)
conda env create -f KinConfBench/env.yml -n kinconfbench
conda activate kinconfbench

# Or prefix install (set KINCONFBENCH_PREFIX for step_3; use KINCONFBENCH_ENV="" )
# conda env create -f KinConfBench/env.yml --prefix /path/to/envs/kinconfbench
# conda activate /path/to/envs/kinconfbench

Install the package (editable)

pip install -e KinConfBench

For prefix installs with step_3_kinase_cofolding_analysis.py, set KINCONFBENCH_PREFIX and KINCONFBENCH_ENV="" so generated scripts use conda run --prefix <path>.

---

Pipeline

flowchart LR
  A[1.0 MSAs] --> B[1.1 Ligands]
  B --> C[1.2 Config]
  C --> D[2 Inference]
  D --> E[3 Kinase labels]
  E --> F[Analysis optional]

Loading

Step	Script	Role
1.0	step_1_0_generate_all_msas.py	Build method-specific MSA layouts for Boltz / Chai / Protenix.
1.1	step_1_1_generate_ligand_pkl.py	Turn ligand tables (SMILES or CCD-style) into a merged ligands.pkl.
1.2	step_1_2_generate_config.py	Protein–ligand combinations and YAML inference config.
2	step_2_run_all_inferences.py	Run inference (--methods boltz chai protenix), optional affinity; organize CIFs via cofolding_inference/gather_structural_files.py.
3	step_3_kinase_cofolding_analysis.py	CIF preprocessing and KinCoRe-style CSVs per prediction set (wraps kinase_benchmark).

Downstream analysis (geometric filters, successful subsets, plots, pair metrics) lives under analysis/, driven by analysis/run_pipeline.py.

---

Inference outputs (Figshare)

Precomputed cofolding inference outputs for KinConfBench are archived on Figshare: https://doi.org/10.6084/m9.figshare.31986663. Use this to reproduce paper analyses without re-running GPU inference; extract the archive so paths align with your local kinconfbench_inference_data (or equivalent) root expected by Step 3 and analysis/.

---

Quick start (copy-paste)

# 1.0 — MSAs
python KinConfBench/step_1_0_generate_all_msas.py proteins.fasta \
  --output-dir tmp/msa_generation --final-dir tmp/msas

# 1.1 — Ligand dictionary (skip if you already have ligands.pkl)
python KinConfBench/step_1_1_generate_ligand_pkl.py data.csv -o ligands.pkl

# 1.2 — Config + combinations pickle
python KinConfBench/step_1_2_generate_config.py \
  --csv_file data.csv --proteins_path proteins.pkl --msa_folder tmp/msas \
  --ligands_path ligands.pkl --output_path combinations.pkl --config_output_path config.yml

# 2 — Inference (default: boltz, chai, protenix)
python KinConfBench/step_2_run_all_inferences.py config.yml \
  --output-folder inference_results --num-samples 5 --num-gpus 2 --predict-affinity

# 3 — Kinase analysis (kinconfbench env)
python KinConfBench/step_3_kinase_cofolding_analysis.py -d examples -v
bash examples/scripts/run_all_kinase_analyses.sh

If your shell cwd is already KinConfBench/, drop the KinConfBench/ prefix on script paths.

---

Repository layout

High-level map of what lives beside the step scripts:

KinConfBench/
├── step_1_0_generate_all_msas.py
├── step_1_1_generate_ligand_pkl.py
├── step_1_2_generate_config.py
├── step_2_run_all_inferences.py
├── step_3_kinase_cofolding_analysis.py
├── cofolding_inference/          # Backend adapters, env paths, gather CIFs
├── kinase_benchmark/
│   ├── kinase_curation/          # Historical benchmark construction (UniProt → PDB → filters)
│   └── kinase_pipeline/          # Runtime labeling (main_pipeline, KinCoRe-derived code)
├── analysis/                     # Paper-style metrics and plots
├── data/                         # Released tables / small examples
├── env.yml
└── setup.py

---

License

See LICENSE (UC Regents; research and not-for-profit use as described in the file).

---

Acknowledgments

• Mia A. Rosenfeld for contributions to project conception, experimental design, and narrative framing.
• Matthew Welborn, Aleksander Durumeric, and Michael Irvin for helpful suggestions.
• Cofolding backends: Boltz, Chai-1, Protenix.
• Kinase conformation labeling builds on ideas and code paths related to Kincore-standalone2 (see kinase_benchmark/kinase_pipeline/kincore_funcs/README.md).

---

Citation

If you use KinConfBench in your research, please cite the preprint (bioRxiv) and KinCoRe software suites.

@misc{kinconfbench,
	title = {{KinConfBench}: {A} {Curated} {Benchmark} for {Cofolding} {Models} on {Kinase} {Conformational} {States}},
	url = {https://www.biorxiv.org/content/10.64898/2026.04.07.716788v1},
	doi = {10.64898/2026.04.07.716788},
	language = {en},
	publisher = {bioRxiv},
	author = {Sun, Kunyang and Head-Gordon, Teresa},
	month = apr,
	year = {2026},
}

@misc{kincore2,
	title = {{AlphaFold2} models of the active form of all 437 catalytically competent human protein kinase domains},
	url = {http://biorxiv.org/lookup/doi/10.1101/2023.07.21.550125},
	doi = {10.1101/2023.07.21.550125},
	language = {en},
	author = {Faezov, Bulat and Dunbrack, Roland L.},
	month = jul,
	year = {2023},
}

@article{kincore,
	title = {Kincore: a web resource for structural classification of protein kinases and their inhibitors},
	volume = {50},
	copyright = {https://creativecommons.org/licenses/by/4.0/},
	issn = {0305-1048, 1362-4962},
	shorttitle = {Kincore},
	url = {https://academic.oup.com/nar/article/50/D1/D654/6395339},
	doi = {10.1093/nar/gkab920},
	language = {en},
	number = {D1},
	journal = {Nucleic Acids Research},
	author = {Modi, Vivek and Dunbrack, Roland L},
	month = jan,
	year = {2022},
	pages = {D654--D664},
}