ARC: Assessment of Interface Residue Conformation

ARC is a single-model quality assessment method for protein complexes. It uses an ensemble of graph neural networks to assign per-residue scores on modeled chain-chain interfaces, with emphasis on interface-residue identification and local interface quality.

This repository contains the inference code, CASP16 evaluation pipeline, packaged asset workflow, and manuscript figure and table generation used for the ARC study.

---

Quick Start

The minimal workflow to reproduce ARC results:

# 1. Install environment
pixi install
pixi run smoke

# 2. Fetch packaged data and predictions
pixi run python -u scripts/assets/fetch_assets.py

# 3. Run evaluation (uses packaged predictions)
pixi run eval-full

To run inference instead of using packaged predictions:

pixi run run-batch
pixi run eval-full

---

Output Overview

ARC produces two complementary outputs per model:

• LOCAL.json: per-residue interface quality scores, used for residue-level evaluation and interface discrimination.
• QSCORE.json: global interface quality scores, used for model ranking and CASP-style evaluation.

The ARC/ directory contains the ensemble prediction used in the manuscript, while ARC_<suffix>/ directories correspond to individual GNN components.

---

Installation

Step 1

Clone the ARC repository and enter it (all commands assume you are at the repository root):

git clone https://github.com/zwang-bioinformatics/ARC.git
cd ARC

Step 2

Create the ARC environment with pixi (reads pixi.toml in this directory):

pixi install
pixi run smoke

If pixi is not on your PATH, install it from https://pixi.sh/latest/installation/ or use ~/.pixi/bin/pixi.

Important

All commands below are meant to be run from the repository root.

---

Data Setup

Step 3

Populate the repository data layout.

You have two options:

1. Recommended (reproducible): download packaged assets
2. Manual: prepare inputs following the required directory structure

Option 1: Fetch packaged assets

The recommended approach is to fetch the packaged tarballs:

pixi run python -u scripts/assets/fetch_assets.py

This extracts each archive (Zstandard .tar.zst) at the repository root:

arc_eval_inputs_core.tar.zst:
data/
├── raw_16/
├── casp16_ema_reference_results/
├── casp16_targets/
├── casp_model_scores.csv
├── casp16_approx_target_sizes.json
├── ema_local_scores_with_lddt_added_mdl_contacts.csv
└── target_margin_scores.csv

arc_graph_data_casp16.tar.zst:
data/
└── CASP16/
    └── <target>/
        └── <model>/
            ├── meta.json
            └── data.st

predictions_apollo.tar.zst:
outputs/
└── predictions/
    └── CASP16/
        └── <target>/
            └── APOLLO/
                └── LOCAL.json

predictions_arc.tar.zst:
outputs/
└── predictions/
    └── CASP16/
        └── <target>/
            ├── ARC/
            │   ├── LOCAL.json
            │   └── QSCORE.json
            └── ARC_<suffix>/
                ├── LOCAL.json
                └── QSCORE.json

The APOLLO predictions are included as a baseline for comparison during evaluation.

In scripts/assets/fetch_assets.py, each archive is controlled by an ENABLE_* flag (all default to True). Set a flag to False to skip specific downloads.

---

Option 2: Manual data preparation

If you prepare inputs manually, mirror the following layout:

data/
├── CASP16/
│   └── <target>/
│       └── <model>/
│           ├── meta.json
│           └── data.st
├── raw_16/
├── casp16_ema_reference_results/
│   └── <model>_<target>.json
├── casp16_targets/
│   └── <target>.pdb
├── casp_model_scores.csv
├── casp16_approx_target_sizes.json
├── ema_local_scores_with_lddt_added_mdl_contacts.csv
└── target_margin_scores.csv

For predictions, either use the packaged layouts above or run inference to populate:

outputs/predictions/CASP16/

---

Usage

Note

Evaluation requires predictions in outputs/predictions/.

• If you fetched packaged predictions, you can run evaluation directly.
• If not, you must run inference before evaluation.

---

Inference

Single target:

pixi run predict -- -t H1202

Batch over all targets:

pixi run run-batch

Expected input:

data/CASP16/<target>/<model>/
├── meta.json
└── data.st

Outputs:

outputs/predictions/CASP16/<target>/
├── ARC/
│   ├── LOCAL.json
│   └── QSCORE.json
├── ARC_GENConv/
├── ARC_GINEConv/
├── ARC_GLFP/
├── ARC_GeneralConv/
├── ARC_PDNConv/
├── ARC_ResGatedGraphConv/
└── ARC_TransConv/

The ARC/ directory contains the ensemble prediction used in the manuscript. The ARC_<suffix>/ directories correspond to individual GNN components.

---

Evaluation

Run full CASP16 evaluation and manuscript outputs:

pixi run eval-full

---

Manuscript Outputs

Generate figures:

pixi run manuscript-figures

Generate tables:

pixi run manuscript-tables

---

Outputs

outputs/
├── predictions/
│   └── CASP16/<target>/...
└── results/
    ├── eval/
    ├── logs/
    └── local_results/
        ├── arc_ensemble/
        ├── manuscript_figures/
        ├── manuscript_tables/
        ├── per_target_analysis/
        └── pooled_analysis/

---

Citation

If you use ARC, please cite: [TBD]

@article{shrestha2026arc,
  title={ARC: Assessment of Interface Residue Conformation using an Ensemble of Graph Neural Networks},
  author={Shrestha, Bishal and Siciliano, Andrew J. and Huang, Gabriel and Bao, Yifan and Wang, Zheng},
  journal={Proteins: Structure, Function, and Bioinformatics},
  year={2026},
  publisher={Wiley Online Library},
  status={Submitted}
}