Benchmark for Biomolecular Structure Prediction Models

‼️This document describes the dataset provided in the initial PXMeter release (the PXM-Legacy dataset). The evaluation results in this version were generated using the PXMeter code at commit 9830335c68d0c918f7776ad58946cfa5e7690e16.

---

This repository provides evaluation codes for assessing models using our curated evaluation sets:

Dataset	Description	Metrics
RecentPDB	Evaluates RecentPDB low homology subset. Antibody-antigen and monomer subsets reported separately.	DockQ success rates (>0.23), LDDT
AF3-AB	Analyses antibody-antigen subset of AlphaFold3.	DockQ success rates (>0.23), LDDT
dsDNA-Protein	Focuses on intra-DNA chains and DNA-Protein interfaces, aggregating LDDT metrics.	LDDT
RNA-Protein	Evaluates intra-DNA chains and RNA-protein interfaces with LDDT aggregation.	LDDT
PoseBusters	Assesses pocket-aligned RMSD of small molecules (PoseBusters V2).	RMSD success rates (< 2 Å)

💡 Usage

0. Download the dataset

The benchmark data is licensed under CC0.

Before running the code for the first time, download the necessary dataset files:

wget https://pxmeter.tos-cn-beijing.volces.com/evaluation_supported_data.tar.gz
tar xzvf evaluation_supported_data.tar.gz -C [your_path]

To download the full dataset used in our article, including model input files, inference output structures, evaluation output JSON files, and summaries:

wget https://pxmeter.tos-cn-beijing.volces.com/evaluation_full_data.tar.gz
tar xzvf evaluation_full_data.tar.gz -C [your_path]

You can place the extracted evaluation folder in the directory from which you run the benchmark code, or specify its location via an environment variable:

export PXM_EVAL_DATA_ROOT_PATH="your/path/to/evaluation"

1. Evaluate inference results

Run the evaluation script as follows:

python benchmark/run_eval.py -i [infer_dir] -o [output_dir] -d [dataset] -c [chunk_str] -m [model] -n [num_cpu]

• infer_dir: Directory containing inference results (CIF files and confidence files).
• output_dir: Directory where evaluation results will be saved.
• dataset: Dataset to evaluate (options: "RecentPDB", "PoseBusters", "dsDNA-Protein", "RNA-Protein", "AF3-AB").
• chunk_str: Chunk string for distributed evaluation (e.g., '1of5'), used when running evaluations across multiple machines. The default is None.
• model: Model name should be one of the supported options: "protenix", "boltz", "chai" or "af2m", as defined in "benchmark.run_eval.run_batch_eval", which calls corresponding Evaluators.
• num_cpu: Number of CPU cores to use. Default is 1.

2. Create JSON describing evaluation results paths

Example JSON structure:

{
  "name": {
    "model": "model name (protenix, af2m, chai, boltz)",
    "seeds": [101, 102, "..."],
    "dataset_path": {
      "RecentPDB": "path/to/eval_results/RecentPDB",
      "PoseBusters": "path/to/eval_results/PoseBusters",
      "AF3-AB": "path/to/eval_results/AF3_AB"
    }
  }
}

Allowed keys for dataset_path are: "RecentPDB", "PoseBusters", "dsDNA-Protein", "RNA-Protein", and "AF3-AB". Each dataset can include only a subset of these keys.

3. Aggregate and display evaluation results

Run the aggregation script: python benchmark/show_intersection_results.py -i [input_json] -o [output_path] -d [dataset_names] -n [num_cpu] -c [subset_csv] --overwrite_agg

Parameters:

• input_json: Path to JSON file with evaluation results (created in step 2).
• output_path (optional): Directory for final aggregated CSV output. Defaults to "./pxm_results".
• dataset_names (optional): Comma-separated dataset names to compare using intersections of "chain" and "interface". Defaults to all names in JSON.
• num_cpu (optional): Number of CPU cores for aggregation. Defaults to all available cores.
• overwrite_agg (optional): Overwrite existing aggregated CSV files. Defaults to False.
• subset_csv (optional): CSV file with columns ["type", "entry_id", "chain_id_1", "chain_id_2"]. Used to subset results. "type" can be "chain" or "interface".

Output structure

The script outputs summary CSV files in the directories specified by dataset_path in the JSON and creates a consolidated CSV with aggregated metrics in output_path:

pxm_results
├── DockQ_details.csv
├── DockQ_results.csv
├── LDDT_details.csv
├── LDDT_results.csv
├── RMSD_details.csv
├── RMSD_results.csv
├── Summary_table.csv
└── Summary_table.txt

• Summary_table.csv and Summary_table.txt provide a concise overview of key metrics such as DockQ success rate, PoseBusters success rate, and LDDT (with protein-protein interfaces indicated as prot_prot).
• *_results.csv files provide a full view of aggregated evaluation metrics.
• *_details.csv files allow exploration of sample-specific metrics selected by each ranker.

🔄 Reproduction of RecentPDB Low Homology Dataset Construction

This section outlines steps to construct the RecentPDB Low Homology dataset. File paths required for scripts are specified in benchmark.configs.data_config. It’s recommended to specify new output paths in scripts to prevent overwriting default files.

1. Filter to RecentPDB

Run this command to filter RecentPDB entries:

python benchmark/scripts/filter_for_recentpdb.py -c [mmcif_dir] -o [chain_interface_csv] -m [meta_csv] -n [num_cpu]

• mmcif_dir: Directory with MMCIF files downloaded from RCSB PDB.
• chain_interface_csv: Output CSV file recording filtered chain and interface information for RecentPDB dataset.
• meta_csv: Output CSV file containing meta information of filtered entries.
• num_cpu: Number of CPU cores for processing. Recommended CPU-to-memory (GB) ratio is 1:4.

2. Filter to Low Homology Subset

Run this command to filter to the Low Homology subset:

python benchmark/scripts/filter_to_lowh.py -c [chain_interface_csv] -t [test_to_train_json] -o [output_protein_lowh_csv] -n [output_nuc_lowh_csv]

• chain_interface_csv: Input CSV with filtered chain and interface info from RecentPDB.
• test_to_train_json: A JSON file containing a dictionary where keys are entities from the test set in the format '{PDB ID}_{Entity ID}', and values are lists of similar entities (high homology) from the train set in the same format.
• output_protein_lowh_csv: Output CSV with filtered chain and interface info for RecentPDB Low Homology Protein subset.
• output_nuc_lowh_csv: Output CSV with filtered chain and interface info for RecentPDB Low Homology Nucleic Acids subset.