Performance in OpenFold3 NIM

NIM Accuracy

OpenFold3 is an all-atom biomolecular complex structure prediction model from the OpenFold Consortium and the AlQuraishi Laboratory. OpenFold3 is a PyTorch implementation of the jax-based AlphaFold3 reported in Accurate structure prediction of biomolecular interactions with AlphaFold 3, and like AlphaFold3, OpenFold3 extends protein structure prediction capabilities to model complete biomolecular complexes including proteins, DNA, RNA, and small molecule ligands.

The OpenFold3 NIM’s accuracy should match that of the AlQuraishi Laboratory implementation of OpenFold3, when using equivalent parameters and inputs.

Note

Running on hardware that is not listed as supported in the prerequisites section may produce results that deviate from the expected accuracy.

The accuracy of the NIM is measured by structural quality metrics such as lddt (local distance difference test). These scores help assess the reliability of the predicted structures.

Factors Affecting NIM Performance

The performance of the OpenFold3 NIM is determined by several key factors:

Hardware Factors

• GPU type and memory: Different GPU architectures provide different performance levels

• System RAM: Larger proteins and complexes require more system memory

• Storage speed: Fast NVMe SSD storage improves model loading and caching performance

Input Complexity

• Sequence length: Runtime increases with total sequence length

• Number of chains: Multi-chain complexes require more computation than single chains

• MSA size: Larger MSAs can improve accuracy but increase memory usage and computation time

• Ligands, DNA, RNA: Additional molecular components increase computational cost

Model Configuration

• Inference backend: TensorRT + cuEquivariance provides significant speedups over PyTorch

• Diffusion samples: Multiple samples provide diversity but multiply computational cost

Performance Characteristics

Typical Runtimes

For reference, approximate runtimes on high-end hardware (NVIDIA H100 80GB):

Structure Prediction (OpenFold3 v1.0.0 on H100 with TensorRT + cuEquivariance):

• ~200 residues: 11.8-14.5 seconds

• ~300-400 residues: 15.0-20.4 seconds

• ~500-600 residues: 21.5-25.7 seconds

• ~800-900 residues: 32.7 seconds

• ~1300-1500 residues: 46.5-67.9 seconds

• ~1700-1900 residues: 88.7-101.8 seconds

• Memory usage: Varies with sequence length, typically 40-80GB GPU memory

Performance Notes The total runtime for structure prediction depends on:

• Total number of residues in the complex

• Total number of atoms in the complex

• Number of molecules and chains

• Number of sequences in the MSAs

• Number of diffusion samples requested

Performance Results on H100 (NVIDIA H100 80GB)

Structure Prediction Performance

The following table shows runtime performance for OpenFold3 NIM on NVIDIA H100 GPUs across different inference backends. The rightmost column shows speedup at settings where structural template processing is not active. These backends are described in Backend Selection and in Backend Optimization Options.

Note

Structural template support is available starting from version 1.1.0.

Test ID	Sequence Length	PyTorch + cuEquivariance (s)	Open source OF3 from openfold consortium (baseline) (s)	OF3 NIM (TRT + cuEquivariance) (s)	OF3 NIM with Templates (TRT + cuEquivariance) (s)	Speedup (NIM vs Baseline)
8eil	186	17.06	16.79	11.78	15.52	1.42x
7r6r	203	18.82	19.26	14.53	16.60	1.33x
1a3n	287	26.76	29.13	23.55	31.24	1.24x
8c4d	331	20.08	18.85	15.04	15.37	1.25x
7qsj	375	20.79	19.92	16.06	18.10	1.24x
8cpk	384	26.85	23.50	20.37	22.12	1.15x
8are	530	26.31	27.42	21.89	23.82	1.25x
8owf	575	27.54	28.86	22.90	25.02	1.26x
8aw3	590	41.96	45.64	37.69	41.40	1.21x
7tpu	616	25.69	28.61	21.48	23.48	1.33x
7ylz	623	31.80	35.66	27.51	30.38	1.30x
8gpp	628	29.92	33.81	25.71	34.91	1.32x
8clz	684	30.68	35.35	26.41	30.41	1.34x
8k7x	858	37.39	44.76	32.75	37.87	1.37x
8ibx	1286	52.12	63.54	46.51	48.06	1.37x
8gi1	1464	74.42	99.45	60.56	62.28	1.64x
8sm6	1496	83.33	110.55	67.92	72.79	1.63x
8pso	1499	72.68	96.87	57.41	61.76	1.69x
8jue	1657	98.09	125.53	78.04	85.66	1.61x
8bsh	1762	110.98	144.03	88.68	96.94	1.62x
5xgo	1869	127.79	163.83	101.79	111.89	1.61x

All runtimes are in seconds for end-to-end structure prediction with a single diffusion sample. The template measurements include an average of 4 CIF template files per protein chain. The additional time is primarily attributed to CIF parsing.

Performance Analysis

Key Observations:

• OpenFold3 NIM Optimization: Provides consistent 1.15x to 1.69x speedup over the open source OF3 baseline

• cuEquivariance Acceleration: PyTorch + cuEquivariance shows speedups for larger proteins, demonstrating the value of cuEquivariance optimization

• Scaling Behavior: Runtime increases with sequence length

• Best Performance: Largest speedups (1.6x-1.69x) observed for proteins in the 1400-1500 residue range

• Small Proteins: Still achieve significant speedups (1.2x-1.4x) even for sequences under 400 residues

Template Processing Impact:

• Template overhead: Structural templates add modest overhead (typically 0.3-10 seconds) depending on sequence length

• Small proteins (<400 residues): Template overhead is ~0.3-4 seconds

• Medium proteins (400-1000 residues): Template overhead is ~2-9 seconds

• Large proteins (>1000 residues): Template overhead is ~1.5-10 seconds

• Primary cost: The additional time is mostly attributed to CIF parsing (average 4 templates per chain)

• Overall impact: Templates provide structural guidance with minimal performance cost

Backend Comparison:

• Open source OF3 (PyTorch + DeepSpeed): Baseline implementation from the openfold consortium, good for development and debugging

• PyTorch + cuEquivariance: Improved performance, especially for larger proteins, while maintaining PyTorch flexibility

• OpenFold3 NIM (TensorRT + cuEquivariance): Best performance across all sequence lengths, recommended for production deployments

Recommended Configuration:

• For development/testing: Use open source OF3 (PyTorch + DeepSpeed) for easier debugging and flexibility

• For production: Use OpenFold3 NIM (TensorRT + cuEquivariance, default) for optimal performance

• For large proteins (>1500 residues): OpenFold3 NIM (TensorRT + cuEquivariance) provides the best speedups (1.6x-1.7x)

• For sequence lengths outside 4-2048 range: Use PyTorch backend as TensorRT has length limitations

• For template-guided predictions: Structural templates add minimal performance overhead

Configuration

The benchmarks use the following configuration:

Parameter	Setting
diffusion_samples	1
output_format	pdb
GPU	H100 80GB
structural_templates	Average 4 CIF files per chain

Performance Metrics

The following tables contain end-to-end runtime (seconds) for the OpenFold3 NIM across supported NVIDIA hardware units, optimization backends, and structural template configuration. Inputs are the same 21 benchmark cases, arranged by sequence length and annotated with PDB ID and sequence length.

Table 1: Performance Across the Supported NVIDIA Hardware Units

The table shows the prediction times for tests with varying sequence lengths for hardware units. It features the following:

• Hardware: Nine supported GPU SKUs, including NVIDIA A100 80GB HBM3, B200, H100 80GB HBM3, H200, GB200, L40s, GB10, RTX6000, and GH200. One tab per SKU. Refer to the tables below.

• Metric: End-to-end predict time (seconds).

• Configuration: Default model; no structural templates; default optimization backend (TensorRT + cuEquivariance). Inputs sorted by sequence length and annotated with PDB ID and sequence length.

NVIDIA A100 80GB HBM3

Test ID	Seq Length	predict_time (s)
8eil	186	14.03
7r6r	203	15.08
1a3n	287	26.57
8c4d	331	16.28
7qsj	375	17.93
8cpk	384	20.47
8are	530	24.51
8owf	575	26.39
8aw3	590	41.71
7tpu	616	25.28
7ylz	623	31.70
8gpp	628	30.04
8clz	684	31.71
8k7x	858	40.16
8ibx	1286	53.05
8gi1	1464	86.39
8sm6	1496	94.46
8pso	1499	82.84
8jue	1657	115.30
8bsh	1762	131.57
5xgo	1869	150.10

NVIDIA B200

Test ID	Seq Length	predict_time (s)
8eil	186	10.05
7r6r	203	10.63
1a3n	287	19.15
8c4d	331	11.57
7qsj	375	12.89
8cpk	384	15.67
8are	530	17.63
8owf	575	18.75
8aw3	590	29.59
7tpu	616	17.07
7ylz	623	23.03
8gpp	628	21.68
8clz	684	22.17
8k7x	858	27.39
8ibx	1286	37.95
8gi1	1464	54.49
8sm6	1496	60.43
8pso	1499	54.29
8jue	1657	71.84
8bsh	1762	81.14
5xgo	1869	92.93

NVIDIA H100 80GB HBM3

Test ID	Seq Length	predict_time (s)
8eil	186	12.20
7r6r	203	12.72
1a3n	287	22.54
8c4d	331	13.80
7qsj	375	15.24
8cpk	384	18.33
8are	530	20.60
8owf	575	21.57
8aw3	590	34.50
7tpu	616	19.78
7ylz	623	26.78
8gpp	628	25.15
8clz	684	25.64
8k7x	858	30.69
8ibx	1286	39.58
8gi1	1464	56.50
8sm6	1496	63.45
8pso	1499	54.01
8jue	1657	73.99
8bsh	1762	85.16
5xgo	1869	97.41

NVIDIA GB200

Test ID	Seq Length	predict_time (s)
8eil	186	20.24
7r6r	203	20.81
1a3n	287	27.13
8c4d	331	21.21
7qsj	375	22.12
8cpk	384	23.99
8are	530	26.07
8owf	575	26.79
8aw3	590	35.74
7tpu	616	25.55
7ylz	623	29.52
8gpp	628	28.70
8clz	684	29.58
8k7x	858	34.86
8ibx	1286	44.12
8gi1	1464	59.46
8sm6	1496	63.62
8pso	1499	58.95
8jue	1657	74.93
8bsh	1762	82.36
5xgo	1869	93.29

NVIDIA H200

Test ID	Seq Length	predict_time (s)
8eil	186	10.38
7r6r	203	10.86
1a3n	287	19.52
8c4d	331	12.02
7qsj	375	13.14
8cpk	384	15.80
8are	530	17.77
8owf	575	18.78
8aw3	590	30.08
7tpu	616	17.17
7ylz	623	23.28
8gpp	628	21.72
8clz	684	22.35
8k7x	858	26.87
8ibx	1286	35.22
8gi1	1464	50.12
8sm6	1496	56.29
8pso	1499	48.47
8jue	1657	65.19
8bsh	1762	75.22
5xgo	1869	87.13

NVIDIA L40s

Test ID	Seq Length	predict_time (s)
8eil	186	12.34
7r6r	203	13.09
1a3n	287	28.57
8c4d	331	14.88
7qsj	375	16.89
8cpk	384	20.17
8are	530	25.78
8owf	575	27.83
8aw3	590	45.18
7tpu	616	27.33
7ylz	623	34.09
8gpp	628	32.77
8clz	684	34.85
8k7x	858	44.29
8ibx	1286	65.76
8gi1	1464	103.55
8sm6	1496	112.75
8pso	1499	98.41
8jue	1657	—
8bsh	1762	—
5xgo	1869	—

NVIDIA DGX Spark (GB10)

Test ID	Seq Length	predict_time (s)
8eil	186	13.52
7r6r	203	18.96
1a3n	287	73.84
8c4d	331	27.44
7qsj	375	32.58
8cpk	384	34.30
8are	530	62.91
8owf	575	73.95
8aw3	590	127.72
7tpu	616	81.40
7ylz	623	88.03
8gpp	628	87.25
8clz	684	103.79
8k7x	858	145.92
8ibx	1286	259.91
8gi1	1464	615.61
8sm6	1496	657.24
8pso	1499	567.58
8jue	1657	—
8bsh	1762	—
5xgo	1869	—

NVIDIA RTX6000-Ada

Test ID	Seq Length	predict_time (s)
8eil	186	8.90
7r6r	203	9.67
1a3n	287	19.52
8c4d	331	10.71
7qsj	375	11.83
8cpk	384	13.96
8are	530	17.56
8owf	575	19.08
8aw3	590	30.86
7tpu	616	18.51
7ylz	623	23.38
8gpp	628	22.31
8clz	684	23.81
8k7x	858	31.06
8ibx	1286	45.42
8gi1	1464	70.29
8sm6	1496	76.67
8pso	1499	68.76
8jue	1657	92.27
8bsh	1762	102.93
5xgo	1869	118.07

NVIDIA GH200

Test ID	Seq Length	predict_time (s)
8eil	186	19.18
7r6r	203	19.63
1a3n	287	25.52
8c4d	331	20.26
7qsj	375	21.23
8cpk	384	22.75
8are	530	24.87
8owf	575	25.31
8aw3	590	34.04
7tpu	616	24.15
7ylz	623	28.26
8gpp	628	27.37
8clz	684	28.32
8k7x	858	32.78
8ibx	1286	39.84
8gi1	1464	53.61
8sm6	1496	58.12
8pso	1499	52.10
8jue	1657	67.16
8bsh	1762	76.03
5xgo	1869	86.87

Table 2: Performance Across Optimization Backends

The table shows the prediction times for tests with varying sequence lengths for optimized backends. It features the following:

• Hardware: NVIDIA H100 80GB HBM3. The same comparison is supported on the other SKUs listed in Table 1.

• Metric: End-to-end predict time (seconds).

• Configuration: Default model; no structural templates. Inputs sorted by sequence length and annotated with PDB ID and sequence length.

Test ID	Seq Length	torch_baseline (s)	torch (s)	trt (s)	trt-speedup-over-baseline
8eil	186	17.40	17.34	12.20	1.43x
7r6r	203	17.88	17.83	12.72	1.41x
1a3n	287	28.00	26.05	22.54	1.24x
8c4d	331	18.50	18.14	13.80	1.34x
7qsj	375	19.88	19.32	15.24	1.30x
8cpk	384	23.03	22.61	18.33	1.26x
8are	530	25.95	24.38	20.60	1.26x
8owf	575	27.26	25.26	21.57	1.26x
8aw3	590	41.02	37.46	34.50	1.19x
7tpu	616	25.79	23.53	19.78	1.30x
7ylz	623	32.43	30.24	26.78	1.21x
8gpp	628	31.10	28.71	25.15	1.24x
8clz	684	32.16	28.98	25.64	1.25x
8k7x	858	39.10	33.46	30.69	1.27x
8ibx	1286	58.03	47.14	39.58	1.47x
8gi1	1464	92.36	69.92	56.50	1.63x
8sm6	1496	101.92	77.46	63.45	1.61x
8pso	1499	88.69	68.09	54.01	1.64x
8jue	1657	119.95	93.32	73.99	1.62x
8bsh	1762	137.31	105.20	85.16	1.61x
5xgo	1869	156.68	122.00	97.41	1.61x

Table 3: Performance Impact From Structural Templates

The table shows the prediction times for tests with varying sequence lengths for structural templates. It features the following:

• Hardware: NVIDIA H100 80GB HBM3. The same comparison is supported on the other SKUs listed in Table 1.

• Metric: End-to-end predict time (seconds).

• Configuration: Default model; default optimization backend (TensorRT + cuEquivariance). Inputs sorted by sequence length and annotated with PDB ID and sequence length.

Test ID	Seq Length	Without structural templates (s)	With structural templates (s)
8eil	186	12.20	17.08
7r6r	203	12.72	15.39
1a3n	287	22.54	31.28
8c4d	331	13.80	15.24
7qsj	375	15.24	17.28
8cpk	384	18.33	21.35
8are	530	20.60	22.62
8owf	575	21.57	23.24
8aw3	590	34.50	37.10
7tpu	616	19.78	21.93
7ylz	623	26.78	28.70
8gpp	628	25.15	31.92
8clz	684	25.64	27.67
8k7x	858	30.69	33.49
8ibx	1286	39.58	44.57
8gi1	1464	56.50	58.45
8sm6	1496	63.45	68.29
8pso	1499	54.01	58.25
8jue	1657	73.99	83.12
8bsh	1762	85.16	93.68
5xgo	1869	97.41	107.87

Performance Optimization Tips

Backend Selection

• Default (TensorRT + cuEquivariance): Best for most use cases

  # This is the default, no environment variable needed
  

• PyTorch + cuEquivariance: For flexibility with good performance

  export NIM_OPTIMIZED_BACKEND=torch
  

• PyTorch + DeepSpeed: For debugging or sequences outside TRT range

  export NIM_OPTIMIZED_BACKEND=torch_baseline
  

General Optimization Tips

• GPU Selection: Use H100, H200, or B200 GPUs for optimal performance. A100 and L40S GPUs are also supported.

• TRT Sequence Length Limits: TensorRT mode (default) supports sequences between 4 and 2048 residues. For sequences outside this range, use PyTorch backend.

• Sequence Length: Performance scales with sequence length.

• Multiple Samples: Setting diffusion_samples > 1 will increase runtime in affine fashion (input featurization time is independent of diffusion_samples).

• MSA Size: While larger MSAs can improve accuracy, they also increase memory usage and computation time. Consider filtering MSAs for very large proteins.

• Structural Templates: Templates add modest overhead (1-10 seconds) but can significantly improve prediction accuracy. See Template Processing for guidance.

• Batch Processing: For multiple independent predictions, process them sequentially or use multiple NIM instances.

• Memory Management: Ensure adequate GPU memory for your target sequence lengths. Very long sequences (>1800 residues) may require GPUs with 80GB+ memory (H100, H200, A100, or B200).

Reproducing Performance Benchmarks

Overview

This section provides scripts and instructions for reproducing the performance metrics reported above. The benchmarking process is as follows:

1. Run inference with OpenFold3 NIM to generate predicted structures

2. Use OpenStructure (OST) to compare predictions against reference structures

3. Extract accuracy metrics like lDDT (local Distance Difference Test)

Prerequisites

Run the OpenFold3 NIM

Before running benchmarks, ensure that OpenFold3 NIM is deployed and running. The benchmarking scripts will send inference requests to the NIM service.

Verify NIM is Running:

# Check if NIM is accessible and ready
curl http://localhost:8000/v1/health/ready

# Expected response:
# {"object":"health.response","message":"ready","status":"ready"}

Note

The default NIM URL is http://localhost:8000. If your NIM is running on a different host or port, you’ll need to specify it using the --nim-url parameter when running benchmarks. For detailed deployment instructions and configuration options, refer to the Getting Started guide.

Install the OpenStructure Docker Image

OpenStructure is a computational structural biology framework that provides tools for structure comparison and validation. You’ll need the OpenStructure Docker image for running benchmarks:

# Pull the latest OpenStructure image from the OST registry
docker pull registry.scicore.unibas.ch/schwede/openstructure:latest

# Verify the installation
docker run --rm -v $(pwd):/home registry.scicore.unibas.ch/schwede/openstructure:latest --version

# Expected response:
# OpenStructure 2.11.1

Download Reference Structures

You’ll need reference structures (ground truth) in CIF format for comparison. These are experimentally determined structures from the Protein Data Bank (PDB):

• CIF (Crystallographic Information File): A standard format for representing molecular structures, including atomic coordinates, experimental metadata, and structural annotations

• Obtaining reference structures: Download from RCSB PDB using the PDB IDs from the performance table (e.g., 8eil, 7r6r, 1a3n)

Example:

# reate directories
mkdir -p references

# Download a reference structure (e.g., 8eil) to references folder
wget https://files.rcsb.org/download/8eil.cif -O references/8eil.cif

Preparing Input JSON Files

What are Input JSON Files?

Input JSON files contain the sequence information needed for OpenFold3 NIM to make predictions. Each file specifies:

• Protein sequences (amino acid chains)

• Chain IDs

• Optional: DNA, RNA, ligands, MSAs, templates

Example Input JSON Structure

{
  "name": "8eil",
  "sequences": [
    {
      "protein": {
        "id": "A",
        "sequence": "MKQHKAMIVALIVICITAVVAALVTRKDLCEVHIRTGQTEVAVF..."
      }
    }
  ]
}

Create Input JSON Files

You can create input JSON files by extracting sequences from PDB structures. Save this script as generate_inputs.py:

#!/usr/bin/env python3
"""Generate input JSON files from PDB structures."""
import argparse
import json
import requests
import sys
from pathlib import Path

def fetch_sequences_from_pdb(pdb_id):
    """Fetch protein sequences from RCSB PDB FASTA endpoint."""
    url = f"https://www.rcsb.org/fasta/entry/{pdb_id}"
    response = requests.get(url)
    if response.status_code != 200:
        raise Exception(f"Failed to fetch FASTA for {pdb_id}")
    
    sequences = []
    current_seq = []
    chain_index = 0
    
    for line in response.text.strip().split('\n'):
        if line.startswith('>'):
            if current_seq:
                # Use sequential chain IDs: A, B, C, D, ...
                chain_id = chr(65 + chain_index)
                sequences.append({
                    "protein": {
                        "id": chain_id,
                        "sequence": ''.join(current_seq)
                    }
                })
                chain_index += 1
            current_seq = []
        else:
            current_seq.append(line.strip())
    
    # Don't forget the last sequence
    if current_seq:
        chain_id = chr(65 + chain_index)
        sequences.append({
            "protein": {
                "id": chain_id,
                "sequence": ''.join(current_seq)
            }
        })
    
    return sequences

def create_input_json(pdb_id):
    """Create OpenFold3 input JSON file for a PDB structure."""
    output_dir = Path("inputs")
    output_dir.mkdir(parents=True, exist_ok=True)
    
    print(f"Fetching sequences for {pdb_id}...")
    sequences = fetch_sequences_from_pdb(pdb_id)
    
    input_data = {
        "name": pdb_id,
        "sequences": sequences
    }
    
    output_path = output_dir / f"{pdb_id}_input.json"
    with open(output_path, 'w') as f:
        json.dump(input_data, f, indent=2)
    
    print(f"Created {output_path} ({len(sequences)} chain(s))")
    return output_path

if __name__ == "__main__":
    # All benchmark PDB IDs from performance table
    benchmark_pdbs = [
        "8eil", "7r6r", "1a3n", "8c4d", "7qsj", "8cpk", 
        "8are", "8owf", "8aw3", "7tpu", "7ylz", "8gpp", 
        "8clz", "8k7x", "8ibx", "8gi1", "8sm6", "8pso", 
        "8jue", "8bsh", "5xgo"
    ]
    
    parser = argparse.ArgumentParser(
        description='Generate OpenFold3 input JSON files from PDB structures',
        epilog='Example: python generate_inputs.py --pdb 8eil'
    )
    parser.add_argument('--pdb', type=str, 
                        help='Specific PDB ID to generate input for (e.g., 8eil)')
    parser.add_argument('--all', action='store_true',
                        help='Generate input files for all benchmark PDB IDs')
    
    args = parser.parse_args()
    
    # For backwards compatibility: if no arguments provided, generate all
    if not args.pdb and not args.all:
        args.all = True
    
    if args.pdb:
        # Generate input for specific PDB ID
        try:
            create_input_json(args.pdb)
        except Exception as e:
            print(f"Error processing {args.pdb}: {e}", file=sys.stderr)
            sys.exit(1)
    elif args.all:
        # Generate all benchmark inputs
        print("Generating input files for all benchmark cases...")
        failed = []
        for pdb_id in benchmark_pdbs:
            try:
                create_input_json(pdb_id)
            except Exception as e:
                print(f"Error processing {pdb_id}: {e}")
                failed.append(pdb_id)
        
        if failed:
            print(f"\nFailed to generate inputs for: {', '.join(failed)}", file=sys.stderr)
            sys.exit(1)

Benchmarking Script

The following is a complete script to benchmark OpenFold3 NIM predictions. Save this as benchmark_openfold3.py:

#!/usr/bin/env python3
"""
Benchmark OpenFold3 NIM predictions against reference structures.
"""

import argparse
import json
import subprocess
import time
from pathlib import Path
import requests

def run_inference(nim_url, input_json, output_dir):
    """
    Run OpenFold3 NIM inference and save the predicted structure.
    
    Args:
        nim_url: URL of the OpenFold3 NIM service (e.g., http://localhost:8000)
        input_json: Path to input JSON file with sequence information
        output_dir: Directory to save predicted structures
    
    Returns:
        tuple: (predicted_pdb_path, inference_time_seconds)
    """
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    
    # Read input configuration
    with open(input_json, 'r') as f:
        input_data = json.load(f)
    
    # Convert to NIM API format
    molecules = []
    for seq in input_data.get("sequences", []):
        if "protein" in seq:
            protein = seq["protein"]
            # Create minimal MSA with just the query sequence
            msa_csv = f"key,sequence\n-1,{protein['sequence']}"
            molecules.append({
                "type": "protein",
                "id": [protein["id"]],
                "sequence": protein["sequence"],
                "msa": {
                    "main_db": {
                        "csv": {
                            "alignment": msa_csv,
                            "format": "csv"
                        }
                    }
                }
            })
    
    nim_request = {
        "inputs": [{
            "input_id": input_data.get("name", "prediction"),
            "molecules": molecules,
            "output_format": "pdb"
        }]
    }
    
    # Start timing
    start_time = time.time()
    
    # Run inference
    response = requests.post(
        f"{nim_url}/biology/openfold/openfold3/predict",
        json=nim_request,
        headers={"Content-Type": "application/json"}
    )
    
    # End timing
    inference_time = time.time() - start_time
    
    if response.status_code != 200:
        raise Exception(f"Inference failed: {response.text}")
    
    # Extract PDB data from response
    result = response.json()
    
    outputs = result.get('outputs', [])
    if not outputs:
        raise Exception(f"No outputs in response")
    
    structures = outputs[0].get('structures_with_scores', [])
    if not structures:
        raise Exception(f"No structures in response")
    
    # Get the first (best) structure
    pdb_content = structures[0].get('structure', '')
    if not pdb_content:
        raise Exception(f"No structure content in response")
    
    # Save predicted structure
    pdb_id = input_data.get('name', 'prediction')
    pred_path = output_dir / f"{pdb_id}_pred.pdb"
    
    with open(pred_path, 'w') as f:
        f.write(pdb_content)
    
    return pred_path, inference_time

def compare_structures(pred_pdb, reference_cif, output_dir):
    """
    Compare predicted structure against reference using OpenStructure.
    
    Args:
        pred_pdb: Path to predicted structure (PDB format)
        reference_cif: Path to reference structure (CIF format)
        output_dir: Directory to save comparison results
    
    Returns:
        dict: Comparison metrics including lDDT score
    """
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    
    out_path = output_dir / "comparison_results.json"
    
    # Build OpenStructure comparison command
    cmd = [
        "compare-structures",
        "-m", str(pred_pdb),           # Model (predicted structure)
        "-r", str(reference_cif),      # Reference (ground truth)
        "--fault-tolerant",            # Handle minor structural differences
        "--min-pep-length", "4",       # Minimum peptide chain length
        "--min-nuc-length", "4",       # Minimum nucleotide chain length
        "-o", str(out_path),           # Output JSON file
        "--lddt",                      # Calculate lDDT metric
    ]
    
    # Run comparison using OpenStructure Docker container
    docker_cmd = [
        "docker", "run", "--rm",
        "-v", f"{Path.cwd()}:/home",
        "registry.scicore.unibas.ch/schwede/openstructure:latest"
    ] + cmd
    
    result = subprocess.run(
        docker_cmd,
        capture_output=True,
        text=True
    )
    
    if result.returncode != 0:
        raise Exception(f"Structure comparison failed: {result.stderr}")
    
    # Read comparison results
    with open(out_path, 'r') as f:
        metrics = json.load(f)
    
    return metrics

def extract_lddt(comparison_results):
    """
    Extract lDDT score from comparison results.
    
    Args:
        comparison_results: Dictionary containing comparison metrics
    
    Returns:
        float: lDDT score (0.0 to 1.0, higher is better)
    """
    # lDDT is directly a float value in the comparison results
    lddt_score = comparison_results.get('lddt', 0.0)
    return lddt_score

def benchmark_structure(nim_url, input_json, reference_cif, output_dir):
    """
    Complete benchmark pipeline for a single structure.
    
    Args:
        nim_url: URL of OpenFold3 NIM service
        input_json: Input configuration for prediction
        reference_cif: Reference structure for validation
        output_dir: Output directory for results
    
    Returns:
        dict: Benchmark results including timing and accuracy
    """
    output_dir = Path(output_dir)
    
    print(f"Running inference...")
    pred_path, inference_time = run_inference(nim_url, input_json, output_dir)
    print(f"Inference completed in {inference_time:.2f} seconds")
    
    print(f"Comparing structures...")
    metrics = compare_structures(pred_path, reference_cif, output_dir)
    lddt_score = extract_lddt(metrics)
    print(f"lDDT score: {lddt_score:.4f}")
    
    return {
        'inference_time': inference_time,
        'lddt_score': lddt_score,
        'predicted_structure': str(pred_path),
        'full_metrics': metrics
    }

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='Benchmark OpenFold3 NIM predictions')
    parser.add_argument('--nim-url', default='http://localhost:8000', 
                        help='URL of the OpenFold3 NIM service')
    parser.add_argument('--input', required=True, 
                        help='Path to input JSON file')
    parser.add_argument('--reference', required=True, 
                        help='Path to reference CIF file')
    parser.add_argument('--output', required=True, 
                        help='Output directory for results')
    
    args = parser.parse_args()
    
    results = benchmark_structure(args.nim_url, args.input, args.reference, args.output)
    
    print("\nBenchmark Results:")
    print(f"  Inference Time: {results['inference_time']:.2f}s")
    print(f"  lDDT Score: {results['lddt_score']:.4f}")
    print(f"  Predicted Structure: {results['predicted_structure']}")

Running a Single Benchmark

To run a benchmark for a single structure, follow these steps in order:

# Step 1: Create directories
mkdir -p inputs results

# Step 2: Generate the input JSON file
python3 generate_inputs.py --pdb 8eil

# Step 3: Verify files exist before running benchmark
ls -l inputs/8eil_input.json references/8eil.cif

# Step 4: Run the benchmark (requires OpenFold3 NIM to be running)
python3 benchmark_openfold3.py \
    --nim-url http://localhost:8000 \
    --input inputs/8eil_input.json \
    --reference references/8eil.cif \
    --output results/8eil

Note

Each step must complete successfully before proceeding to the next. Verify that input and reference files exist (Step 4) before running the benchmark (Step 5).

Understanding lDDT Metric

lDDT (local Distance Difference Test) is a robust metric for assessing the quality of predicted protein structures:

• Score Range: 0.0 to 1.0 (or 0 to 100 when expressed as percentage)

• Interpretation:

  • lDDT > 0.90: Excellent accuracy, very high confidence

  • lDDT 0.70-0.90: Good accuracy, reliable predictions

  • lDDT 0.50-0.70: Moderate accuracy, some structural features correct

  • lDDT < 0.50: Low accuracy, limited reliability

How lDDT Works:

• Measures local geometric agreement between predicted and reference structures

• Evaluates distances between atoms within local neighborhoods (typically 15Å radius)

• More robust to domain movements and flexible regions compared to global metrics like RMSD

• Focuses on local structural correctness rather than global superposition

Why lDDT for OpenFold3:

• AlphaFold3 and OpenFold3 models are trained to optimize lDDT during the training process

• Well-suited for evaluating multi-chain complexes and structures with flexible regions

• Provides per-residue metrics in addition to global metrics

• Standard metric used in CASP (Critical Assessment of protein Structure Prediction) competitions

Running Benchmarks for Performance Table

To reproduce the performance metrics from the table above, follow these steps in order:

Important

Complete Steps 1 and 2 before running Step 3. The benchmark workflow requires:

1. Both Python scripts (generate_inputs.py and benchmark_openfold3.py) saved as files

2. The bash script (run_benchmarks.sh) saved as a file

3. All scripts must be in the same directory before execution

Running steps out of order will result in “File Not Found” errors.

Step 1: Prepare the Python scripts

Save the two Python scripts provided earlier in this document:

• generate_inputs.py - Creates input JSON files from PDB (see Preparing Input JSON Files)

• benchmark_openfold3.py - Runs inference and compares structures (see Benchmarking Script)

Step 2: Create the benchmark workflow script

Save the following as run_benchmarks.sh:

#!/bin/bash
# Complete benchmark workflow for all test cases

NIM_URL="http://localhost:8000"
PDB_IDS=("8eil" "7r6r" "1a3n" "8c4d" "7qsj" "8cpk" "8are" "8owf" "8aw3" "7tpu" "7ylz" "8gpp" "8clz" "8k7x" "8ibx" "8gi1" "8sm6" "8pso" "8jue" "8bsh" "5xgo")

# Create directories
mkdir -p references inputs results

# Step 1: Generate input JSON files from PDB
echo "Step 1: Generating input JSON files..."
python3 generate_inputs.py
echo "✓ Input files created in inputs/"

# Step 2: Download reference structures
echo "Step 2: Downloading reference structures..."
for pdb_id in "${PDB_IDS[@]}"; do
    if [ ! -f "references/${pdb_id}.cif" ]; then
        echo "  Downloading ${pdb_id}.cif..."
        wget -q "https://files.rcsb.org/download/${pdb_id}.cif" -O "references/${pdb_id}.cif"
    fi
done
echo "✓ Reference structures downloaded"

# Step 3: Verify NIM is running
echo "Step 3: Checking NIM availability..."
curl -sf "${NIM_URL}/v1/health/ready" > /dev/null || {
    echo "Error: NIM not accessible at ${NIM_URL}"
    exit 1
}
echo "✓ NIM is ready"

# Step 4: Run benchmarks
echo "Step 4: Running benchmarks..."
for pdb_id in "${PDB_IDS[@]}"; do
    echo "  Benchmarking ${pdb_id}..."
    python3 benchmark_openfold3.py \
        --nim-url "$NIM_URL" \
        --input "inputs/${pdb_id}_input.json" \
        --reference "references/${pdb_id}.cif" \
        --output "results/${pdb_id}"
done

echo ""
echo "Benchmarking complete! Results saved in results/"

Step 3: Run the complete workflow

# 1. Install dependencies
pip3 install requests

# 2. Make the benchmark script executable
chmod +x run_benchmarks.sh

# 3. Run the complete workflow (generates inputs, downloads references, and runs benchmarks)
./run_benchmarks.sh

Note

The run_benchmarks.sh script handles all steps automatically: generating input files, downloading reference structures, and running benchmarks for all test cases. Make sure the OpenFold3 NIM is running before executing the script.

Expected Output Format

The comparison results JSON file contains detailed metrics:

{
  "lddt": 0.9234,
  "chain_mapping": {
    "A": "A",
    "B": "B"
  },
  "aln": [
    ">reference:A\nMKQLYGHSTI...",
    ">model:A\nMKQLYGHSTI..."
  ],
  "model_clashes": [],
  "model_bad_bonds": [],
  "model_bad_angles": [],
  "reference_clashes": [],
  "reference_bad_bonds": [],
  "reference_bad_angles": [],
  "status": "SUCCESS",
  "ost_version": "2.11.1"
}

The primary metric of interest is lddt, which is a float value (0.0 to 1.0) representing the overall structural quality of the prediction. Higher values indicate better agreement with the reference structure.

Troubleshooting

General Performance Issues

• Out of memory errors: Reduce MSA depth, decrease sequence length, or upgrade to GPUs with more memory

• Slow performance: Ensure fast storage (NVMe SSD), sufficient CPU cores, and adequate system RAM

• Poor quality predictions: Check input sequence quality, increase MSA depth if available, or adjust diffusion parameters

Backend-Specific Issues

• TensorRT errors with long sequences: Use PyTorch backend for sequences >2048 residues

  export NIM_OPTIMIZED_BACKEND=torch
  

• TensorRT errors with short sequences: Use PyTorch backend for sequences <4 residues

  export NIM_OPTIMIZED_BACKEND=torch
  

• Inconsistent results between backends: This is expected; TensorRT uses optimizations that may produce slightly different numerical results while maintaining accuracy

Benchmarking Issues

The following are “File not found” errors.

• “Input file not found”: Generate input JSON files first using generate_inputs.py

  python3 generate_inputs.py --all  # Creates all input files
  python3 generate_inputs.py --pdb 8eil  # Or for specific PDB
  

• “Reference file not found”: Download CIF files from RCSB PDB

  wget https://files.rcsb.org/download/8eil.cif -O references/8eil.cif
  

The following are structure comparison issues.

• OpenStructure comparison failures: Ensure both predicted and reference structures have compatible chain IDs and residue numbering

• Missing atoms in predictions: Use --fault-tolerant flag to handle incomplete structures (already included in script)

• lDDT score of 0.0: Check that sequences match between prediction and reference; may indicate alignment failure

The following are NIM connection issues.

• “Connection refused”: Ensure OpenFold3 NIM is running

  curl http://localhost:8000/v1/health/ready  # Should return {"object":"health.response","message":"ready","status":"ready"}
  

Note

For detailed performance tuning guidance specific to your deployment, refer to the documentation on configuration and optimization.