Output Format

Understanding AlphaFold 3 output structure and files

Output Directory Structure

AlphaFold 3 creates an organized directory structure for each job. The directory name is the sanitized version of your job name.

For job name “My first fold (TEST)”, outputs are written to My_first_fold_TEST. If the directory exists, a timestamp is appended unless --force_output_dir is used.

Example Directory Layout

For a job named “Hello Fold” with 1 seed and 5 samples:

hello_fold/
├── seed-1234_distogram/                        # If --save_distogram=true
│   └── hello_fold_seed-1234_distogram.npz
├── seed-1234_embeddings/                       # If --save_embeddings=true
│   └── hello_fold_seed-1234_embeddings.npz
├── seed-1234_sample-0/
│   ├── hello_fold_seed-1234_sample-0_confidences.json
│   ├── hello_fold_seed-1234_sample-0_model.cif
│   └── hello_fold_seed-1234_sample-0_summary_confidences.json
├── seed-1234_sample-1/
│   ├── hello_fold_seed-1234_sample-1_confidences.json
│   ├── hello_fold_seed-1234_sample-1_model.cif
│   └── hello_fold_seed-1234_sample-1_summary_confidences.json
├── seed-1234_sample-2/
│   ├── hello_fold_seed-1234_sample-2_confidences.json
│   ├── hello_fold_seed-1234_sample-2_model.cif
│   └── hello_fold_seed-1234_sample-2_summary_confidences.json
├── seed-1234_sample-3/
│   ├── hello_fold_seed-1234_sample-3_confidences.json
│   ├── hello_fold_seed-1234_sample-3_model.cif
│   └── hello_fold_seed-1234_sample-3_summary_confidences.json
├── seed-1234_sample-4/
│   ├── hello_fold_seed-1234_sample-4_confidences.json
│   ├── hello_fold_seed-1234_sample-4_model.cif
│   └── hello_fold_seed-1234_sample-4_summary_confidences.json
├── TERMS_OF_USE.md
├── hello_fold_confidences.json
├── hello_fold_data.json
├── hello_fold_model.cif
├── hello_fold_ranking_scores.csv
└── hello_fold_summary_confidences.json

Output Files

Top-Level Files

model.cif

Top-ranked prediction structure

mmCIF format compatible with structural biology tools. No PDB format provided (convert CIF if needed).

confidences.json

Detailed confidence metrics

Full 1D/2D arrays of pLDDT, PAE, and contact probabilities for top prediction.

summary_confidences.json

Summary confidence scores

Scalar metrics like pTM, ipTM, ranking scores for top prediction.

data.json

Input with MSA/templates

Original input JSON augmented with MSA and template data from pipeline.

ranking_scores.csv · file

CSV file ranking all predictions. Highest ranking prediction is included in root directory.

TERMS_OF_USE.md · file

License and usage terms for AlphaFold 3 outputs

Per-Sample Subdirectories

For each seed and sample combination (seed-<seed>_sample-<n>), three files are generated:

1. Structure File

   <job>_seed-<seed>_sample-<n>_model.cif - Predicted structure in mmCIF format

2. Detailed Confidences

   <job>_seed-<seed>_sample-<n>_confidences.json - Full confidence arrays

3. Summary Confidences

   <job>_seed-<seed>_sample-<n>_summary_confidences.json - Scalar metrics

Optional Output Files

Distogram

Enabled with: --save_distogram=true

Location: seed-<seed>_distogram/distogram.npz

NumPy zip file containing distance predictions:

• Key: distogram
• Shape: (num_tokens, num_tokens, 64)
• Dtype: np.float16
• Size: ~3 GiB for 5,000 tokens

Embeddings

Enabled with: --save_embeddings=true

Location: seed-<seed>_embeddings/embeddings.npz

NumPy zip file with two keys:

• single_embeddings: (num_tokens, 384), np.float16
• pair_embeddings: (num_tokens, num_tokens, 128), np.float16
• Size: ~6 GiB for 5,000 tokens

import numpy as np

with open('embeddings.npz', 'rb') as f:
    embeddings = np.load(f)
    single = embeddings['single_embeddings']
    pair = embeddings['pair_embeddings']

Multi-Seed and Multi-Sample Results

By default, AlphaFold 3 generates 5 samples per seed. The top-ranked prediction across all samples and seeds is placed in the root directory.

Ranking Predictions

For ranking the full complex, use the ranking_score (higher is better):

ranking_score = 0.8 × ipTM + 0.2 × pTM + 0.5 × disorder - 100 × has_clash

This score includes:

• Structure confidence (pTM and ipTM)
• Disorder penalty for spurious helices
• Clash penalty for atomic conflicts

The ranking_score is designed for ranking only. Use pTM, ipTM, PAE, or pLDDT for quality assessment.

Chain-Specific Ranking

If interested in specific entities or interactions, rank by:

• chain_ptm: Confidence in individual chain structure
• chain_iptm: Confidence in chain interfaces with all other chains
• chain_pair_iptm: Confidence in specific two-chain interfaces
• chain_pair_pae_min: Minimum PAE between chain pairs (correlates with binding)

Confidence JSON Files

Two JSON files provide confidence metrics for each prediction:

Summary Confidences JSON

Scalar and per-chain/per-chain-pair metrics:

ptm · number

Predicted TM-score for full structure (0-1). Values >0.5 indicate correct overall fold.

iptm · number

Interface predicted TM-score (0-1). Values >0.8 = high quality, <0.6 = likely failed, 0.6-0.8 = uncertain.

fraction_disordered · number

Fraction of structure that is disordered (0-1), measured by accessible surface area.

has_clash · boolean

True if >50% of a chain has clashes, or >100 clashing atoms in any chain.

ranking_score · number

Composite score for ranking predictions (-100 to 1.5).

chain_ptm · array<number>

Per-chain pTM scores. Element i is pTM restricted to chain i.

chain_iptm · array<number>

Per-chain interface confidence. Average ipTM between each chain and all others.

chain_pair_iptm · array<array<number>>

[num_chains, num_chains] matrix. Off-diagonal (i,j) = ipTM for chains i-j interface. Diagonal (i,i) = pTM for chain i.

chain_pair_pae_min · array<array<number>>

[num_chains, num_chains] matrix. Element (i,j) = minimum PAE from chain i to chain j. Correlates with binding interactions.

Full Confidences JSON

Detailed per-atom and per-token arrays:

atom_plddts · array<number>

[num_atoms] array of per-atom predicted lDDT scores (0-100). Higher = more confident.

pae · array<array<number>>

[num_tokens, num_tokens] matrix. Element (i,j) = predicted error in position of token j when aligned using token i’s frame.

contact_probs · array<array<number>>

[num_tokens, num_tokens] matrix. Element (i,j) = probability tokens i and j are within 8Å.

token_chain_ids · array<string>

[num_tokens] array mapping tokens to chain IDs.

atom_chain_ids · array<string>

[num_atoms] array mapping atoms to chain IDs.

mmCIF Structure Files

The .cif files contain predicted 3D coordinates in the standard mmCIF format.

AlphaFold 3 does not output PDB format. Use standard tools to convert mmCIF to PDB if needed:

# Using PyMOL
pymol -c -d "load model.cif; save model.pdb"

# Using gemmi
gemmi convert model.cif model.pdb

Viewing Structures

Compatible with most structural biology tools:

• PyMOL: pymol model.cif
• ChimeraX: chimerax model.cif
• VMD: vmd model.cif
• Mol*: Web-based viewer at https://molstar.org

Data JSON File

The <job>_data.json file contains your original input augmented with:

• MSAs generated by genetic search
• Structural templates found by template search
• Other data pipeline outputs

This file can be reused as input with --norun_data_pipeline to skip expensive genetic searches.

Chirality Checks

For ligand predictions, chirality errors can be assessed using the provided utility:

from alphafold3.model.scoring.chirality import compare_chirality

# Compare predicted structure chirality against CCD reference
result = compare_chirality(predicted_mol, reference_mol)

In the AlphaFold 3 paper, a chirality penalty was applied to ranking scores for ligands. Running multiple seeds with chiral-aware ranking greatly reduces error rates.

File Sizes

Typical file sizes for a 5,000-token prediction:

File Type	Approximate Size
mmCIF structure	5-20 MB
Confidence JSON	200-500 MB
Summary JSON	1-10 KB
Distogram (optional)	~3 GB
Embeddings (optional)	~6 GB

Next Steps

Confidence Metrics

Deep dive into pLDDT, PAE, pTM, and ipTM

Input Format

Learn how to create input JSON files