IntelliFold 2

What is IntelliFold 2?

IntelliFold 2 is an all-atom biomolecular structure prediction model for proteins, ligands, DNA, RNA, and mixed complexes. ProteinIQ runs the open-source IntelliFold 2.0.4 release on its vendored AlphaFold 3 JAX engine.

The model generates several diffusion samples for each seed, scores them, and returns the complete native result set. Optional physical steering can guide ligand poses using differentiable geometric and chemical potentials.

Applications

• Single-chain and multimer protein structure prediction
• Protein-ligand complex modeling
• Protein-DNA and protein-RNA complex modeling
• Alternative-conformation sampling across seeds and diffusion samples
• Physically guided ligand-pose generation

How to use IntelliFold 2 online

Inputs

At least one protein, DNA, or RNA chain is required. The combined limit is 3,000 residues after molecule copies are counted.

Input	Description
Protein	One FASTA sequence per entry, supplied as text, .fasta/.fa, or an RCSB sequence fetch.
MSA	Optional protein alignment in A3M format. Multiple A3M files are matched to protein inputs in submission order.
Ligand	One SMILES string per entry, entered directly or fetched from PubChem.
CCD ligand	One PDB Chemical Component Dictionary code, such as ATP, NAD, or SAH.
DNA	One DNA sequence per FASTA entry.
RNA	One RNA sequence per FASTA entry.

Molecule copies, chain IDs, and supported protein or nucleic-acid modifications are carried into the AlphaFold 3 version 4 input contract. Cyclic polymers are rejected because IntelliFold 2.0.4 does not define a cyclic-chain input field.

When no A3M is supplied, ProteinIQ runs the source-supported single-sequence path. It does not claim to generate an MSA remotely; the 2.0.4 data pipeline requires large local sequence databases that are not part of this deployment.

Settings

Setting	Description
Diffusion samples	Structures generated per seed (1-10, default 5).
Recycling iterations	Model recycling passes (1-12, default 10).
Seeds	One integer seed or up to 10 unique comma-separated seeds. Duplicate numeric seeds are removed.
Enable steering	Turns on IntelliFold's physical and chemical diffusion guidance. Off by default.
Steering optimization steps	Gradient-descent iterations per denoising step when steering is enabled (default 20).
Steering weight scale	Global multiplier for the steering potentials (default 1.0).

IntelliFold 2.0.4 uses one published v2 checkpoint and writes mmCIF. Earlier v1, v2, v2-Flash, sampling-step, MMseqs2, and PDB-output controls belonged to the previous inference implementation and are not options in the current release.

Output

ProteinIQ returns every native text artifact produced by the run:

Output	Description
Ranked sample CIF files	One predicted structure for every seed and diffusion sample.
Top-ranked CIF copy	The source-selected best model at the root of the result directory.
Summary confidence JSON	Ranking score and model-level confidence metrics for each structure.
Full confidence JSON (.zst)	Per-atom and pairwise confidence data produced by the model, using the source-supported Zstandard compression.
Ranking CSV	Native seed, sample, and ranking-score rows in source order.
Processed input JSON	The exact AlphaFold 3 input consumed by inference.
Terms file	The native output terms supplied by the AlphaFold 3 engine.

The data table uses the native ranking_score; higher values rank first in the structure viewer. The original ranking CSV remains downloadable without rewriting its row order.

How does IntelliFold 2 work?

The 2.0.4 release runs IntelliFold's converted v2 parameters through an AlphaFold 3 JAX inference engine. Protein, nucleic-acid, ligand, MSA, copy, chain, and modification data are serialized to the official AlphaFold 3 JSON dialect before invoking the published intellifold predict command.

Diffusion sampling and recycling

Each seed initializes an independent model trajectory. Diffusion samples broaden the generated ensemble, while recycling feeds intermediate representations through the network again for refinement. Increasing either control raises compute cost.

Optional physical steering

Steering adds gradient guidance for physical and chemical constraints during diffusion, including ligand geometry and clash-related potentials. Because steering changes the scientific run and increases computation, it is disabled unless explicitly selected.

Interpreting results

Start with the native ranking score, then inspect the summary and full confidence JSON for the selected sample. Confidence metrics may include pTM, ipTM, disorder, and clash-related values depending on the system. Agreement among high-ranked samples is often more informative than one structure alone.

Limitations

• Predictions are computational hypotheses, not experimental measurements.
• Flexible regions, weak interfaces, novel chemistry, and sparse sequence information can reduce reliability.
• This deployment supports uploaded A3M data or single-sequence inference; it does not run the source database search pipeline.
• The current source release writes CIF structures, not PDB files.
• Total protein, DNA, and RNA length is limited to 3,000 residues after copies.