Structure Prediction

Structure prediction takes a sequence (protein, DNA, RNA, or a small molecule as SMILES or a CCD code) and predicts its 3D structure. Results are returned as PDB or CIF files, along with confidence scores. Predicted structures are used throughout molecular discovery: preparing targets, designing and screening binders, docking small molecules, and checking developability.

Quick to

Section	Jump to
Choose a tool	Find structure prediction tools · Recommended tools · Tools for specific cases · Conformations and experimental data
How it works	Multiple sequence alignment (MSA) · Confidence metrics · Input limits and speed
Use cases	AlphaFold2 · AlphaFold3-class models · Antibodies · Peptides · Linkers
Walkthrough	Example workflow: protein–ligand complex with affinity · Read your results

Find structure prediction tools

On All Tools, open the Function filter (1) and check Structure Prediction (2). The filter appears as a chip (3). Click Clear all to remove it.

Recommended tools

For most proteins and complexes, start with one of these. AlphaFold3-class models (all-atom co-folding) can include proteins, nucleic acids, small molecules, and ions in the same prediction.

Tool	Inputs	Key outputs	Good for
Boltz-2	Protein, DNA/RNA, small molecule (SMILES or CCD)	Confidence score, pTM/ipTM, pLDDT, PAE/PDE, binding affinity	General complexes; protein–ligand complexes with affinity prediction; restraints
Chai-1	Protein, DNA/RNA, small molecule (SMILES)	Aggregate score, pLDDT, pTM/ipTM, PAE/PDE	General complexes; restraints; peptides
Protenix-v2	Protein, DNA/RNA, small molecule (CCD)	pLDDT, pTM/ipTM, PAE/PDE	Complexes, including antibody–antigen interfaces
OpenFold3	Protein, DNA/RNA, small molecule	pLDDT, pTM/ipTM, PAE	Open-source AlphaFold3 reproduction (Apache 2.0)
IntelliFold 2	Protein, DNA/RNA, small molecule	Confidence metrics	General complexes
ESMFold2	Protein, DNA/RNA, ligand	Confidence metrics	Protein, nucleic acid, and ligand complexes
OpenDDE	Protein, DNA/RNA, small molecule	Confidence metrics	Complexes, with a focus on interfaces
RosettaFold-3	Protein, DNA/RNA, small molecule	Confidence metrics	AlphaFold3 reproduction
AlphaFold (AlphaFold2 / ColabFold)	Protein	pLDDT, pTM/ipTM, PAE	Proteins and protein complexes; high-throughput protein–protein interaction screens

About these models:

Boltz-2 (MIT and Recursion, 2025) predicts structure and binding affinity together. Its developers report affinity accuracy approaching physics-based free energy methods at a fraction of the time and cost (Bio-IT World). It is released under the MIT license.
OpenFold3 was released by the OpenFold consortium in October 2025 under the Apache 2.0 license, and is available hosted on Tamarind (Business Wire).
Protenix is ByteDance’s open AlphaFold3 reproduction (bioRxiv).

Some models are also available as NVIDIA NIM versions (Boltz-2, OpenFold3, AlphaFold2-Multimer).

Tools for specific cases

Use case	Tools
Antibodies, nanobodies, TCRs (fast, no antigen)	ImmuneBuilder, ABodyBuilder3, FlashABB, NbForge (nanobodies), ABB4-STEROIDS (antibody conformational ensembles)
TCR and peptide–MHC	TCRModel2, TCRdock, SwiftMHC, MHC-Fine, MHC-Diff, PMGen
Cyclic peptides	Highfold, or Boltz-2 with the Cyclic option
Fast, no MSA	ESMFold, OmegaFold, DeCAF
Large complexes	CombFold (assembles large complexes from subunit predictions)
GPCRs	AF MultiState
RNA secondary structure	ViennaRNA, RNA-FM
Protein–ligand complexes	OpenBind (built on OpenFold3), Boltz-2 with affinity
Custom training	Boltz Finetune (fine-tune Boltz on your own CIF files)
Binder filtering and design	Promera

Conformations and experimental data

Goal	Tools
Multiple conformations	AFSample2, AlphaFlow, AF Cluster, SF-Cluster, AF-Traj
Folding pathway	PathFold
Use templates or multimeric templates	AF Unmasked
Crosslinking mass spec (XL-MS) or interface restraints	AlphaLink2, GRASP
Cryo-EM or X-ray density maps	CryoBoltz, Sampleworks

AFSample2 uses AlphaFold’s jackhmmer MSAs instead of MMseqs2, so runs can take several hours.

Multiple sequence alignment (MSA)

An MSA collects sequences related to your input from large databases. The evolutionary information it captures helps models predict the structure. The result is an .a3m file that’s passed to the model.

On by default. MSAs are generated automatically for structure prediction tools.
Turning it off. AlphaFold2 and Chai-1 let you disable MSAs. For Boltz-2, you can get a similar effect by limiting the MSA depth to one sequence.
How Tamarind builds MSAs. Tamarind uses the ColabFold MMseqs2 method to search UniRef30, BFD, and the ColabFold environmental database. Tamarind runs its own MSA servers with the databases loaded into memory, which makes MSA generation much faster.

Single-sequence (no MSA) predictions are faster, and can help for designed proteins or sequences with few natural relatives. For natural proteins, keep the MSA on.

Confidence metrics

Every prediction comes with confidence scores. Use them to decide which models to trust.

Metric	What it measures	How to read it
pLDDT	Confidence in the local structure of each residue (or atom), 0–100	>90 very high · 70–90 confident · 50–70 low · <50 very low (often disordered or wrong)
pTM	Confidence in the overall fold or complex, 0–1	>0.5 suggests the overall fold is likely correct
ipTM	Confidence in how chains are positioned relative to each other, 0–1	>0.8 confident interaction · 0.6–0.8 uncertain · <0.6 likely incorrect
PAE	Expected position error between pairs of residues, in Å	Lower is better. Low PAE between two chains or domains means their relative position is confident
PDE	Expected distance error between pairs of residues (AlphaFold3-class models)	Lower is better
Confidence score / aggregate score	A combined ranking score (model-specific)	Use it to rank samples from the same run

pTM is less useful for very small structures, and large disordered regions lower pTM and ipTM. A complex can still be predicted correctly with lower values if PAE between the ordered parts is low (EMBL-EBI).

Input limits and speed

	AlphaFold3-class models	AlphaFold2
Max size	2,048 tokens (residues)	About 5,000 residues
Typical runtime	About 10 minutes	30 minutes to several hours for large jobs
Runtime scales with	Number of samples	Number of recycles

Each tool page shows an Estimated Runtime and Est. Compute Cost before you submit. See Compute Hours.

AlphaFold2

AlphaFold2 moved the field from physics-based modeling to deep learning, using attention-based networks and evolutionary information from MSAs. Its training data cutoff is April 30, 2018, which is useful to know when benchmarking. Settings:

Models: AlphaFold2 has 5 sets of weights, so it returns up to 5 models, ranked with the best first.
Recycles: 3 by default, up to 20. By default, recycling stops early if the structure barely changes. Turn on No Stop to always run every recycle.
Relaxation: Uses ColabFold’s AMBER relaxation. For further energy minimization, use Tamarind’s OpenMM molecular dynamics tool.

When to use it: AlphaFold2 works well for high-throughput protein–protein interaction screens (pulldowns), up to hundreds of thousands of structures. Its confidence metrics are well calibrated as a binding indicator for protein–protein interactions.

AlphaFold3-class models

AlphaFold3 itself is not available for commercial use. Tamarind offers open models that reproduce or exceed its accuracy, including Boltz-2, Chai-1, Protenix-v2, OpenFold3, IntelliFold 2, and RosettaFold-3.

All-atom. These models can include small molecules, nucleic acids, ions, and modified residues.
Generative. Each run can produce many samples. To find a specific pose, generate more samples, then pick by confidence score or by clustering.
Ranking. The overall confidence score is the most useful metric for ranking samples.

Antibodies

With the antigen: Use Boltz-2, Chai-1, or Protenix-v2. If you know the epitope or paratope, add restraints and generate many samples.
Without the antigen: ImmuneBuilder, ABodyBuilder3, and FlashABB are very fast and give reasonable structures for large numbers of antibodies or nanobodies.
TCRs: Use TCRModel2 for class I, class II, and unbound TCRs.

Restraint types (Boltz-2 and Chai-1):

Restraint	What it does	Between
Contact	Keeps a residue on one chain close to a residue on another chain	Two protein chains
Pocket	Keeps a set of residues on one chain (the pocket) close to an entire other chain or ligand (the binder)	Two protein chains, or a protein and a ligand
Covalent	Creates a covalent bond between specific atoms	Two chains, or a protein and a ligand

Peptides

Boltz-2 and Chai-1 perform best for peptide–protein complexes. Try both and choose the one that best matches your data.
For cyclic peptides, use Boltz-2 and check Cyclic on the peptide chain, or use Highfold.

Linkers

Model linkers as part of the chain. For example, to join two proteins with a flexible linker, include the linker in the sequence and predict it as one chain with any tool:

AAAAAAGGGGSGGGGSGGGGSGGGGSAAAAAAAA

Example workflow: protein–ligand complex with binding affinity

This walkthrough uses Boltz-2’s built-in example to predict a protein bound to a small molecule and estimate binding affinity.

1. Open Boltz-2

Go to Boltz-2, or click the Boltz-2 card on All Tools.

Tool info — Show Outputs, Show Settings, Example Results, API Docs, Batch Mode, and Details. Click Example Results to see a finished example without running anything.
Examples — load a ready-made input: 8JYR – Antibody complex, 7QF4 – Protein complex, or Protein-Ligand Complex + Affinity.
Chains — each chain has a type (such as Protein or Ligand), a number of Copies, and a chain ID.

Click Protein-Ligand Complex + Affinity.

2. Check the protein chain

Chain A is filled in with a 384-residue protein sequence (1). To use your own protein, paste a sequence, click Import FASTA, or choose from Tamarind Bio proteins.

3. Check the ligand

Chain B is a Ligand (1). Enter it as a SMILES string or a CCD code (2) (use CCD codes for ions). A 2D drawing of the molecule appears below. You can also pick from Tamarind Bio compounds.

4. Add a restraint (optional)

If you know where the ligand binds, add a restraint (1). The example uses a Pocket restraint:

Binder Chain: Chain B (the ligand)
Pocket Chain: Chain A (the protein)
Pocket residues: 45
Max Distance (Å): 5

You can also click and drag over the sequence to select pocket residues. Use Add chain or Import FASTA (2) to add more chains.

5. Turn on affinity and set samples

Check Predict Affinity and set the Binder Chain to the ligand (Chain B). Affinity prediction only works for small-molecule ligands.
Set Number of Samples (the example uses 5). More samples give more chances to find the right pose but take longer.
Check Estimated Runtime and Est. Compute Cost (about 2 minutes and 0.15 weighted hours for this example).
Click Submit.

Optional settings include a Seed (for reproducible results), a Weighted Hours Budget, and Max Runtime.

6. Read your results

Open the job from My Results (or click Example Results on the tool page to see this example’s output).

Job summary — status, tool, time elapsed, compute hours used, source (app or API), and project.
Download / Resubmit — download all outputs, or resubmit with the same or changed settings.
Tabs — Results, Files, Job Settings, Logs, and Citations.
3D viewer — choose a model file and turn on Show pLDDT to color the structure by confidence (blue is low, red is high).

Scroll down to the metrics table:

Viewing — choose which output file to view, such as metrics.csv.
Table tools — Analyze CSV, Create Chart, Add filter, Download CSV, and Full screen.
Metrics — one row per sample, sorted by confidence. Columns include confidence score, pTM, ipTM, complex pLDDT, ligand ipTM, predicted binding affinity, and binding probability.

In this example, the top sample has a confidence score of about 0.93, pTM of about 0.93, and ipTM of about 0.90, so the overall fold and the ligand placement are both confidently predicted.

7. Next steps

Run more inputs: Use Batch Mode to submit many inputs at once (for example, several ligands against the same protein), then compare predicted affinities.
Dock or rescore: Compare with docking tools such as GNINA or Autodock Vina.
Score the structure: Send the best model to Scoring for properties such as solubility, aggregation, or pKa.
Refine: Run molecular dynamics (OpenMM or GROMACS) to relax the structure or check stability.

Learn more

Scoring — score predicted structures for many properties at once
API — run structure prediction programmatically and at scale
Compute Hours
Manage Files

References

MIT and Recursion release Boltz-2 — Bio-IT World, June 2025
OpenFold3 release — Business Wire, October 2025
Protenix: Advancing structure prediction through a comprehensive AlphaFold3 reproduction — bioRxiv, 2025
How to assess the quality of AlphaFold 3 predictions — EMBL-EBI Training

Tasks

Tools

Platform

Agent

Support