Mol-1

Mol-1 webserver overview

Mol-1 predicts the 3D structure of a complex from its parts: protein chains, DNA and RNA strands, small-molecule ligands and ions, in any combination. Give it a 203-residue protein and the two strands of the DNA it binds, and it returns ranked structures of the whole protein-DNA complex with pLDDT, pTM and ipTM confidence scores.

Molence Bio released Mol-1 in October 2026 as an Apache-2.0 model built on the Protenix codebase. ProteinIQ runs the released base model, mol-1-base version 1.0.0, with Molence's default settings. Molence's constraint-conditioned variant, Mol-1-constraint, and its large-assembly workflow are not publicly released and are not available here.

How Mol-1 differs from other AlphaFold3-style models

Mol-1 follows the AlphaFold3 recipe of a pair representation refined by a trunk and a diffusion module that generates all-atom coordinates. Its trunk is split: 8 lighter blocks without triangle attention at 448 pair channels, then 40 Pairformer blocks at 320 channels. Molence reports that this keeps the fitting accuracy of a full 448-channel trunk at about 60 percent of the compute.

The model was also trained with deep recycling, and its default is 20 recycling cycles where Protenix and OpenDDE use 10. Its training structures from the PDB end on 30 September 2021, so benchmark complexes released later were unseen. In Molence's report the base model reaches a 75.0 percent DockQ > 0.23 success rate on the FoldBench antibody-antigen set with five seeds.

For other complex predictors on ProteinIQ, see Boltz-2, Chai-1 and OpenFold3.

Pricing

Mol-1 is billed by the minute of runtime on the GPU its complex size routes to: 22 credits per minute on A10G, 36 on L40S and 46 on A100-80GB. Completed runs are charged in proportion to elapsed time, rounded up to a whole credit. The minimum reservation is one minute at the routed rate, not a minimum final charge, and unused reserved credits are returned.

Larger complexes run on larger GPUs. The tool page shows an estimate before submission; these examples are for one protein chain at the default settings (one seed, five samples, 20 cycles, 200 steps):

Protein residues	GPU	With MSA search	Without MSA search	Three seeds, with MSA
100	A10G	191 credits	77 credits	242 credits
300	A10G	235 credits	121 credits	374 credits
768	A10G	338 credits	224 credits	683 credits
1,000	L40S	546 credits	360 credits	1,098 credits
1,500	A100-80GB	928 credits	690 credits	2,093 credits

MSA search, template search and more seeds, samples, cycles or steps raise the runtime. The final charge follows the measured runtime, not the estimate.

Inputs

Every molecule card on the form belongs to one predicted complex: a protein, a ligand and a magnesium ion entered together are predicted together. Separate FASTA records become separate entities in that complex, and each DNA strand is its own entity, so a duplex needs both strands.

Input	Accepted formats	Notes
Protein	Pasted sequence or FASTA; .fasta, .fa; RCSB fetch	Uppercase one-letter codes for the 20 standard amino acids, plus X for an unknown residue.
DNA	Pasted sequence or FASTA; .fasta, .fa	Uppercase A, C, G, T, plus N, X, I and U. Single strands only.
RNA	Pasted sequence or FASTA; .fasta, .fa	Uppercase A, C, G, U, plus N, X and I.
Ligand	Pasted SMILES; .smiles, .smi, .sdf, .mol, .mol2, .pdb; PubChem fetch	Structure files must carry 3D coordinates; Mol-1 refuses a 2D drawing. A SMILES string needs no coordinates.
CCD ligand	CCD code, with a picker	One Chemical Component Dictionary code in uppercase, such as ATP or HEM.
Ion	CCD code	The ion's CCD code in uppercase, such as MG or ZN.
MSA and template files	.a3m, .hhr, .json	Paired or unpaired MSAs and template hits, attached to a protein chain in MSA and template files. Every uploaded file must be attached.

Uploaded files are limited to 50 MiB each. Free accounts can submit up to 1,000 residues per job, counted across the protein, DNA and RNA sequences; copies, ligands and supporting files do not count. Paid plans have no plan-wide residue cap. The optional Job name labels the saved run.

Component options

Parameter	Type	Default	Description
Copies	integer	1	From 1 to 10 copies of the component.
Residue position	integer	1 when added	Position of a residue or base modification, counted from 1.
Modification CCD	string	optional	Uppercase CCD code for the modified residue without the CCD_ prefix, such as MSE.
MSA	string	optional	An unpaired A3M alignment for a protein component.

A component with modifications, an attached MSA or a template file must contain a single sequence. Mol-1 has no cyclic flag; a head-to-tail cyclic peptide is defined with a covalent bond between its first and last residues.

Settings

Prediction

Parameter	Type	Default	Description
Seeds	string	101	Comma-separated integers from 0 to 4294967295, such as 101,102,103; one prediction set per seed.
Samples per seed	integer	5	Structures sampled for each seed; minimum 1.
Recycling cycles	integer	20	Pairformer recycling cycles; minimum 1.
Diffusion steps	integer	200	Diffusion sampling steps; minimum 1.

Sequence features

Parameter	Type	Default	Description
Search protein MSAs	boolean	true	Searches the ColabFold MMseqs2 server for protein chains that have no supplied MSA.
Use PDB templates	boolean	false	Uses each protein chain's template file and searches the PDB with HMMER for chains without one.

When any protein chain lacks an MSA, Mol-1 searches all protein chains of that job together, which also builds the paired MSA across chains. Without MSAs, predictions of natural proteins are usually much less accurate. Searched template hits and uploaded .a3m or .hhr hits are limited to PDB entries released on or before 30 September 2021; JSON templates are used as given.

RNA alignments are not available. Mol-1 searches them against tens of gigabytes of RNA databases that are not part of its release, and it reads an uploaded RNA alignment only in that mode.

Advanced settings

Parameter	Type	Default	Description
Full confidence files	boolean	false	Writes per-atom pLDDT and the token-level PAE, PDE and contact probability matrices.
Covalent bonds	records	none	Bonds between a polymer and a ligand, or between two ligands.
MSA and template files	records	none	Attaches an uploaded file to a protein chain by its chain label.

Each covalent bond names two atoms by entity, copy, position and atom. Entities are numbered from 1 in component order, with each FASTA record counted separately. Leaving both copy fields blank bonds corresponding copies, which requires both entities to have the same number of copies. A glycan is entered as one CCD ligand card per sugar joined by bonds: with two NAG cards as entities 3 and 4, entity 3 position 1 atom O4 to entity 4 position 1 atom C1 makes a β(1→4) link.

Each MSA or template file row names the chain label shown on a component card, the file type (Unpaired MSA, Paired MSA or Template hits) and the uploaded file name. A chain can take one file of each type.

Outputs

Result	Contents
Viewer	Interactive display of the predicted structures.
Confidence	One row per prediction: job name, seed, rank, ranking_score, pLDDT, pTM, ipTM, GPDE and file name.
Files	Every structure, confidence file, prepared input, alignment, template hit and log the run wrote.
<job>_seed_<seed>_sample_<rank>.cif	Predicted structure; the B-factor column holds each atom's pLDDT on the 0 to 100 scale.
<job>_seed_<seed>_sample_<rank>_confidence.json	Summary scores, per-chain and chain-pair scores and the clash flag.
<job>_seed_<seed>_sample_<rank>_full_confidence.json	Per-atom and per-token confidence arrays when Full confidence files is on.
mol1_input.json, mol1_provenance.json	The input Mol-1 read, and a record of the run's settings and model version.
mol1.log, conversion.log	The prediction log and, for structure input, the conversion log.

Files are named after the job complex. Within each seed, rank 0 is Mol-1's top-ranked sample. Seeds keep the order given, and predictions from different seeds are not reranked against each other.

Understanding results

The table shows the scores most people sort by, and each _confidence.json holds the rest. Two scale differences trip up readers of these files: plddt runs from 0 to 100, while chain_plddt and the per-atom pLDDT in the full confidence file run from 0 to 1.

Field	Meaning
ranking_score	0.8 × ipTM + 0.2 × pTM - 100 × has_clash; Mol-1 orders samples by it. A value near -100 marks a sample with a severe clash.
plddt	Mean predicted local confidence over all atoms, 0 to 100.
ptm	Predicted TM-score of the whole structure, 0 to 1.
iptm	Interface pTM across chains, 0 to 1; exactly 0.0 for a single chain.
gpde	Global predicted distance error in ångström; lower is better.
has_clash	Whether a severe steric clash between different polymer chains was detected.
num_recycles	Recycling cycles actually run.
chain_pair_iptm	ipTM for each pair of chains; the diagonal is 0.0 because a chain has no interface with itself.

pLDDT reads like it does for AlphaFold: above about 90 is a confident backbone and side-chain placement, 70 to 90 a confident backbone, 50 to 70 low confidence and below 50 usually disorder.

For a complex, the question is usually whether one interface is right, and the matching entry of chain_pair_iptm answers that better than the whole-complex ipTM. The ipTM and pTM and structure prediction confidence guides explain how to combine these scores.

As an example, Mol-1 predicted the 7r6r complex, a 203-residue protein bound to a 21-base-pair DNA duplex, with pLDDT 88.9, pTM 0.865 and ipTM 0.927. Its ranking_score of 0.915 is 0.8 × 0.927 + 0.2 × 0.865 with no clash. An ipTM above 0.9 says Mol-1 is confident about where the protein sits on the DNA, which is the question this complex is usually predicted for.

The PAE, PDE and contact matrices in the full confidence file are indexed by token, not by residue number. A token is one standard residue, or one heavy atom of a ligand or modified residue, so atom_to_token_idx and token_asym_id map rows back to atoms and chains.

Re-running the same input with the same seed can give a slightly different structure, so compare several seeds rather than relying on one sample. Before using a pose, check its geometry with PoseBusters or MolProbity.

Sources

• Molence Bio. Mol-1: Open Frontier Biomolecular Intelligence. Technical report, 2026. https://github.com/molence/Mol-1/blob/main/assets/report.pdf
• Molence Bio. Mol-1 source code, version 1.0.0. https://github.com/molence/Mol-1
• Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493-500 (2024). https://doi.org/10.1038/s41586-024-07487-w