Use Mol-1 Online
Commercially Available Mol-1 No-Code Web Server
Mol-1: Open Frontier Biomolecular Structure Prediction for Antibodies, Ligands, Peptides and Molecular Glues
Mol-1 is a biomolecular foundation model for predicting molecular complexes and for incorporating biological knowledge into those predictions, introduced in the technical report "Mol-1: Open Frontier Biomolecular Intelligence" from Molence Bio. It is derived from Protenix and the AlphaFold 3 ecosystem, and it targets key modalities in drug discovery: antibody-antigen, protein-ligand, protein-peptide and molecular-glue complexes.
Mol-1 comprises a base model with a single checkpoint shared across tasks and molecular modalities, plus a fine-tuned constraint-conditioned version, Mol-1-constraint, that accepts binding-site (epitope or interface) information to guide predictions. The base model is released openly on GitHub under the Apache License 2.0, and both versions are available through the developers' web server. The report also describes a long-running agent for very large assemblies and a de novo VHH design workflow built on an earlier Mol-1 checkpoint.
Property | Detail |
|---|---|
Task | Biomolecular complex structure prediction |
Inputs | JSON job file with protein, DNA, RNA, ligand and ion entries, optional covalent bonds, residue modifications and templates |
Variants | Mol-1 (base) and Mol-1-constraint (with binding-site constraints) |
Outputs | CIF structures per sample with confidence JSON (pLDDT, PAE, ranking score) |
Weights | mol-1-base checkpoint, about 2.2 GB |
License | Apache License 2.0 (code repository) |
Mol-1 Benchmark Results
Antibody-Antigen Prediction
Success rates are top-ranked results with a five-seed budget. At DockQ above 0.23, Mol-1 outperforms the strongest open-source baseline on each of four benchmarks, reaching 75.0% on FoldBench-AB and 57.3% on MolBench-AB-26. With epitope input, Mol-1-constraint reaches 88.3%, 84.2%, 77.5% and 79.4% on FoldBench-AB, AF3-AB, PXMeter-AB and MolBench-AB-26. Deeper recycling reduced the number of seeds needed: about 58.7 to 58.8% on MolBench-AB-26 was reached with 30 cycles and 5 seeds, 20 cycles and 10 seeds, or 10 cycles and 40 seeds.
Protein-Ligand, Protein-Peptide and Molecular-Glue Prediction
Benchmark | Metric | Mol-1 | Mol-1-constraint | Comparison |
|---|---|---|---|---|
Runs N' Poses, lowest-similarity bin | Pose success rate | 36.1% | 63.9% | 24.6% (AlphaFold 3) |
MolBench-Peptide-26 | DockQ above 0.23 / 0.49 / 0.80 | 88.84% / 77.28% / 54.07% | 60.16% at DockQ above 0.80 | Exceeds displayed baselines at each threshold |
MGBench (molecular glue), Seen / High / Low homology | Success rate | 92.2% / 74.6% / 28.6% | Not reported | Low homology: 19.1% for the strongest displayed baseline |
On Runs N' Poses, Mol-1 outperformed all baselines across all eight similarity bins. On PoseBusters checks, Mol-1 passed at high rates, though OpenFold 3 performed slightly better on ligand geometry and energy checks.
Case Studies
Albumin-ALB8 VHH (PDB 8Z8V): DockQ of 0.926 and interface RMSD of 0.521 Å versus a DockQ of 0.005 for OpenDDE.
GITR-Ab#1 Fab (PDB 9P8X): DockQ of 0.926 for the antigen-heavy-chain interface versus 0.041 for OpenDDE.
DDB1-CDK12-DS16 glue (PDB 8BUN): DockQ of 0.809 and ligand RMSD of 1.36 Å, where Protenix-v1 misplaced CDK12 (DockQ 0.020).
14-3-3σ-ERα phosphopeptide with a stabilizer (PDB 8BXM): ligand RMSD of 1.10 Å versus 6.69 Å for Protenix-v1.
Large Assemblies and De Novo VHH Design
Long-running agent: built on Mol-1 subcomplex predictions, it predicts assemblies of up to 60,000 tokens on a single NVIDIA H20 GPU over tens of hours. Eight of ten assemblies reached a TM-score of 0.90 or higher, and nine of ten outperformed previously reported results.
VHH design workflow: using an earlier checkpoint, de novo VHHs yielded binding hits across 17 targets, with initial hit rates of 6% to 63%.
Affinity and function: BCMA VHHs reached SPR affinities as low as 43 pM from a design budget of 30 candidates; designs also showed 4-1BB agonism and CXCR6 antagonism, and experimental feedback raised the 4-1BB functional hit rate from 7% to 85% in a second round.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Structure Prediction with Mol-1 on Tamarind Bio
Antibody and VHH complexes: predict antibody-antigen structures and, with Mol-1-constraint, guide poses with known epitope residues.
Protein-ligand complexes: model ligand binding poses for drug discovery targets.
Protein-peptide complexes: predict peptide binding modes.
Molecular glue ternary complexes: model glue-induced protein-protein interfaces.
No local setup: avoid installing the toolchain, downloading weights and managing GPU memory yourself.
How to Use Mol-1 on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and select Mol-1.
Enter your complex: provide protein, DNA, RNA, ligand or ion chains for the complex you want to predict.
Add optional knowledge: where the tool exposes it, supply binding-site or epitope residues to use the constraint-guided model.
Set sampling options: choose the number of seeds and samples; the report used 5 seeds by 5 samples and a higher recycling depth for antibody-antigen tasks.
Run the job: submit and wait for the predicted structures.
Review results: download the CIF files and confidence scores, and use the top-ranked sample for downstream analysis.
Parameters shown on Tamarind may differ from the open-source command line, so refer to the tool page for what is exposed.
Limitations
Default predictions use no templates, and protein MSAs in the open-source release come from the public ColabFold service.
Peak memory grows close to quadratically with token count, and large inputs need specific memory settings.
Molecular-glue prediction in less familiar protein contexts remains challenging, with 28.6% success in the low-homology group.
The large-assembly agent and the VHH design workflow are separate from the released base model; the design workflow used an earlier checkpoint.
Constraint-guided predictions depend on the quality of the supplied binding-site information.
The report is a technical report, and the repository lists the citation as coming soon.
Source: Molence Bio, "Mol-1: Open Frontier Biomolecular Intelligence," technical report. Code and base weights: github.com/molence/Mol-1 (Apache License 2.0).