Multi-Modal Foundation · Merck
MMPT-FM
Generates matched molecular pair transformations and chemical analogues for medicinal chemistry design.
What MMPT-FM is
MMPT-FM (Matched Molecular Pair Transformation Foundation Model) is a generative sequence-to-sequence model developed by Merck & Co., Inc. and Emory University to assist with medicinal chemistry analogue design. Designing small-molecule analogues traditionally depends heavily on human intuition and experimental matched pairs. MMPT-FM models context-independent variable-to-variable chemical modifications, known as matched molecular pair transformations (MMPTs), to capture chemical optimisation patterns across varied compound series.
Alongside the primary MMPT-FM model, the framework includes three matched molecular pair (MMP) variants: molecule-to-molecule (MMP-M2M), molecule-to-transformation (MMP-M2T), and constant-to-variable (MMP-C2V). Together, these models provide a scalable, interpretable method for exploring chemical space, suggesting plausible edits, and supporting both targeted optimisation and automated enumeration workflows.
How it works
The model employs a 220-million-parameter encoder–decoder Transformer architecture initialised from ChemT5, a T5 variant pretrained on chemistry-domain text and molecular structures. Chemical structures and substructural edits are represented using SMILES and SMARTS strings. Training is conducted under a supervised sequence-to-sequence objective using cross-entropy loss, optimised with a batch size of 64 and a learning rate of 5 × 10⁻⁴.
Training examples are generated using the MMPDB package applied to small molecules from the ChEMBL database. The raw data undergo filtering based on drug-likeness rules (rule_of_druglike_soft), a molecular weight minimum of 200 Da, and the removal of unwanted structural alerts via the curated Walters alert collection.
Key facts
- Architecture
- Encoder–decoder Transformer (ChemT5 base)
- Parameter count
- 220M (per variant)
- Chemical representations
- SMILES and SMARTS
- Training data
- ChEMBL database processed with MMPDB
- Training hardware
- 4 × NVIDIA A6000 GPUs (MMPT-FM); 4 × NVIDIA H100 GPUs (MMP variants)
What you can use it for
- Site-specific analogue design
- Proposing chemically valid functional group modifications or variable replacements at a specific, user-defined edit site within a lead compound.
- Transformation and pair generation
- Generating valid matched molecular pair transformations or full molecular pairs for compound ideation.
- High-throughput virtual screening and enumeration
- Integrating into computational chemistry pipelines to systematically expand candidate libraries for screening campaigns.
- Retrieval-augmented generation base model
- Serving as the generative foundation within retrieval-augmented generation frameworks (MMPT-RAG) to condition proposals on specific reference datasets.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Encodes medicinal chemistry intuition into context-independent, generalisable rules across diverse chemical series.
- Evaluated on prospective-style patent series from PMV Pharmaceuticals, demonstrating utility in cross-patent analogue generation.
- Built upon ChemT5 to exploit chemistry-specific sequence representations directly in SMILES and SMARTS formats.
Watch out for
- Performance may degrade when generating analogues within underrepresented or novel chemical domains.
- Relies entirely on historical transformations and inherits the reporting biases of the ChEMBL medicinal chemistry literature.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · mit
The mit licence carries terms worth reading before you depend on it.
Questions
Can I use MMPT-FM commercially?
Check the licence before commercial use. The mit licence carries terms worth reading before you depend on it.
What is the difference between MMPT-FM and its MMP variants?
MMPT-FM learns variable-to-variable transformations. The accompanying variants specialise in alternate input-output formulations: MMP-M2M operates from molecule to molecule, MMP-M2T maps whole molecules to transformations, and MMP-C2V maps fixed molecular constants to variable functional groups.
What quality filters are applied to the training compounds?
Molecules extracted from ChEMBL are restricted to a molecular weight of at least 200 Da, filtered for drug-likeness using soft drug-like rules, and stripped of structural alerts using the Walters alert list.
How are the generated analogues evaluated?
Model outputs are evaluated by measuring chemical validity, novelty relative to the training set, and retrieval recall across held-out ChEMBL test sets and patent benchmarks.
How to cite it
Pang, H.-W., Zhang, P. Z., Pan, B., Zhao, L., Yu, X., & Zhang, L. (2026). Scalable and Generalizable Analog Design via Learning Medicinal Chemistry Intuition from Matched Molecular Pair Transformations. ChemRxiv. doi:10.26434/chemrxiv.15001722
Written from MMPT-FM's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.