MMPT-FM

Medicinal chemists refine drug candidates largely by making small, localized edits to a lead molecule and observing how those edits shift activity, selectivity, and physicochemical properties. Matched molecular pair (MMP) analysis formalizes this practice by cataloguing pairs of compounds that differ by a single well-defined chemical change, but classical MMP methods are bound to the transformation rules and pairs already present in a curated database and generalize poorly beyond them. MMPT-FM (Matched Molecular Pair Transformation Foundation Model) reframes the task as generative: instead of looking up a transformation, it produces chemically valid molecular edits directly, encoding context-independent variable-to-variable modifications so a learned transformation can be applied at a specified site of a new molecule.

MMPT-FM is a chemical language model that reads and writes molecules as SMILES strings and transformations as SMARTS/SMIRKS patterns. It is released alongside three task-specialized siblings — MMP-M2M (molecule-to-molecule), MMP-M2T (molecule-to-transformation), and MMP-C2V (constant-to-variable) — that cover the common framings of analog design: generating a paired analog, proposing the transformation that links two molecules, and completing a fixed scaffold with a new variable fragment. A retrieval-augmented variant, MMPT-RAG, conditions generation on reference analogs to better recapitulate expert medicinal-chemistry intuition.

MMPT-FM was developed by Merck & Co., Inc. (Rahway, New Jersey; known as MSD outside the US and Canada) together with Emory University, and released in 2026 under the MIT license with accompanying code and preprints.

#Key Features

• Generative MMP transformations: The model produces chemically valid variable-to-variable edits directly rather than retrieving them from a fixed rule database, extending matched-pair design beyond precedented transformations.
• Built on ChemT5: MMPT-FM is a 220M-parameter encoder-decoder transformer fine-tuned from ChemT5, a chemistry-domain pretrained T5, giving it a language-model prior over molecular and reaction strings.
• Directed, promptable edits: Users can specify preferred transformation patterns or edit sites during generation, steering analog design toward a desired region of a molecule.
• Retrieval-augmented generation: The MMPT-RAG configuration supplies reference analogs as context to align outputs with observed medicinal-chemistry practice.
• Task-specialized family: Companion M2M, M2T, and C2V models address distinct analog-design formulations from a shared modeling approach.

#Technical Details

MMPT-FM is a 220M-parameter encoder-decoder transformer fine-tuned from ChemT5. Training pairs were extracted from the ChEMBL database after filtering for drug-likeness, a molecular-weight floor of 200 Da, and removal of compounds matching the Walters structural-alert list; molecules and transformations are represented as SMILES and SMARTS/SMIRKS. Training uses supervised sequence-to-sequence learning with a cross-entropy objective (batch size 64, learning rate 5x10^-4) on 4x NVIDIA A6000 GPUs, and inference uses beam search. The models are evaluated with validity, novelty, and recall metrics against a held-out ChEMBL MMPT test set and patent-derived sets from PMV Pharmaceuticals (2017 and 2021), assessing whether generated transformations recover realistic, chemically sound analogs.

#Applications

MMPT-FM targets medicinal-chemistry lead optimization, where teams iterate on a hit by proposing structurally similar analogs with improved properties. Its directed generation supports analog enumeration at specified edit sites, and its outputs can feed enumeration pipelines, virtual-screening workflows, and retrieval-augmented generation systems. The task-specialized variants let practitioners either generate a paired analog, infer the transformation between two known compounds, or grow a fixed scaffold, matching common lead-optimization questions.

#Impact

MMPT-FM brings a foundation-model and retrieval-augmented framing to matched molecular pair analysis, a long-standing workhorse of medicinal-chemistry SAR, and its release from an industrial pharmaceutical group under a permissive MIT license makes the weights and code broadly reusable. The accompanying methods are described in 2026 arXiv preprints awaiting peer review, and reported results are in-silico validity, novelty, and recall metrics rather than experimentally confirmed activity, so generated transformations are candidate edits that still require synthesis and assay-based validation.