Multi-Modal Foundation · IBM Research
MAMMAL
A biomedical foundation model predicting binding affinity, blood–brain barrier permeability, and solubility.
What MAMMAL is
MAMMAL (Molecular Aligned Multi-Modal Architecture and Language) is a biomedical foundation model developed by IBM Research. Designed to operate across multiple biological and chemical modalities, it processes inputs spanning protein sequences, small molecules, and single-cell gene expression data.
The model targets diverse stages of the drug discovery workflow. Through specific fine-tuned checkpoints, it addresses downstream predictive tasks including protein–ligand binding affinity, blood–brain barrier permeability (BBBP), and protein solubility.
MAMMAL relies on an adaptable prompt syntax that combines tokens and scalars dynamically. This setup enables unified handling of classification, regression, and generation tasks, either within a single biological domain or across cross-domain entities.
How it works
MAMMAL is built on a multi-domain architecture that aligns various modalities through a modular tokenisation process. The model's input prompt format allows users to define specific tokeniser types, such as amino acid sequences or molecular notations, together with input data strings. The model processes these token sequences and attention masks to yield either generated token predictions or class representations.
The 458-million-parameter model was pre-trained on more than 2 billion biological samples spanning proteins, small molecules, and single-cell gene measurements.
Key facts
- Parameters
- 458M
- Modalities
- Protein sequences, small molecules, single-cell gene data
- Training data
- Over 2 billion biological samples
- Release date
- 28 October 2024
- Architecture name
- Molecular Aligned Multi-Modal Architecture and Language (MAMMAL)
- Required GPU memory
- 24 GB
What you can use it for
- Protein–Ligand Binding Affinity Prediction
- Estimates the dissociation constant as pKd for a given protein sequence and small-molecule SMILES string using a BindingDB fine-tune.
- Blood–Brain Barrier Permeability Screening
- Predicts whether a small molecule can cross the blood–brain barrier as a binary classification using a fine-tune trained on MoleculeNet BBBP data.
- Protein Solubility Assessment
- Classifies whether a candidate protein sequence is soluble or insoluble using a dedicated sequence-based fine-tune.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
MAMMAL
Strengths and limits
Good at
- Unifies predictions across multiple distinct biomedical modalities, including small molecules, proteins, and single-cell gene data.
- Supports regression, classification, and generative generation tasks using a single flexible prompting syntax.
- Pre-trained on a broad biological corpus comprising more than 2 billion samples.
Watch out for
- Task-specific checkpoints require distinct combinations of inputs, such as both a sequence and SMILES string for binding affinity or a single sequence for solubility.
- Requires substantial hardware, needing 24 GB of GPU memory.
Inputs and outputs
You provide
Taskoptional
Protein sequenceoptional
Single-letter amino acid code. Whitespace and FASTA headers are stripped automatically.
Ligand SMILESoptional
Canonical SMILES for the small molecule.
You get back
Binding affinity in pKd, or a binary BBBP or protein-solubility classification.
Licence
Check the licence before commercial use · Apache-2.0
The Apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use MAMMAL commercially?
Check the licence before commercial use. The Apache-2.0 licence carries terms worth reading before you depend on it.
How do I run MAMMAL without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 6.0s, and a run costs about 112 credits. Visitors without an account get 1,000 credits a month.
What inputs are required to run MAMMAL?
The inputs depend on the chosen task. Predicting binding affinity requires a protein sequence and a small-molecule ligand SMILES string, blood–brain barrier permeability requires a SMILES string, and protein solubility requires only a protein sequence.
What output formats does MAMMAL provide?
Binding affinity predictions are returned as continuous pKd values, whereas blood–brain barrier permeability and protein solubility are returned as binary classification outcomes.
Which modalities was the foundation model trained on?
The foundation model was pre-trained on proteins, small molecules, and single-cell gene data across over 2 billion samples.
How to cite it
Yoel Shoshan, Moshiko Raboh, Michal Ozery-Flato, Vadim Ratner, Alex Golts, Jeffrey K. Weber, Ella Barkan, Simona Rabinovici-Cohen, Sagi Polaczek, Ido Amos, Ben Shapira, Liam Hazan, Matan Ninio, Sivan Ravid, Michael M. Danziger, Joseph A. Morrone, Parthasarathy Suryanarayanan, Michal Rosen-Zvi, Efrat Hexter. MAMMAL -- Molecular Aligned Multi-Modal Architecture and Language. arXiv:2410.22367, 2024.
Written from MAMMAL'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.