Compare · 23 models
Chemistry and ADMET prediction models you can run online
Computational tools and models in this category evaluate small-molecule ADMET profiles, predict physicochemical properties and binding affinities, generate novel structures, and model chemical reactions.
What chemistry and admet prediction models do
Small-molecule drug discovery and chemical evaluation require predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) liabilities, fundamental physicochemical parameters, and target binding affinities before conducting experimental assays. Computational approaches in this domain address these challenges by mapping chemical representations into numerical property predictions, multi-objective safety scores, or synthetic pathways.
Most methods rely on text-based representations such as SMILES strings, chemical graphs, or matched molecular pairs. Implementations range from deterministic cheminformatics calculators and directed message passing neural networks to graph transformers and large chemical language models. Several models also incorporate biological partner data, such as target protein amino-acid sequences, to estimate binding affinities or simulate co-folding.
Compared
- Molecular Property CalculatorCorollary LabsComputes physicochemical descriptors and druglikeness rule compliance for small molecules from SMILES strings.Commercial OKInstant
- MolFormerIBM ResearchA chemical language transformer that extracts molecular embeddings from SMILES strings for property prediction.Commercial OKReady
- MolMIMNVIDIA BioNeMoMolMIM generates and optimises small molecules toward target properties using a smooth latent space.Commercial OKReady
- ADMET-AISwanson et al., StanfordPredicts 52 small-molecule ADMET endpoints simultaneously using Chemprop D-MPNN ensembles.Commercial OKOn demand
- Chemprop D-MPNN — toxicity panelSwanson et al., Stanford · Chemprop, MITPredicts small-molecule toxicity endpoints and structural alerts from chemical structures.Commercial OKOn demand
- ether0futurehouseA 24-billion parameter language model that reasons in English to design and modify molecular structures as SMILES.Check termsOn demand
- MAMMALIBM ResearchA biomedical foundation model predicting binding affinity, blood–brain barrier permeability, and solubility.Check termsOn demand
- MIST 1.8B basemist-modelsA 1.8-billion-parameter foundation model using masked language modelling on molecular SMILES for chemistry prediction.Check termsOn demand
- MIST 28M basemist-modelsA 28M-parameter molecular language model pretrained on SMILES strings for molecular property prediction.Check termsOn demand
- MIST mixturesmist-modelsPredicts thermodynamic excess properties and mixing behaviour of binary chemical mixtures from molecular SMILES.Check termsOn demand
- Reaction prediction and fingerprintsSagawa & Kaneko · IBM RXNPredicts forward reactions and retrosynthetic routes for small molecules, and generates reaction fingerprints.Commercial OKOn demand
- AQAffinitySandboxAQAn open-source protein–ligand binding affinity prediction head built on OpenFold3 for drug discovery.See its pageNot yet
- CYP Inhibition ModelopenadmetA multi-task baseline model predicting pIC50 inhibition values across four cytochrome P450 isoforms.See its pageNot yet
- GROVERTsinghua UniversityA molecular graph transformer predicting ADMET endpoints from small-molecule SMILES strings.Commercial OKNot yet
- LULA-1.1omtxA sequence-only model predicting protein-ligand binding scores from amino-acid sequences and SMILES strings.See its pageNot yet
- MegaMolBARTNVIDIA / BioNeMoGenerates SMILES strings and predicts forward reaction products for chemical design and synthesis planning.Commercial OKNot yet
- MMPT-FMMerckGenerates matched molecular pair transformations and chemical analogues for medicinal chemistry design.See its pageNot yet
- NessorecursionpharmaPredicts protein–ligand binding affinities and binder status from protein sequences and small molecules.See its pageNot yet
- PXR Activation ModelopenadmetA baseline CheMeleon model predicting pregnane X receptor activation potency to assess drug-drug interaction risks.See its pageNot yet
- REINVENT 4AstraZenecaGenerates and optimises small molecules against multi-objective property profiles using reinforcement learning.Commercial OKNot yet
- TxGemma 27B PredictgoogleA 27B language model predicting drug, target, and molecular properties across diverse therapeutic modalities.See its pageNot yet
- TxGemma 2BgoogleA 2-billion-parameter language model fine-tuned for therapeutic property prediction across drug discovery tasks.See its pageNot yet
- TxGemma 9B PredictgoogleA 9B-parameter language model fine-tuned on Therapeutics Data Commons for therapeutic property prediction.See its pageNot yet
Plus 39 task-specific fine-tunes of the models above, each with its own page in the catalogue.
How to choose
For property screening and ADMET profiling, selection depends on required endpoints and hardware constraints. ADMET-AI predicts 52 ADMET endpoints simultaneously using Chemprop D-MPNN ensembles on standard CPU hardware and contextualises outputs against approved drugs. Chemprop D-MPNN — toxicity panel isolates the safety endpoints from this ensemble, covering Ames, hERG, DILI, ClinTox, LD50, and Tox21 assays alongside structural alert filters like PAINS and Brenk. If only deterministic physicochemical rules are needed, Molecular Property Calculator computes Lipinski and Veber rule compliance directly from SMILES without model uncertainty. For graph-based ADMET predictions requiring GPU acceleration, GROVER uses a graph transformer pre-trained on 10 million molecules but requires 16 GB of GPU memory. For metabolism-specific baseline risks, the openadmet CYP Inhibition Model and PXR Activation Model predict pIC50 and pEC50 values, although developers note their performance is poor and they lack uncertainty estimates.
For general therapeutic prediction across diverse biological and chemical tasks, Google's TxGemma models process structured Therapeutics Data Commons prompts across up to 66 tasks. TxGemma 2B offers a compact footprint suitable for fine-tuning, while TxGemma 9B Predict and TxGemma 27B Predict offer higher benchmark performance across 50 of the 66 tasks without conversational dialogue overhead. When building custom property-prediction heads, MolFormer provides pre-trained linear attention embeddings from 1.1 billion molecules for structures up to 200 atoms and 202 tokens. Alternatively, MIST 28M base and MIST 1.8B base offer pre-trained representations from Enamine REAL Space, though both require strictly Kekulised SMILES and a Rust-dependent SMIRK tokenizer, with MIST 1.8B needing 24 GB of GPU memory. For binary liquid systems, MIST mixtures models thermodynamic excess properties and mixing behaviour.
When predicting protein-ligand interactions, tools vary significantly in their structural requirements. LULA-1.1 predicts binding scores from primary protein amino-acid sequences and ligand SMILES without 3D coordinate inputs, docking steps, or folding pipelines. Nesso evaluates interactions by employing coarse-grained co-folding directly from protein sequences and ligands formatted as SMILES, CCD, or SDF. AQAffinity acts as an OpenFold3 affinity head replicating the Boltz-2 module, relying on OpenFold3 checkpoints and kalign2 sequence alignments. MAMMAL handles multi-modal biological inputs, including amino-acid sequences, SMILES, and single-cell gene expression data for binding affinity, blood-brain barrier permeability, and solubility, requiring 24 GB of GPU memory.
For molecular generation and synthesis planning, ether0 applies a 24-billion-parameter language model to generate SMILES conforming to complex multi-property constraints, though it cannot answer forward numerical property queries. MolMIM navigates a smooth mutual-information latent space from a seed SMILES to optimise toward targets such as QED or logP, whereas REINVENT 4 uses reinforcement learning over a SMILES generator to optimise against custom multi-objective scoring functions; both require 16 GB of GPU memory. MMPT-FM generates chemical analogues by applying matched molecular pair transformations trained on ChEMBL data. For synthetic routes, Reaction prediction and fingerprints provides sequence-to-sequence forward reaction prediction and ranked retrosynthetic proposals trained on USPTO patents, while MegaMolBART combines forward reaction prediction with autoregressive de novo generation.
Questions
Can I run these models without a GPU or installing anything?
11 of the 23 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
Which models predict ADMET profiles or chemical descriptors on standard CPU hardware?
ADMET-AI and Chemprop D-MPNN — toxicity panel both run on standard CPU hardware without requiring GPU acceleration. Molecular Property Calculator also executes deterministically on CPU to calculate physicochemical descriptors and rule compliance.
How do LULA-1.1, Nesso, and AQAffinity differ in predicting protein-ligand binding?
LULA-1.1 is a sequence-only model that scores binding directly from the target amino-acid sequence and ligand SMILES without structural coordinates, docking, or protein folding. Nesso uses coarse-grained co-folding from protein sequences and ligands provided in SMILES, CCD, or SDF formats. AQAffinity acts as an affinity prediction head integrated into OpenFold3 and requires structural checkpoints as well as external kalign2 alignment software.
What input formatting is required to use the MIST family models?
MIST 28M base, MIST 1.8B base, and MIST mixtures all strictly require SMILES strings to be converted into Kekulised form before tokenisation. Their SMIRK tokenizer also requires a local Rust installation to build and run dependencies.
Can ether0 be used to predict the numerical property values of existing molecules?
No. ether0 is designed to generate molecular structures that satisfy multi-property criteria, but it cannot answer direct forward property prediction queries, such as predicting the numerical pKa of an existing structure.
What is the difference between MolMIM and REINVENT 4 for molecular design?
MolMIM navigates a smooth latent space from a seed molecule toward predefined properties such as QED and logP. REINVENT 4 uses reinforcement learning to guide a SMILES generator against user-defined, composite multi-objective scoring functions. Both models require 16 GB of GPU memory.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.