Description
Uni-Mol is a comprehensive framework for 3D molecular representation learning, designed to enhance performance across a wide spectrum of downstream applications in drug design and molecular science. The core Uni-Mol model is a universal 3D molecular pretraining framework that significantly expands representation capacity and application scope. It comprises two pre-trained models: one trained on 209 million molecular 3D conformations and another on 3 million candidate protein pocket data. These models can be used independently or combined for protein-ligand binding tasks. Uni-Mol has demonstrated superior performance on 14 out of 15 molecular property prediction tasks and achieved exceptional accuracy in 3D spatial tasks, including protein-ligand binding pose prediction and molecular conformation generation.
Read More
Uni-Mol+ focuses on highly accurate quantum chemical property prediction. It takes a 2D molecular graph, generates an initial 3D conformation, iteratively optimizes it to its equilibrium conformation, and then predicts quantum chemical properties. Uni-Mol+ has shown significant improvements over previous state-of-the-art methods on benchmarks like PCQM4MV2 and OC20.
Uni-Mol Tools provides user-friendly wrappers for automatic property prediction, representation, and other downstream tasks using Uni-Mol. It can be installed via pip (`pip install unimol-tools`) and offers tools for molecular property prediction, representation generation, and various other applications.
Uni-Mol Docking V2 advances protein-ligand binding pose prediction, achieving industry-leading performance comparable to AlphaFold3. It accurately predicts binding poses for a high percentage of ligands, circumventing common issues like chirality inversions and steric clashes found in previous machine learning models. A service for Uni-Mol Docking V2 is available.
Uni-Mol2 is a flexible and scalable molecular pretraining model, featuring a two-track transformer architecture that integrates atomic, graph, and geometric structural features. It offers models ranging from 84 million to 1.1 billion parameters, allowing users to select the size best suited for their molecular tasks. Uni-Mol2 has been scaled up to 1.1 billion parameters, making it one of the largest molecular pretraining models to date, trained on 800 million conformations. Its availability in `unimol_tools` allows for easy utilization of its representation and fine-tuning capabilities.
- CategoryScientific Discovery & Lab Tools
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Uni-Mol's Core Features
Universal 3D molecular representation learning framework
Pre-trained models for molecular and protein pocket data
High performance in molecular property prediction tasks
Accurate 3D spatial task performance (binding pose, conformation generation)
Quantum chemical property prediction with Uni-Mol+
Iterative 3D conformation optimization for QC prediction
Easy-to-use wrappers for property prediction and representation with Uni-Mol Tools
Industry-leading protein-ligand binding pose prediction with Uni-Mol Docking V2
Scalable molecular pretraining models (Uni-Mol2) from 84M to 1.1B parameters
Two-track transformer architecture integrating atomic, graph, and geometric features
Support for various downstream molecular tasks
Released with multiple research papers and publications
Getting Started with Uni-Mol
Access model: Obtain access to the Uni-Mol models and tools.
Set up environment: Configure your local environment with necessary dependencies.
Integrate via API: Utilize provided APIs or wrappers for programmatic access.
Load pre-trained models: Load specific Uni-Mol models for your chosen task.
Prepare input data: Format your molecular or protein-ligand data according to requirements.
Run predictions: Execute property predictions, docking simulations, or representation generation.
Evaluate results: Analyze the output and assess performance for your application.
Uni-Mol's Use Cases
- Molecular Property Prediction
- Binding Pose Prediction
- Quantum Chemical Modeling
- Conformation Generation
- Molecular Representation Learning
- Drug Design
- Materials Science
