Uni-Mol

Overview

Uni-Mol is a molecular pretraining framework intended to bring three-dimensional structural information into molecular modeling and drug-design workflows. Its official repository also hosts related but distinct methods: Uni-Mol+, Uni-Mol Tools, Uni-Mol Docking V2 and Uni-Mol2. These components serve different purposes, so selecting a task-specific implementation is more useful than treating the repository as a single prediction application.

The original Uni-Mol framework includes separate pretrained models for molecules and protein pockets. The README describes training on 209 million molecular conformations and 3 million candidate protein pockets. The models can be used independently or combined for protein–ligand binding tasks. Documented applications include molecular property prediction, molecular representation extraction, binding-pose prediction and molecular conformation generation. Depending on the component, results therefore include representations, predicted properties or proposed three-dimensional structures.

For quantum chemical modeling, Uni-Mol+ starts from a two-dimensional molecular graph, generates an initial three-dimensional conformation using an inexpensive method such as RDKit, iteratively refines that conformation and predicts quantum chemical properties. Uni-Mol Tools provides wrappers for representation and property-prediction workflows; Uni-Mol Docking targets ligand docking with supplied pockets. Uni-Mol2 uses a two-track transformer to integrate atomic, graph and geometric information.

The supplied top-level README is a navigation and research overview, not a complete input-schema or deployment specification. Component-specific requirements, supported property targets and computational needs must be checked in the linked documentation before evaluation. Repository code licensing should not be treated as confirmation of terms for every dataset, model weight or hosted service.

Key Features

• Separate molecular and protein-pocket pretrained models that can be used independently or combined for protein–ligand binding tasks.
• Molecular property prediction, molecular conformation generation and protein–ligand binding-pose prediction within the original Uni-Mol framework.
• Uni-Mol+ workflow from a 2D molecular graph through initial 3D conformation generation and iterative geometry refinement to quantum chemical property prediction.
• Uni-Mol Tools wrappers for molecular representation extraction, property prediction and downstream workflows, with README-reported support for Uni-Mol2 representations and fine-tuning.
• Uni-Mol Docking V2 for predicting ligand binding poses with given protein pockets.
• Uni-Mol2 molecular pretraining architecture combining atomic, graph and geometric features through a two-track transformer.

Use Cases

• Intended evaluation: assess Uni-Mol Tools representations and property predictions on a molecular dataset with known labels and a predefined validation split.
• Intended evaluation: examine Uni-Mol+ geometry refinement and quantum chemical property predictions against reference structures and properties.
• Intended evaluation: assess Uni-Mol Docking V2 poses for ligands with supplied protein pockets, comparing predicted structures with reference complexes.
• Intended evaluation: compare original Uni-Mol and Uni-Mol2 representations for a selected downstream molecular task using the same evaluation protocol.

How to Use

1. Start with the official repository. Use its navigation table to choose the original Uni-Mol framework, Uni-Mol+, Uni-Mol Tools, Uni-Mol Docking V2 or Uni-Mol2 according to the intended task.
2. Read the selected component’s linked subfolder instructions before preparing inputs. The top-level excerpt does not specify complete file formats, dependencies or hardware requirements, so confirm these rather than assuming a shared setup.
3. For representation extraction or property prediction, consult the Uni-Mol Tools documentation. Follow its documented setup and examples, and check which model and property workflow apply to your dataset.
4. Prepare a small evaluation set with suitable references: measured properties, reference conformations or protein–ligand complexes. For Uni-Mol+, account for its graph-to-conformation workflow; for docking, supply the required pocket information.
5. Evaluate outputs against the references using a predefined protocol. Consult the Uni-Mol paper for research context and the repository LICENSE for code terms; check weights, data and service terms separately.