Resources | Xu Lab

E-CloudBind

Mon, 01 Jan 0001 00:00:00 +0000

What you can use

E-CloudBind combines electron-density point clouds and molecular graphs to predict protein–ligand binding affinity. The public repository provides the model implementation, a pretrained weight file, five preprocessed protein–ligand examples and a workflow for generating input features.

Use the original repository for code and instructions. The Nature Communications article also links processed research data on figshare and a fixed code archive on Zenodo.

Start with the provided examples

Read the environment requirements. The repository lists its Python packages and GPU dependencies. Its current example entry point selects an NVIDIA CUDA device.
Inspect an example complex. The five prepared cases show how molecular graphs and point clouds are organized for the model.
Follow the repository’s inference demo. The example loads the provided weights and reports evaluation metrics for the labeled examples.
Read the preprocessing workflow before using new structures. Feature generation includes coordinate preparation, ligand electron-density calculations, point-cloud processing and graph construction.

The supplied demo is an evaluation example. Its reported RMSE and correlation describe those example cases; they are not performance estimates for an arbitrary new dataset.

Read the example script · Find the pretrained weights

Preparing a new input

The documented workflow starts with a protein PDB file and a ligand MOL2 file. It requires three-dimensional structural input, with the protein and ligand prepared in the appropriate coordinate context. A protein sequence and a SMILES string alone are not the documented input to the supplied demo.

Ligand electron-density generation uses xTB and Multiwfn. Protein-pocket point clouds use the representation described in the paper. Review each preprocessing step and its file conventions before extending the examples to a new complex.

The current demo reads known labels to calculate evaluation metrics. To predict unlabeled complexes, an inference workflow must retain the model predictions as per-complex outputs instead of requiring reference labels for evaluation.

Explore the preprocessing scripts

Data, versions and citation

The article’s data statement links the processed training/validation and test data. Follow its references for the underlying datasets and their usage conditions. Cite the paper for the method and record the code version used for a reproducible analysis.

The GitHub repository contains the working project; the Zenodo DOI identifies the archived code release. This page provides an entry into those original resources.

Read the paper and its availability statements

Photosensitizer Dataset & Active-Learning Resources

Mon, 01 Jan 0001 00:00:00 +0000

Explore the molecular data

The public repository contains 14,164 molecular records in two tables, plus 238 donor, acceptor and bridge records used to define molecular building blocks and connections. The molecular tables pair SMILES representations with quantum-chemical energy labels.

File	Records	Contents
Photosensitizers_DA.csv	8,051	Donor–acceptor molecular records and energy labels
Photosensitizers_DAD.csv	6,113	Donor–acceptor–donor molecular records and energy labels
Donor_Acceptor_Bridge_Substructures.csv	238	Building-block SMILES or connection definitions, with type labels

These counts refer to the repository files checked on 7 September 2026. Some bridge entries specify connections, such as a single bond, rather than standalone molecules.

Understand the labels

Column	Meaning	Unit
Index	Record identifier within its source table	—
SMILES	Molecular structure representation	—
ST Gap	S1–T1 energy gap	eV
HL Gap	HOMO–LUMO energy gap	eV
S1	First singlet excited-state energy	eV
T1	First triplet excited-state energy	eV

The repository describes ground-state optimization at the B3LYP/6-31G(d) level, with excited-state characteristics calculated using TD-DFT at the same level of theory. These are computational labels. Keep the source table identity alongside each Index when combining DA and DAD records.

Connect the data to the discovery study

The study explored a molecular space of more than seven million molecules and reported 5,357 potential high-performance photosensitizers. Four photosensitizers were subsequently synthesized and evaluated.

The 14,164 labeled records, the 5,357 shortlisted candidates and the experimentally tested compounds describe different sets. The two data tables should not be presented as a catalog of experimentally validated materials or as the complete candidate shortlist.

The paper’s Supporting Information includes a photosensitizer database and substructure-list archive, together with experimental details.

Open the JACS paper and Supporting Information

Work with the code

The repository includes molecular-space construction, graph-convolutional prediction models, active-learning code and analysis scripts. Its prediction Notebook is a useful entry point for understanding how prepared models are used.

Follow the repository’s model-download and environment instructions before running its examples. For reuse, retain the original records and keep any cleaned or reformatted data as a separately documented derivative.

Credit and cite

Code and data were contributed by Jiali Li and Pengfei Cai. Cite the JACS paper for the study and identify the code/data version used in your work. The Zenodo V1.0.0 archive provides a fixed historical release alongside the original GitHub repository.

Teaching & Tutorials

Mon, 01 Jan 0001 00:00:00 +0000

Courses

Shidang Xu teaches the following courses:

Neural Networks and Deep Learning
Data Mining and Big Data

Teaching and Mentoring

I teach courses in neural networks and deep learning, and in data mining and big data. My mentoring aims to help students become independent researchers who can formulate scientific questions, design methods, and develop meaningful evaluation criteria.

I encourage undergraduate students to explore scientific questions that spark their curiosity and to take an active role in shaping their research. I guide them in turning initial ideas into clear questions, developing suitable methods, and evaluating their findings with care.

PI profile · Our Lab Charter

Research tutorials

Molecular data and active discovery

Start here if you want to understand how a molecular design problem becomes a learning problem. Some familiarity with molecular structures, Python and tabular data will help.

Understand the records. Open the photosensitizer resource guide. Compare the DA and DAD tables, identify the prediction targets and check their units and computational origin.
Trace the representation. Inspect the donor, acceptor and bridge table and the molecular-space construction code. Ask which design choices define the space that can be searched.
Read a prediction example. Follow the Model Predictions Notebook, using the original repository’s model and environment instructions if you run it.
Examine the next-query decision. Read the active-learning implementation alongside the paper. Identify how new calculations enter the learning process.

Question to carry through the example: how would changing the available molecular building blocks change what the discovery system can find, even if the learning algorithm stayed the same?

Protein–ligand representations and prediction

Start here if you want to understand how structural information becomes input to an affinity model. Familiarity with Python, molecular structures and neural-network inference will help.

Inspect the inputs. Read the E-CloudBind resource guide and compare its protein, ligand, point-cloud and graph files.
Trace the feature construction. Follow the preprocessing workflow. Distinguish ligand electron-density generation from protein-pocket point-cloud construction.
Read the inference example. Inspect the demo script. Identify where the model produces predictions and where known labels enter the evaluation.
Separate a prediction from an evaluation. Before adapting the code to a new complex, decide which outputs you need and which checks can be performed without knowing its measured affinity.

Question to carry through the example: which features depend on the supplied three-dimensional geometry, and what would you need to check when that geometry is uncertain?

These routes use the authors’ existing research materials. Consult each original repository for its current dependencies, examples and usage instructions.