PaddleHelix Documentation

PaddleHelix is a biological computing toolkit that leverages machine learning, specifically deep neural networks, to advance several key domains:

• Drug Discovery: Provides large-scale pre-trained models for compounds and proteins, along with applications for molecular property prediction, drug-target affinity prediction, and molecular generation.
• Vaccine Design: Provides RNA design algorithms, including LinearFold and LinearPartition.
• Precision Medicine: Provides applications for drug-drug synergy prediction.

PaddleHelix is a bio-computing tool that leverages machine learning, specifically deep neural networks, to facilitate research and development in three primary areas:

• Drug Discovery: Includes large-scale pre-training models for compounds and proteins, and applications such as molecular property prediction, drug-target affinity prediction, and molecular generation.
• Vaccine Design: Provides RNA design algorithms, specifically LinearFold and LinearPartition.
• Precision Medicine: Provides applications for studying drug-drug synergy.

PaddleHelix provides several specialized molecular generation models for drug discovery and molecular design tasks. The available models include:

• JT-VAE: Junction Tree Variational Autoencoder.
• SD-VAE: Structure-based Diffusion Variational Autoencoder.
• seq-VAE: Sequence-based Variational Autoencoder.

Each model has its own dedicated documentation and implementation details within the repository.

PretrainGNNs is a module within PaddleHelix designed to improve molecular informatics and drug development efficiency. It addresses the limitations of traditional machine learning in handling irregular molecular graphs and the challenges of missing labels or distribution shifts in training data.

By implementing strategies from the paper "Strategies for Pre-training Graph Neural Networks", this module allows users to:

1. Pre-train Graph Neural Network (GNN) models on data-rich tasks at both the node level and the graph level.
2. Fine-tune these pre-trained models on specific downstream tasks, such as chemical molecular property prediction.

Supported GNN architectures include GIN, GAT, and GCN.

The HelixProtX application consists of the following key files:

• model.py: Contains the model definition.
• model_config.py: Contains model configuration classes.
• build_model.py: Script to create a HelixProtX model checkpoint.
• inference.py: Script for performing inference.

HelixFold3 is a biomolecular structure prediction tool designed to replicate the capabilities of AlphaFold3. It is capable of predicting the structures of proteins, small molecule ligands, and nucleic acids (including DNA and RNA) with accuracy comparable to AlphaFold3.

Users can access HelixFold3 via the HelixFold3 server on the PaddleHelix website in two modes:

1. Visualized interactive interface: For intuitive, user-friendly structural exploration.
2. API-based access: For high-throughput, large-scale screening and design workflows.

Licensing Note:

• Free version: Restricted to non-commercial use.
• Paid version: Supports unrestricted commercial applications.

The drug_drug_synergy application in PaddleHelix provides methods for predicting the synergy effects of drug combinations. It currently supports two primary prediction models:

• RGCN: A Relational Graph Convolutional Network approach.
• DTSyn: A specific drug synergy prediction method.

Detailed documentation for each method, including installation and usage, can be found in their respective subdirectories.

ChemRL is a framework focused on effective molecular representation learning to facilitate molecular property prediction, a key task in the drug and material industries. It utilizes advanced neural network architectures and large-scale self-supervised learning strategies to improve performance.

One of the primary implementations is GEM, which is designed to fully utilize molecular geometries for both the network architecture and the self-supervised learning strategy.

HelixFold is an efficient and improved implementation of the AlphaFold 2 protein structure prediction pipeline using PaddlePaddle. It supports both training and inference on GPU and DCU (Deep Computing Unit) hardware.

Key advantages include:

• Reduced Training Time: Can reduce training from ~11 days to 5.12 days, or as low as 2.89 days using hybrid parallelism.
• High Accuracy: Achieves competitive accuracy with the original AlphaFold 2 when training from scratch.
• Long Protein Support: Supports prediction for ultra-long monomer proteins of approximately 6600 amino acids (AA).
• Parallelism Strategies: Utilizes Branch Parallelism (BP), Dynamic Axial Parallelism (DAP), and Data Parallelism (DP) to accelerate computation.