Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

Drug combination therapy is a well-established strategy for disease treatment.
Computational approaches, specifically deep learning models, can be used to discover synergistic combinations.
Data from various datasets was collected and used to generate informative representations and features.
A message-passing graph was built to propagate information and graph structure learning flexibility.
State-of-the-art results were achieved in comparison with other deep learning-based methods.
10% improvement over previous methods was observed in a test on an independent set.
Robust against unseen drugs and surpasses almost 15% AU ROC compared to the second-best model.

Drug combination therapy has advantages such as improved efficacy, reduced side effects and host toxicity, and can overcome drug resistance
Used to treat complex diseases such as HIV, virus infections, and cancer
Combination of two drugs, CBS and NAC, suppresses SARS-CoV-2 virus and reduces viral loads
Traditional methods for drug combination discovery are time-consuming and costly
High-throughput drug screening (HTS) is a sensitive and fast synchronous experiment
Public databases provide antifungal drug combinations data and a large HTS synergy study
In vivo and in vitro experiments are not always consistent
Machine learning models and neural networks are effective and promising in finding novel drug combinations
Deep learning models and traditional machine learning models are used to compare and evaluate drug combinations

Preprocessing section: datasets manipulation and feature pre-training
Heterogeneous graph section: graph construction, graph neural network, and synergistic prediction head information
Graph structure learning section: Drug-Target predictive module, Drug-Drug interaction module, and graph structure learning details
Self-training and inference section: self-training strategy and inference

Constructed a heterogeneous graph with nine edge types and three node types
Drug-Drug Similarity edge computed by measuring distance between KPGT drug embeddings and fingerprints Tanimoto similarity
Graph neural networks built upon constructed graph to conduct message passing
MLP modules used to form unified vectors with length 512
Graph attention network used to propagate information
Synergistic Prediction Head used to extract final drug representations and cell line representations

Graph Structure Learning (GSL) is a technique for learning adaptive graphs when only limited data is available.
GSL is used to generate pseudo edges to conduct inductive inference for unseen drugs.
Drug-Target Interaction (DTI) and Drug-Drug Interaction (DDI) are the two most informative edge types for unseen drugs.
DTI and DDI modules are pre-trained separately and fine-tuned into the framework.
Refined graph is passed through a graph neural network and enters a synergistic prediction head.
Framework is tuned according to a binary cross entropy function.

Self-training can be used to improve limited-data supervised learning tasks.
There are over 0.7 billion possible cases in the combinatorial search space, but only 0.1% of the space is labeled data.
A new training set is created by merging the original dataset with entries whose confidence scores are greater than 0.8.
The model is retrained until it cannot gain improvement.

Model is a fully differentiable end-to-end method to predict drug synergistic combinations from drug sequences and cell line representation
Techniques used include large-scale pre-training, adaptive self-training, and prediction module
Evaluated model against state-of-the-art methods on DrugComb dataset
Model achieved best result across all metrics except for ACC and precision
Model outperformed DeepDDS by around 2% and classical ml methods by almost 20%
Model achieved highest prediction score and DeepDDs/MR-GNN achieved similar second-best values
Evaluated model on AstraZeneca dataset, model achieved highest result across all metrics
Model outperformed DeepDDS by around 14%
Evaluated model on independent drugs and cell lines, model achieved high AU ROC and AU PRC over 80%
Ablation study conducted to evaluate effectiveness of submodules