Graph Neural Network Pre-training for ADMET Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ADMET prediction platforms for drug development face limitations due to the need for large amounts of annotated training data, which can be insufficient, leading to limited model improvement.
Innovation Solution
A method involving pre-training a graph neural network with unannotated compound data and performing multi-task training using annotated data sets to predict compound properties, including ADMET properties, thereby reducing data acquisition costs and improving prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a deep learning model is applied for ADMET prediction, then prediction capability is provided, but the model requires large amounts of annotated training data which are insufficient
Solution Approach 1:
The patent applies preliminary action by pre-training the graph neural network on large-scale unannotated compound data before the actual ADMET prediction task. This pre-training stage prepares the model to learn general chemical structure representations and patterns, so that when fine-tuned on the limited annotated ADMET data, the model can achieve good prediction performance without requiring large amounts of annotated training data for the final task.
Solution Approach 2:
The training process is segmented into two distinct stages: a pre-training stage using unannotated data to learn general chemical knowledge, and a fine-tuning stage using annotated data to specialize in ADMET predictions. This segmentation allows the model to leverage both large unannotated data for general capabilities and small annotated data for task-specific accuracy, resolving the contradiction between prediction capability and data quantity requirements.
2Measurement precision
If more annotated training data is collected, then model accuracy improves, but data acquisition cost and time increase
Solution Approach 1:
By performing preliminary pre-training on readily available unannotated compound data, the patent reduces the dependency on time-consuming annotation processes. The pre-trained model captures essential chemical structure features that can be directly leveraged for prediction, significantly reducing the amount of annotated data needed and thus accelerating the overall model development timeline while maintaining high accuracy.
Solution Approach 2:
Unannotated compound data serves as an intermediary resource that bridges the gap between available data and required annotated data. The pre-training stage uses this intermediary unannotated data to build a foundation model, which then requires much less annotated data for fine-tuning, effectively mediating between data availability and model accuracy requirements.
3Reliability
If existing deep learning models are used, then ADMET prediction is achieved, but the models are limited by insufficient training data
Solution Approach 1:
The patent creates a universal pre-trained graph neural network that can handle multiple chemical prediction tasks through a single pre-training phase on diverse unannotated compound data. This universal model serves as a foundation that can be adapted to various specific ADMET prediction tasks through fine-tuning, significantly enhancing the model's versatility and improvement potential compared to task-specific models that require separate training for each property.
Solution Approach 2:
The preliminary pre-training on diverse unannotated data creates a robust foundation model with general chemical understanding, enabling the model to be effectively adapted to multiple specific ADMET prediction tasks. This preliminary action establishes a versatile baseline that can be fine-tuned for different properties, greatly enhancing the model's adaptability and future improvement potential.
Data Source
AI summary
A method and apparatus for training a compound property prediction model, a device, a storage medium and a program product. A implementation of the method comprises: acquiring an unannotated compound data set; pre-training a graph neural network using the unannotated compound data set to obtain a pre-trained graph neural network; acquiring a plurality of annotated compound data sets, each annotated compound data set being annotated with one kind of compound property; and performing multi-task training on the pre-trained graph neural network using the plurality of annotated compound data sets, to obtain a compound property prediction model, the compound property prediction model being used to predict a plurality kinds of properties of a compound.