A group of researchers from Tsinghua University, Fudan University’s School of Pharmacy, the National University of Singapore, and Zheijang University have developed a novel AI tool to predict drugs’ pharmaceutical properties. The researchers present this tool called MolMapNet in a paper published in Nature Machine Intelligence. The tool analyzes the human-knowledge-based molecular representations to predict drugs’ pharmaceutical properties and can even be used by people with less understanding of computer science or biology.
In recent years, Scientists have built plenty of deep learning tools for various applications such as pharmaceutical drug analysis. Researchers have been training the deep learning models predicting pharmaceuticals’ properties to analyze and learn molecular representations.
The team states that pharmaceutical investigations require learning many molecular characters, especially the rich collection of molecular properties derived from human knowledge. However, it is not easy to teach these molecular properties to an AI. AI tools are generally good at identifying spatially ordered images, but they do not perform well on unordered data such as molecular properties, which significantly impairs their pharmaceuticals analysis performance.
Addressing this issue, the team wanted to improve deep-learning models’ performance for predicting pharmaceutical properties. Improving AI architectures with limited pharmaceutical data is challenging. So they explored the idea of enhancing the way AI reads molecular properties. Their vision is to map unordered molecular properties into ordered images for AI to recognize molecular properties more efficiently.
This tool is also accessible to non-expert users since it does not require fine-tuning of parameters. It has outperformed SOTA AI tools on most of the 26 pharmaceutical benchmark datasets.
The approach follows three steps for improved deep learning prediction of pharmaceutical properties:
- Broadly learning the intrinsic relationships of molecular properties from over 8 million molecules. These relationships may be linked to and thus indicators of many pharmaceutical properties.
- Using a newly developed data transformation technique to map pharmaceuticals’ molecular properties into 2D images, where the pixel layouts indicate the intrinsic relationships between them. These pixel layouts contain essential indicators of pharmaceutical properties that a well-trained deep learning model can capture.
- Training an image-recognition tool to learn the 2D images and using them to predict pharmaceutical properties. The AI tool can capture specific pixel layout patterns characterizing specific pharmaceutical properties, similarly to how AI methods might distinguish between males and females in a picture by looking at some gender-related features.
The team hopes that this newly developed model could significantly speed up pharmaceutical research and help scientists predict different properties efficiently. In the future, they aim to upgrade their model further to apply it in biomedical studies.