What, PyTorch can also develop new drugs? Harvard launched this toolkit, 10 lines of code to train the "Drug God" model

Edited by Xiao Xiao
Quantum Bit Report | Public Account QbitAI

Recently, researchers from Harvard University and other institutions have developed an AI "Drug God" toolkit to help accelerate the development of new drugs under the COVID-19 pandemic .

The toolkit, called DeepPurpose , contains not only a COVID-19 biometric dataset, but also 56 cutting-edge AI models .

As a PyTorch-based toolkit, DeepPurpose only requires less than 10 lines of code to train the AI ​​"Medicine God" model.

These models can not only perform virtual screening, but also discover new functions of existing drugs (for example, high blood pressure drugs can treat Alzheimer's disease) .

Let’s take a look at how it works.

56 cutting-edge models with full functionality

DeepPurpose consists of two encoders, which are used to generate embeddings of drug molecules and proteins, respectively , which is the mapping in the deep learning process.

Subsequently, these two encoders are concatenated into the decoder to predict the binding affinity of the two, as shown in the figure below.

During this period, the input of the model is a drug-target pair , and the output is a score indicating the binding activity of the drug-target pair.

Of course, DeepPurpose is a toolkit after all, so whether it is a drug molecule or a protein, there is more than one type of encoder .

For drug molecules, DeepPurpose provides 8 encoders .

Among these encoders, there are those used to construct molecular structure diagrams, those that convert drawn molecules into binary numbers, those that are used to obtain sequence order information, etc. The models are different.

For target proteins, DeepPurpose also provides 7 encoders . Compared with the chemistry and informatics of drugs, the encoders' conversion of target proteins focuses more on biological information.

In other words, DeepPurpose can provide a total of 7*8=56 models, many of which are very novel and cutting-edge and worth purchasing.

So, how do you get started with DeepPurpose?

Get started with AI "Drug God" in 10 steps

In fact, training a new drug development model requires the following steps, each of which can be implemented with only one line of code , and all these steps added up to no more than 10 steps .

Let’s take a look at the steps this model goes through:

1. Data loading
2. Specify encoder
3. Split data set, encode
4. Generate model configuration file
5. Initialize model
6. Train model
7. Repurposing old drugs/virtual screening
8. Save/load model

Among them, the two most critical functions of DeepPurpose, new uses of old drugs and virtual screening can be realized after training. It can be seen that DeepPurpose automatically generates the affinity of drugs and sorts them from low to high.

In this way, the screening scope of high-throughput molecules can be quickly narrowed down (if the affinity is 0, then you really don't need to consider it) .

As for virtual screening, it works similarly and generates a ranking list similar to the one above.

Not only that, this AI model also includes several other cases, such as new uses of old drugs for SARS-CoV2 3CLPro, pre-trained models, etc.

In addition, in response to the recent COVID-19 pandemic, DeepPurpose also includes the COVID-19 open source dataset collected by MIT .

For these data, there are corresponding functions in the toolkit that can be directly referenced.

The framework of this toolkit is based on the principles of drug development .

Target protein: the object of drug action

The most fundamental principle of drug screening is usually to determine the affinity between drug molecules and target proteins (the target of drug action) .

Why protein?

In fact, this is because the causes of some diseases (such as cancer and tumors) are usually related to a certain type of protein. If this protein can be found and "regulated" with drugs, the disease can be cured.

△ The picture comes from flickr

For example, the communication between cells depends on the glycoproteins on the cell membrane. The cause of a certain disease may be the overexpression of glycoproteins on a certain type of cell .

This glycoprotein is called the target protein in the disease process .

However, it is not easy to find drugs that can be used to regulate a certain target protein. After all, not every compound can stick well to the target protein.

On this basis, the researchers developed DeepPurpose, a toolkit that can be used to predict the affinity between drug molecules and target proteins. The professional academic term is drug-target interaction (DTI) , abbreviated as DTI .

There are reasons behind choosing to use AI to assist in new drug development.

AI helps new drug development

In fact, it takes about 15 years , or even longer, for a pharmaceutical company to develop a new drug .

During this period, the research and development phase alone will take 2-10 years .

The purpose of the research and development stage is to screen out new compounds with therapeutic potential. That is to say, each compound needs to be experimented with through trial and error.

This process is not only tedious, but also requires a huge amount of work, requiring a lot of manpower and financial resources.

If AI is used to complete the drug screening process, it will play a significant role in accelerating the development of new drugs.

about the author

The first author of the paper, Huang Kexin, received a double degree in mathematics and computer science from New York University. He is currently studying for a master's degree in medical big data at Harvard University.

Huang Kexin's research direction is mainly the application of graph neural networks (GNN) in new drug development and medical texts (such as electronic medical records) .

In addition, Tianfan Fu, Lucas Glass, Marinka Zitnik, Cao Xiao and Jimeng Sun also participated in the research.

Portal

Paper link:
https://arxiv.org/abs/2004.08919

Project link:
https://github.com/kexinhuang12345/DeepPurpose

Huang Kexin's homepage:
https://www.kexinhuang.com/