A Graph Convolutional Neural Network Approach to Antibiotic Discovery

Publié dans Coder stories

15 avr. 2020

6min

A Graph Convolutional Neural Network Approach to Antibiotic Discovery
auteur.e
Andy Favell

Web editor, digital and content strategist, copywriter and journalist

In an age when bacterial infections are developing resistance to common antibiotics, the discovery of a new and potentially powerful antibiotic is news in itself. But what makes a recent breakthrough truly revolutionary is that the promising molecule—called halicin—was discovered using deep learning.

Chemprop, the neural network behind this breakthrough, was developed by Kyle Swanson and Kevin Yang in 2018, when they were studying for their master’s degrees at MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). It was built from scratch and trained using everyday open-source Python-based technologies.

Swanson and Yang trained their network to identify patterns in the chemical composition of molecules that had common properties, such as solubility or toxicity, and thus to predict if other molecules shared those properties based on their chemical composition. Their original research study was published in July 2019.

The two co-leads at J-Clinic, a collaboration of computer science and life science labs at MIT that focuses on developing AI for healthcare—CSAIL professor Regina Barzilay and bioengineering professor Jim Collins—wanted to investigate if Chemprop could help to discover new antibiotic treatments. To test this hypothesis, Swanson and Yang, from Barzilay’s lab, were hooked up with Jonathan Stokes, a postdoc antimicrobial biochemist working in Collins’s lab, who would lead the investigation.

Initially, their work looked at whether Chemprop could predict if molecules could inhibit the growth of the bacterial infection Escherichia coli (E. coli). Certain strains of E. coli cause severe “stomach” upsets, and outbreaks are common, contagious, and increasingly resistant to common antibiotics. Chemprop not only rose to the challenge but also delivered results that surpassed all participants’ expectations.

Under the hood of Chemprop

Speaking to Behind the Code, Swanson and Yang explained how Chemprop works. “The core model is a type of deep neural network called a graph convolutional neural network, or message passing neural network. These types of neural networks are specifically designed to operate on objects that can be represented as graphs, such as molecules. In the case of molecules, atoms are represented by nodes and bonds are represented by edges in the graph.

“Chemprop’s goal is to build a representation of the molecule that can be used to predict properties like solubility or toxicity. The model works by first identifying simple features of the [molecule’s] atoms and bonds, such as the type of atom—carbon, oxygen, and so on—and the type of bond, such as a single or double bond.

“The model then performs a number of ‘message passing’ steps, where it uses neural network layers to ‘pass messages’ between adjacent atoms and bonds in the molecule. This message passes aggregate information across the molecule and eventually results in a single, global representation of the molecule that incorporates all of the information from the individual atoms and bonds and their connections. This global representation, which is a single vector representing the molecule, is then used by another neural network that learns to predict the relevant property from this representation.”

Clever as Chemprop undoubtedly is, it is important to remember that it was developed using everyday open-source Python-based technologies.

“All of the code is written in Python, and we used PyTorch for the neural network components. We also occasionally used the NumPy, SciPy, and scikit-learn libraries throughout our code. We used the Python cheminformatics toolkit RDKit to initially process the molecules and convert them into the graphs that are processed by our graph convolutional network,” Swanson and Yang explained.

Training Chemprop to spot E. coli inhibitors

The team trained the Chemprop deep learning model to recognize E. coli inhibition as a property of a molecule, in the same way Chemprop had identified toxicity or solubility as a property in the original research.

They did this using a dataset of molecules known to inhibit or fail to inhibit the growth of E. coli, represented as a 1 or 0. They then tested it on a different dataset to evaluate how well it could predict E. coli inhibitors. Finally, they ran the model on a dataset where E. coli inhibition was unknown. The molecules for which Chemprop gave a high prediction of E. coli inhibition were tested in the lab by Stokes.

In total, Chemprop was tested on 2,335 molecules, of which 120 inhibited E. coli growth.

“Since ordering molecules for lab testing is a lot more expensive and time-consuming than running them through a computational model, we filtered out molecules that are similar to known antibiotics and clustered our predictions to ensure that we tested a diverse set of molecules and had the best chance of finding promising new antibiotics,” recalled Swanson and Yang.

Chemprop gets put through its paces

When the team was satisfied with the training, Chemprop was put to work searching through more than 107 million molecules—using datasets from the Broad Institute’s Drug Repurposing Hub and ZINC15—to look for ones with similar profiles to the inhibitors identified in testing.

The Drug Repurposing Hub is a database of more than 6,000 drugs that have been launched and approved by regulators or have undergone some clinical development. From these, Chemprop revealed a shortlist of 99 compounds with strong predictions, and empirical (laboratory) testing revealed that 51 on the shortlist did inhibit the growth of E. coli.

“We prioritized strongly predicted molecules for follow-up based on those that were structurally least similar to known clinical antibiotics and also, in certain cases, low predicted toxicity,” says Stokes, who was the lead experimenter on the project, as well as lead author of the research paper (published February 2020). “We also prioritized molecules that we were able to source or synthesize fairly easily.”

The compound that most satisfied these criteria was a drug called c-Jun N-terminal kinase inhibitor SU3327, which had previously been investigated as a treatment for diabetes. The team renamed it halicin, after HAL 9000, the AI computer in the 1968 film 2001: A Space Odyssey.

Lab tests conducted by Stokes showed that halicin would inhibit the growth of E. coli, which was the goal of the investigation. But further lab tests revealed that halicin would also inhibit the growth of a range of other serious pathogens, including Mycobacterium tuberculosis, which causes tuberculosis, as well as “hospital superbugs” including Clostridioides difficile (C. difficile).

“These are not similar types of bacteria, which is why we were quite surprised that halicin had activity across these different types,” says Stokes. “We are continuing to test for toxicity to human cells by halicin, as well as understand [its] additional drug-like properties. If these studies go well, we hope to partner with a biotech, pharmaceutical, or nonprofit organization to help move halicin into clinical trials.”

Other possible antibiotics

While halicin has great potential as a broad and powerful antibiotic, it is not a panacea. Stokes pointed out that trials did not show it to have significant impact on the bacteria pseudomonas, one type of which, P. aeruginosa, is another hospital superbug, which can occur after surgery and cause infections of the blood and lungs (pneumonia) as well as other complications. However, the team has other encouraging molecules to investigate.

As mentioned, in addition to the Drug Repurposing Hub dataset, the team also put Chemprop to work on the ZINC15 database, which contains hundreds of millions of purchasable “drug-like” compounds. Chemprop provided a shortlist of just over 1,000 molecules with a prediction score of >0.9 (1 is the score of a known inhibitor).

Filtering out those with high toxicity or similarity to common antibiotics whittled this down from 1,000 to 23. Lab tests revealed that eight of these would inhibit one or more of the five target pathogens, while further research showed that two molecules were particularly powerful.

The combined deep learning and life sciences teams at MIT are now starting an investigation into finding antivirals to combat COVID-19, and Chemprop is one of the tools in their arsenal.

The search for drugs that inhibit COVID-19

So the big question right now, of course, is could Chemprop be used to find drugs that inhibit COVID-19? “Yes, it could be,” says Stokes. “Provided that the appropriate training data is available, Chemprop could be applied in a similar way to how we used it for antibiotics to find antiviral molecules with activity against Sars-CoV-2 [the virus that causes COVID-19].”

MIT’s J-Clinic has recently launched an AI Cures team, with a website of resources available for collaborating researchers. These include datasets pertaining to both COVID-19 (Sars-CoV-2) and related viruses for which there is more data, such as Sars (Sars-CoV-1), which broke out in Asia in 2002. This will coordinate MIT’s efforts to use and develop AI technologies to find antiviral molecules that may be effective with COVID-19 infections.

“We aim to develop, apply, and make available AI tools designed to help identify or generate molecules or molecular cocktails against viral, bacterial, or other similar threats,” explains Tommi Jaakkola, professor at MIT CSAIL and joint lead of the AI Cures team. “Traditional approaches to drug development are expensive and too slow to react to pandemics like COVID-19.

“Regarding COVID-19 specifically, to contribute in the short term, our goal is to search for effective and safe antiviral therapies from among already-approved drugs or those that have already passed phase one clinical trials. In particular, we are looking for effective combinations or cocktails of such drugs.”

And it’s looking like Chemprop could help.“It’s a starting point for us, a baseline method,” confirms Jaakkola. “It is still useful for predictive screening of safe and effective single compounds, similar to its original use for discovering antibacterial drugs. We are also developing a number of more sophisticated tools.”

We wait with bated breath to see whether Chemprop and other deep learning technologies can help scientists deliver the same breakthroughs against the deadly virus as Chemprop did against deadly bacterial infections.

This article is part of Behind the Code, the media for developers, by developers. Discover more articles and videos by visiting Behind the Code!

Want to contribute? Get published!

Follow us on Twitter to stay tuned!

Illustration by Blok

Les thématiques abordées