Artificial intelligence may change the rules of the drug development game

Artificial intelligence algorithms can be used to specifically design active ingredients that have the same effects as natural substances but have a simpler structure. In the rapid design, manufacturing, testing, and analysis cycle, automated, rule-based molecular construction is well combined with machine learning and experimental verification.

Recently, scientists from the Swiss Federal Institute of Technology Zurich (ETH) published an article in the journal Advanced Science, introducing how to use artificial intelligence (AI) to develop new drugs based on natural examples. Artificial intelligence can not only identify the biological activity of natural substances, but also help to find molecules that have the same effects as natural substances, but are easier to manufacture. This method can make it easier to design new, patent-free molecular structures in the future, and may change the rules of the game for pharmaceutical research and development.

Image source: Visual China

Natural substances are an important source of innovative drugs

The use of natural substances for drug design is an effective way to develop modern innovative drugs. According to statistics, between 1939 and 2016, more than 50% of the listed drugs approved by the U.S. Food and Drug Administration (FDA) contained molecular fragments of natural substances, or were directly derived from natural substances. Compared with chemically synthesized small molecule drugs, natural substances have obvious advantages in terms of structural novelty, biocompatibility, and functional diversity, and have undergone natural selection and optimization in the long-term evolution process.

The target molecules of natural substances are potential drug targets. Determining the target protein and mechanism of action of active natural substances is the key to the development of new drugs. However, it is not easy to find drug targets from as many as 400,000 different human proteins. Therefore, Professor Gisbert Schneider of ETH Zurich uses artificial intelligence programs to help find possible target molecules of natural substances, so as to identify related compounds in pharmacology. Schneider emphasized: "In this way, the opportunity to find a combination of medically important active ingredients and target proteins is much greater than traditional screening."

Artificial intelligence algorithm narrows the range of protein targets

The researchers chose Marinopyrrole A, a bispyrrole compound extracted from marine Streptomyces, to verify their artificial intelligence algorithm. Marinopyrrole A not only has anti-bacterial properties, but also has strong anti-cancer activity. Through the machine learning model, the researchers compared the pharmacologically significant parts of Marinopyrrole A with the corresponding active ingredient patterns and analyzed which target proteins they might be attached to.

Based on pattern matching, the researchers identified eight human receptors and enzymes that bacterial molecules can attach to. They are related to inflammation, pain, and the immune system. Experiments have confirmed that Marinopyrrole A does have measurable interactions with most predicted proteins. Schneider pointed out: "Our artificial intelligence method can narrow the range of protein targets of natural substances, and the reliability is usually more than 50%, thereby simplifying the search for active pharmaceutical ingredients."

Looking for alternatives that have the same effect but are simpler

Due to the relatively complex structure of many natural substances, laboratory synthesis is difficult and expensive. Therefore, Professor Schneider’s research team further developed another artificial intelligence program to find alternatives to natural substances that have the same effect but are simpler and cheaper to manufacture. This artificial intelligence program is equivalent to a "virtual chemist", which can find molecules with different structures from the natural model but with equivalent chemical functions. According to the algorithm design, such molecules must also be able to be produced in up to 3 synthesis steps, so they are relatively easy and cheap.
 In order to determine the synthetic route, this program has a catalog containing more than 200 starting materials, 25,000 commercially available chemical building blocks, and 58 established reaction schemes. After each reaction step, the program selects these variants as starting materials for the next step.

Also taking Marinopyrrole A as an example, the program found 802 suitable molecules based on 334 different basic structures. The researchers made the best 4 in the laboratory, and these molecules actually show activity very similar to the natural model. They have considerable influence on 7 of the 8 target proteins determined by the algorithm.

The researchers then examined the most promising molecules in detail. X-ray structure analysis shows that the calculated compound attaches itself to the active site of the target protein in a similar way to known inhibitors of the enzyme. In other words, although the structure is different, the molecules discovered by the artificial intelligence program have the same mechanism of action as the target model.

Designing molecular structures will become easier

In fact, the integrated method proposed by Professor Schneider and his team combines automated, rule-based molecular construction with machine learning and experimental verification in a rapid design, manufacturing, testing, and analysis cycle. Professor Schneider said: "Our work has proved that artificial intelligence algorithms can be used to specifically design active ingredients with the same effect but with a simpler structure. On the one hand, this helps to develop new drugs; on the other hand, it also puts us in medical chemistry research. The beginning of a fundamental change may occur."

It is worth noting that with the help of ETH Zurich's artificial intelligence methods, people can find alternatives to existing drugs that are equally effective but based on different structures. This can make it easier to design new, patent-free molecular structures in the future.
But this has also led to more intense debates: On the one hand, to what extent can artificial intelligence systematically circumvent drug patent protection? On the other hand, can the molecules of "creative" artificial intelligence design be patented? With the further improvement of this method in the future, the pharmaceutical industry will have to adjust its research strategy to adapt to the new rules of the game. 

(German correspondent Li Shan)

Post a Comment

Previous Post Next Post