Shulou(Shulou.com)06/02 Report--

No matter how many words of praise that antibiotics bring to human beings, they cannot be overstated. At the initial stage of application, antibiotics were regarded as a panacea for almost all diseases, and became one of the greatest achievements of mankind in the 20th century.

(photo title: thank you for penicillin, give me the way home)

Today, however, about 50% of antibiotics are abused worldwide each year, and the proportion is even close to 80% in China. People with children may have more personal experience. As long as the children have a headache and fever, they always go to the hospital for an intravenous drip, and most of them use antibiotics. The direct consequence of people's obsession with "quick results" is to make bacteria more drug-resistant, and antibiotics that worked in the past may fail as a result of increased drug resistance.

Ordinary people who have high hopes for antibiotics may think: scientists just need to find stronger antibiotics than drug-resistant bacteria. In fact, scientists do, but the study of antibiotics is far from as simple as people think.

For onlookers who are more anxious and obsessed with "silver bullet thinking", they always want to break the casserole and ask: is it possible to find an antibiotic that can effectively kill a large number of harmful bacteria? In terms of biological laws, this is a bit wishful thinking. But there must always be a goal. What if we get close?

Ask for what you can get. On February 20th, Cell, a famous natural science journal in the world, published a research paper entitled "A deep learning method for discovering antibiotics". It was reported that the research team of Massachusetts Institute of Technology (MIT) discovered a super antibiotic-Halicin using the deep learning model. Through experimental study, this new antibiotic compound has bactericidal effect on many kinds of drug-resistant bacteria and has broad-spectrum antibacterial effect. Of course, the highlight of this study lies in the new research method, that is, through AI's deep learning algorithm to find this unique molecular structure from a large library of synthetic chemistry molecules.

AI's ability to develop drugs is not uncommon, but this is the first time antibiotics have been discovered. Curiosity will certainly drive us to ask: why is the rate of human development of antibiotics not outpacing the growth of bacterial resistance? what does it mean for medicine that AI pioneered the path of antibiotic research and development?

The puzzle discovered by antibiotics, why is the human road getting narrower and narrower?

In the decades since Fleming discovered penicillin, the first antibiotic, in 1928, scientists have discovered more than 100 antibiotics. But for 28 years from 1987 to 2015, no new antibiotics were found.

Why is the discovery of new antibiotics getting slower and slower, even for such a long gap period? It can be said that the main antibiotics we use at present were found in the 1940s and 1950s and were screened and cultured from microbial populations in the soil. About 99% of bacterial populations could not previously be cultured separately in the laboratory, a restriction that makes it impossible to isolate potentially effective antibiotics produced by bacteria.

Since the 1960s, the search for new antibiotics from microbial populations has slowed significantly and has been replaced by "semisynthetic antibiotics". Nowadays, β-lactam antibiotics, mainly penicillin and cephalosporin, have become the most important chemotherapeutic agents.

It was not until 2015 that American scientists discovered a new antibiotic, Teixobactin. This antibiotic has achieved a major breakthrough in the field of antimicrobials, killing a variety of deadly pathogens, while pathogens are difficult to develop resistance to them. This scientific progress benefits from the progress in the technology of bacterial cultivation. Through an electronic chip device called iChip, the researchers were able to isolate and grow the target bacteria in a more natural soil environment than the laboratory, raising the number of bacteria with only a 1 per cent chance of growth to 50 per cent.

In spite of this, the efficiency of finding new antibiotics is still too low, and the antibiotics cultured on a large scale are likely to be old antibiotics or do not have good antibacterial activity. So why can AI do it this time?

Peep and see Leopard: a New path of AI for New antibiotics

Because of the inefficiency of traditional antibiotic drug cultivation, it is a new idea to screen the molecular structure of antibiotics from a large synthetic chemical library. However, these libraries may contain hundreds of thousands to millions of chemical molecules, but the chemical diversity of molecular formulas is limited and cannot reflect the possible chemical properties of antibiotic molecules. On the one hand, the workload of manual screening is too high, and the second is the limitation that it is difficult to verify the diversity of compounds.

Screen the molecules that can inhibit certain bacteria from the molecular library, and then verify the difference between these molecules and the antibiotics that have been used. It involves massive calculations and repeated experiments, which makes this approach really unfriendly to human researchers.

But this time, AI introduced screening of these synthetic chemical molecules as a new research tool, leading to a major turnaround in the discovery of antibiotics. The researchers proposed the idea of combining structural analysis with screening, and used machine learning algorithms to predict compounds with potential antibacterial properties from molecular properties. It was a great success this time.

First of all, the researchers screened 2335 drug molecules or natural compound molecules approved by FDA with the goal of inhibiting certain E. coli, and then divided them into bacteriostatic and non-bacteriostatic categories according to the 80% growth inhibition rate as the training data set of the neural network.

At the same time, they use a "directed message passing depth neural network" (DMP-DNN) algorithm, which repeatedly transmits the atomic and bonding information of compounds in the form of continuous vectors, so as to obtain more advanced characterization results. This neural network characterizes molecules through self-learning vectors without the need for artificial labeling of specific molecular structures.

The directed information transmission network can predict the properties of molecules directly from the graph structure of molecules. After a fixed number of information transmission steps, the resulting single vector of molecules can predict the inhibition rate of bacteria. This way of seeing leopards greatly shortens the screening path and reduces the cost of computing power, time, and so on.

(except Pseudomonas aeruginosa (bottom blue), Halicin showed good broad-spectrum antibacterial activity in several drug-resistant bacteria tests.

Subsequently, the trained deep learning model was applied to the Baird Institute's library of DRH compounds with a molecular weight of about 100 million. The model selected 99 compounds most likely to have antimicrobial activity from 6111 drug molecules in the research stage.

Finally, the experimental results show that 51 of them can significantly inhibit the growth of this kind of Escherichia coli. This compound, named Halicin, was selected as the most potential new antibiotic because of its low toxicity and structural novelty.

Since then, the researchers applied the retrained deep learning model to the larger ZINC15 drug small molecule database, predicted and screened about 100 million compounds, and discovered new potential antibiotic structures. This screening process took only three days.

It can be seen that with the help of in-depth learning, 6111 drug molecules have been screened into 99, and the efficiency of these 99 drug molecules will naturally be much faster.

This groundbreaking study marks a paradigm shift in the discovery of antibiotics and even the more common methods of mining new drugs. So what more value will AI bring after all?

The battlefield of new antibiotic drug research and development, AI is attacking by the whole army

Generally speaking, the research and development of traditional drugs mainly goes through these four stages:

1. Selection and confirmation of targets. 2. Discovery and optimization of lead compounds. 3. Preclinical studies. 4. Clinical trials. Before it can be officially approved for listing. A new drug faces three "10" tests from research and development to market: 10 years, $1 billion, and a 10% success rate, and this trend is even more serious. For the research and development of antibiotics, it is also faced with the same problem of high input, low output and low profit for a long time.

With the help of AI, especially the deep learning algorithm, drug research and development is ushering in a new turning point. From the new drug development that AI has been put into use, and this time AI has shown a very obvious effect in the discovery of new antibiotic compound molecules. It can be said that in the future, AI will play a great auxiliary role in new drug mining, compound screening, target discovery and drug effectiveness prediction, to some extent, improve the efficiency of research and development, save money and reduce the risk of failure in clinical trials.

The most important thing in modern drug research and development is the search and identification of drug targets. Drug targets refer to the binding sites of drugs in the body, including gene sites, receptors, enzymes, ion channels, nucleic acids and other biological macromolecules. The selection of novel and effective drug targets is the primary task of new drug development. Before the introduction of AI technology, the screening of new targets of traditional antibiotics was based on genome, inhibition of protein synthesis, synthase and so on. The difficulty lies in the low efficiency and slow progress of manual screening test. Through the intervention of AI's deep learning model, we can more quickly count tens of millions of unstructured data in scientific literature and find possible suitable targets in the genetic molecules of these organisms. Then through the selection of different targets and verification, in order to understand the biological characteristics of the targets, real-time interaction to get the results of evidence, to achieve the discovery of the corresponding drug targets.

In addition, AI is equally efficient and accurate in screening compounds for new drugs. The discovery of Halicin, a highly active compound, is based on the specific drug targeting requirements, using flexible model algorithms to effectively screen possible drug molecular structures among thousands of possible molecular compounds, which greatly saves screening time and cost.

In addition, in addition to the discovery of new drug targets and highly active compounds, AI has practical applications in the design of automatic synthesis routes for small molecular drugs, the simulation and prediction of new drug effects, and even the prediction of new drug molecules. This will bring great opportunities for both antibiotic drugs and the more general drug research and development.

There is an image saying in the field of new drug research and development that those "drooping fruits" have been picked, and future new drug research and development needs to go to fruit trees blocked by thick branches and leaves to pick fruit. And AI may be the best way to reach these branches to pick fruit.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.