Neural network created at SRNU MEPhI, thanks to a combination of AI and quantum chemical modeling, reveals huge opportunities in pharmaceuticals.
“The pandemic has clearly shown how important it is to have self-sufficient pharmaceutical industry. Now it is obvious to everyone that this is a strategic industry that ensures the security of the country. We coped with the release of vaccines and drugs, but partially these successes are based on imported components. Therefore, we should begin immediately the formation of a technological independence on biological and chemical substances, raw materials, and equipment”, Russian Prime Minister Mikhail Mishustin said, speaking at the Innoprom International Industrial Exhibition.
A neural network created by scientists of SRNU MEPhI can become one of these steps to the formation of technological independence.
Fullerenes are spherical hollow clusters with a diameter of slightly less than one nanometer. The most famous fullerene C60 consists of 60 carbon atoms and is similar to a soccer ball by its structure. It has been known since 1984, and in 1996 its discovery was recognized by the Nobel Prize in Chemistry. Researches of the last three years have proved that this fullerene is interesting not only to chemists, but although to doctors. It turned out that it can adsorb many anti-cancer drugs and increase their activity several times. Fulleren not only forces drugs to act stronger, but also provides them with the best penetration into cancer cells.
At the same time, it is completely non-poisonous and easy to excrete from the body. Considering the possibilities of inexpensive and good reproducible synthesis of this cluster compared to other nanoparticles, it has become a common “carrier” used for drug delivery. Its scope is expanding: now it delivers medicines not only for cancer, but also for diabetes, asthma, as well as heart vascular and respiratory diseases, including COVID. Drug delivery reduces the required dose several times, thereby reducing the cost of treatment and minimizing side effects.
Fluorinated fullerenes deserve special attention. They are easier trackable inside the body using nuclear magnetic resonance, and they can release drugs in a controlled manner due to high activity in near-infrared. However, despite all the advantages of fluorinated fullerenes, they are not absolutely universal carriers: some drugs can be efficiently delivered to them, while others can not.
“To investigate the interaction of each drug with fullerenes means huge labor costs, no resources will be enough for this, – Konstantin Katin, Dr. Phys.-Mat. sciences, professor at SRNU MEPhI, explained. – Therefore, we have applied to the aid of quantum chemical modeling and neural networks. It allowed us to consider not one or two drugs, as is usually done in such studies, but forty popular drugs at once. This quantity turned out to be enough to carry out statistical analysis: we identified “leaders” and “outsiders” among drugs (of course, we are talking only about their ability to be delivered on fullerenes).
To draw conclusions about the rest of the drugs, including those that are still to be opened in the coming years, we tried to establish patterns describing their interactions with fullerenes. To do this we used, among other things, neural networks, which can sometimes find non-obvious patterns that a person cannot notice.
The network we trained can predict how reliably a drug can be “anchored” to fullerene. Its accuracy is limited, but it works very quickly, which makes it useful at least for primary drug analysis. To evaluate the interaction of fullerene with medicine, it takes a few moments, while classic quantum chemical calculations of even one drug can last weeks and even months.”
The study carried out illustrates the huge possibilities that appear in pharmaceuticals thanks to a combination of AI and quantum chemical modeling. Number of possible molecules is incredibly large, so just a negligible part of them was synthesized and investigated to date. That is why every year we read reports about discovery of a variety of new drugs.
Modern chemists can synthesize almost any molecule. The problem is that scientists still do not know how to predict reliably whether this or that “drug-like” molecule will be truly effective medicine. Even quantum chemical simulations can’t cope with a huge queue of applicant molecules selecting among them a small number of really useful compounds. Here the help of neural networks cannot be overestimated: they are already “wiser” so much that they can summarize the results of quantum chemical calculations and separate the “false” medicines for “real” ones.
Another problem is that most quantum-chemical programs are created in Western countries, and access to them may be limited.
“We can create our own, domestic, quantum chemical programs for the pharmaceutical industry”, Konstantin Katin says. – Many existing foreign programs have a long history, so you have to adapt them to modern computers. Some parts of the old code are no longer advisable to redo. Therefore, analogues created “from scratch” may be more effective than projects developing a long time ago.
In addition to fullerenes, other carriers are often used in pharmaceuticals: metal and composite nanoparticles, polymers, boron-nitride clusters, etc. All of them have their own advantages and disadvantages, important in treatment of specific diseases. For each newly created drug, you have to select a suitable medium, which can take several months and slows down the trial and implementation process. Considering that the cost of developing one drug can reach up to one billion dollars, such delay can lead to losses calculated in tens of millions of dollars.
The program we created, trained on a variety of already known drug delivery systems, will be able to select instantly the appropriate carrier for this drug, and also a drug suitable for the carrier. It will become possible thanks to the combination of traditional quantum chemical approach with training and use of neural networks.”
Links to articles in scientific journals: