Thu. Aug 18th, 2022
Hands in blue surgical gloves hold brightly colored drugs, including antibiotics.
enlarge Here, take some antibiotics.

Biochemists have had some success designing drugs to meet specific goals. But much of drug development remains a tedious grind, screening hundreds to thousands of chemicals for a “hit” that has the effect you’re looking for. There have been several attempts to perform this grind in silico, using computers to analyze chemicals, but they had mixed results. Now, a US-Canadian team reports that it modified a neural network to deal with chemistry and used it to identify a potential new antibiotic.

Artificial neurons meet chemicals

Two factors greatly influence the success of neural networks: the structure of the network itself and the training it undergoes. In this case, the training was quite minimalistic. The research team did the training using a group of 1,760 drugs previously approved by the US FDA, along with another 800 or so natural products. Most of these are not antibiotics; they target different conditions and consist of largely unrelated molecules. The researchers simply tested whether it could control the growth of E coli† While many of them were partially effective, the researchers set a limit and used that to give a yes or no answer.

This approach has some advantages in that it should not bias the resulting neural network for any given chemical structure. But with such a small data set, it is likely that some specific functional chemical groups have been completely omitted from the training set. Success was also very rare, with only 120 molecules coming above the limit. And since the frontier was a “works” or “doesn’t work” binary, the network had no way of identifying trends that could help it predict which chemicals might be more active.

If that part of the experiments seems a little underexposed, it’s in stark contrast to the work that’s gone into structuring the neural network. Normally, the individual functional units of a neural network perform a series of simple tasks: input from other ‘neurons’, perform their own calculations, and pass the results on to the next neurons down the road. In this case, the neurons were set to match a representation of a molecule, and each relayed messages representing its chemistry to all the neurons to which it was linked via a chemical bond.

With sufficient message transfer, the final network output messages were a representation of the whole molecule, and the messages are combined to create a vector representation of the molecule’s chemistry. This view was supplemented by the output of a simpler algorithm that evaluated the chemistry of the molecule in question. The neural network then used these values ​​to compare the molecule to what it had learned from its training.

To make sure it worked, the authors compared the evaluations to those of several other algorithms, including other neural networks trained with the same training data. All the promising-looking chemicals were also evaluated using an algorithm that predicts their likely toxicity in humans.

But does it work?

Evidently! After using the network on a small library of chemicals, it identified 99 molecules that looked promising. Tests of this showed that more than half inhibited the growth of bacteria. And, perhaps more importantly, there was a nice correlation between the score for the molecule generated by the neural network and its performance when tested against real bacteria.

After a few more tests, the researchers tackled a big one: a selection from a massive database of more than 100 million molecules (107,349,233 to be exact). It took the system four days to go through it, which is a lot faster than the “probably never” needed to screen that number of molecules in real life. Not surprisingly, a number of molecules came out of that screen, and the authors describe a few tests of two of them. Both had a broad spectrum and killed a wide variety of bacterial species – one brought the growing bacterial cultures to a halt in just four hours.

But most of the attention went to a molecule they call halicin (according to a press release, in honor of 2001‘s AI, HAL 9000). Halicin was originally developed to target a human protein in the hopes that it would help treat diabetes. Given that background, it should come as no surprise to us that halicin is nothing like known antibiotics. (This was true for most of the molecules identified in the various screens.)

Halicin was effective against a wide variety of bacterial species (though not all) and is effective against known drug-resistant strains. The researchers also created wound infections that they successfully treated with halicin. It’s also cleared up C. difference infections, a common cause of drug-resistant digestive tract problems. Crucially, halicin also killed cells that were not undergoing cell division — going silent is one way many bacteria survive antibiotic treatments.

The researchers decided to find out how halicin worked by developing a resistant strain. Amazingly, they failed to do so, which is of course a positive. Instead, they looked at the genes that were active in bacteria exposed to halicin. These gave a hint about how halicin works: by disrupting the balance of protons in the cell. Bacteria normally use their energy to pump protons out of the cell, using their return to stimulate ATP production and move the flagella that propels them through water. With halicin present, the protons find their way back into the cell without doing anything useful.

Intelligent Chemistry

This approach is obviously very promising. We are quickly running out of antibiotics and the methods we have used to produce new candidates have not yielded anything new lately. Not only is this a different approach, but it also contains none of the biases that would normally influence human-driven discovery. In addition, the same general approach could be taken for a wide variety of diseases, especially cancer. And things should only improve, as researchers who manually screen drug panels regularly publish new data that can be used for further training or redirecting the system to new conditions.

That said, it’s important to emphasize that even if this system continues to work, it’s only a partial solution. Not all molecules in these databases will be free of toxicity or off-target effects, and some just won’t work. Then there is the question of whether they can be produced using standard reaction techniques and in a manner that works with industrial practices as well as health and safety standards.

But to some extent, it’s astonishing that this limited dataset can produce such useful results. Hopefully the authors have used a neural network that is controllable so that we can get an idea of ​​what chemistry it is looking at.

Cell, 2020. DOI: 10.116/j.cell.200.01.021 (About DOIs).

By akfire1

Leave a Reply

Your email address will not be published.