Enlarge /. Here are some antibiotics.
Biochemists have had some success in developing drugs that are geared towards specific goals. But much of drug development remains a difficult task, where hundreds to thousands of chemicals are screened for a "hit" that has the desired effect. There have been several attempts to do this silico grinding using computers to analyze chemicals, but they have had mixed results. A Canadian team is now reporting that it has modified a neural network to deal with chemistry and used it to identify a potential new antibiotic.
Artificial neurons meet chemicals
Two factors have a major impact on the success of neural networks: the structure of the network itself and the training it goes through. In this case, the training was pretty minimal. The research team conducted the training on a group of 1,760 drugs previously approved by the United States FDA, along with another 800 or so natural products. Most of them are not antibiotics; They target a variety of conditions and consist of largely unrelated molecules. The researchers only tested whether these slowed the growth of E. coli. Although many of them were partially effective, the researchers set a limit and used it to give a yes or no answer.
This approach has some advantages because it should not affect the resulting neural network for a particular chemical structure. With such a small data set, however, it is likely that some specific functional chemical groups have been completely left out of the training set. Success was also very rare, as only 120 molecules arrived above the cutoff. And since the cutoff was a binary "working" or "not working", the network was unable to identify trends that could help it project which chemicals might be more active.
If this part of the experiment seems somewhat exaggerated, it stands in sharp contrast to the work that was put into structuring the neural network. Normally, the individual functional units of a neural network carry out a series of simple tasks: input from other "neurons", performing your own calculations and transmitting the results to the next neurons across the board. In this case, the neurons were set up to correspond to a representation of a molecule, and any messages representing their chemistry sent to all neurons to which they were chemically bonded.
With sufficient messaging, the network's final output messages were a representation of the entire molecule, and the messages are combined to provide a vector representation of the chemistry of the molecule. This representation 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 with what it had learned from its training.
To ensure that it worked, the authors compared their scores to those created by a variety of other algorithms, including other neural networks trained with the same training data. All promising-looking chemicals have also been evaluated using an algorithm that predicts their likely toxicity in humans.
But does it work?
Apparently! After the network was used in a small library of chemicals, 99 molecules that looked promising were identified. Tests of these showed that more than half inhibited the growth of bacteria. And perhaps more importantly, there was a good correlation between the score for the molecule generated by the neural network and its performance in testing against actual bacteria.
After a few more tests, the researchers tackled a major one: a selection from a huge database of over 100 million molecules (107,349,233 to be exact). It took four days to go through these systems, which is much faster than "probably never" that would be required to screen this number of molecules in real life. Not surprisingly, a number of molecules came out of this screen, and the authors describe some tests from two of them. Both had a wide range and killed a wide variety of bacteria – one of them brought growing bacterial cultures to a standstill in just four hours.
However, most attention has been devoted to a molecule called halicin (according to a 2001 press release in honor of the AI, HAL 9000). Halicin was originally developed to target a human protein in the hope that it would help treat diabetes. With this in mind, we shouldn't be surprised that halicin doesn't look like well-known antibiotics. (This was true for most of the molecules identified on the various screens.)
Halicin has been active against a variety of bacteria (though not all) and is effective against known drug-resistant strains. The researchers also created wound infections that they successfully treated with halicin. It also eliminated C. diff infections, a common cause of drug-resistant digestive problems. What matters is that halicin also kills cells that don't divide – resting is one way many bacteria can survive antibiotic treatments.
The researchers decided to find out how halicin works by developing a resistant strain. Amazingly, they didn't make it, which is obviously positive. Instead, they examined the genes that were active in bacteria that were exposed to halicin. These provided an indication of how halicin works: by disrupting the balance of protons in the cell. Bacteria usually use their energy to pump protons out of the cell and use their return to drive ATP production and move the flagella that drives them through water. If halicin is present, the protons will return to the cell without doing anything useful.
Intelligent chemistry
This approach is obviously extremely promising. We quickly run out of antibiotics and the methods we used to create new candidates have not really brought anything new lately. Not only is this a different approach, it also doesn't contain any of the distortions that would normally affect the human-made discovery. In addition, the same general approach could be followed for a variety of diseases, particularly cancer. And things should only improve as researchers who manually examine drug panels regularly publish new data that could be used to further train or redirect the system to new disorders.
However, it is important to emphasize that this system, even if it continues to work, is only a partial solution. Not all molecules in these databases are free of toxicity or out-of-target effects, and some simply do not work. Then the question arises whether they can be made using standard reaction techniques and in a way that works with both industrial practices and health and safety standards.
To some extent, however, it is surprising that this limited data set can provide such useful results. Hopefully the authors have used a neural network that is verifiable so that we can get an idea of what chemistry it is looking at.
Cell, 2020. DOI: 10.1016 / j.cell.2020.01.021 (About DOIs).