How I got into AI: Small nasty bugs

"Hi Andreas, my name is Marco and I work in chemical biology. Would you have time to explain to me a bit how machine learning could help me? I got a ton of images of bacteria from a recent screen and frankly I don’t know what to do with that…”

My journey in artificial intelligence started with this call to Andreas Maunz, a bright scientist working at Roche. It was a cold February day back in 2015.

A couple of months earlier, Kurt Amrein and I were seated side by side at our monthly department review committee meeting. Roche was big in antibiotics research at that time and several high throughput screening campaigns were launched in search of novel antibiotics.

The results were quite puzzling. Out of the 1.5 million compounds of the Roche compound library tested on 4 Gram-negative pathogens, only a handful of them were active, all already known and available at the pharmacy round the corner.

Compounds were evaluated on their ability to reduce bacterial growth (in lay terms: killing the bugs). I dropped him a note on a post-it: “R W makin’ mistakes here?”.

After the meeting was over, we had a crazy idea: what if we screen for a different readout? I don’t know, maybe cell shape change…

The reasoning behind this was that maybe we were simply not giving a “fair” chance to compounds, many of them in fact capable to modulate/perturb bacteria biology but not potent enough to kill the bugs.

It was definitely a crazy idea worth exploring. More importantly, there was a way forward to test the hypothesis. By reducing the dose of the compounds killing the bugs, we would have expected to see changes occurring to the bacteria when imaged under a microscope.

We needed someone capable to do that.

We put a meeting in Sannah Zoffmann diary, at that time managing the cell imaging facility. Sannah was (and is) a master of imaging cells under the microscope. It took us some hours to convince her but she enthusiastically adopted the idea.

Her team started to test the hypothesis. When diluted, at non killing concentrations, the “killer” antibiotics were in fact inducing specific changes to bacteria shape. The idea was not as crazy as we thought.?We could now apply to the entire Roche compound library.

There was a problem...we underestimated the amount of data such an approach would have generated when applied to a 1.5 million compound collection. Petabytes of data were sitting in our drives and nobody had an idea how to analyze them.

At this point of the story I was lost. At lunch I told the story to a colleague that told me: “Wow, you got to speak to Andreas, he will love it, it seems perfect for his machine learning algorithms”.?I called Andreas.

Andreas was gracious enough to listen to me, comprehended my problem and explained to me that this was a classical case where application of machine learning (a subspecialty of artificial intelligence) would have been productive.

It took us several years of work to making it working properly. A team work performed in collaboration with novel key collaborators, Fethallah Benmansour, a philosopher of imaging and the microbiologist Maarten Vercruysse (part of Kurt Amrein 's team).

The tool really delivered what we were hoping: identify novel hit compounds likely to become antibiotics. But this were still not antibiotics (aka: not killing bacteria).

To transform these hits into true antibiotics we thought we had to do the exact opposite of what we did with the true antibiotics when we diluted to come up with our AI-assisted approach. We had to make them more potent and show they could kill the nasty bugs.

The only way to do that was to improve their potency. Chemists started to modify these compounds.

Or, if you work in Pharma or Biotech, you know the reluctancy chemists have to change their pretty tiny molecules without having a clear measurement of the molecular changes occurring in their reactions. This is specifically true when you screen compounds without knowing what the molecular target is (this is called phenotypic drug discovery). This was exactly our case...

While seeking potency, in fact, a single molecular change can transform a molecule acting in a certain way into another one acting completely differently. All that irrespective of potency.

While this is not a problem when you have a precise measurement of activity of your molecule on your target protein this becomes a HUGE problem when you screen for very general characteristic (we call that a phenotype) such as killing the nasty bugs or cell shape change.

And here comes the real plus value of the use of AI.

Fethallah and Andreas quickly realized they could use the machine learning algorithms to come up with a kind of identikit of each compound, inspired by a whisky they called that an archetype.

In the meantime, Tobias Heckel , one of the brightest guy working in sequencing, joined the team. Tobias work (sequencing) was added to capture even more biology to the compound archetype.

Almost magically, AI enabled chemists to seek potency while at the same time make sure the newly synthetized compounds would still belong to the same archetype. You seen an example below.

You can read the full story of our work in Cell Chem Biol.

Next article will be on Artificial intelligence and clinical development.

Stay tuned.

#Artifical Intelligence #AI #DrugDiscovery #HCS #highcontent #screening