Machine Learning Identifies Anti-Aging Drugs

It is not (yet?) the age of machines, but it is the age of machine learning.

That is, if we are to believe the science headlines of the past five years or so. Machine learning — I’ll use it here interchangeably with ‘narrow’ AI, which is not exactly correct, but makes for easier reading — has been making inroads into many fields.

Several recent machine learning advances are mentioned in some of my earlier posts.

(Warning: link fest! The previous posts I am referring to: general science and art, but also more specifically, the use of AI in historical research, genetic enhancement, mental health, aging research (including the development of ‘aging clocks’), video game ecology, Hollywood, astrobiology, epidemiology, stock markets, and the job market.)

In the field of molecular biology, AI is being used to, among others, comb through databases of molecules and infer possible not-yet-known functions.

Consider the deep learning model that was given a training dataset of over 2,000 molecules with a known antibacterial working. After training, it was set loose in a database of over 100 million other molecules without any further assumptions about their structure or function. The system was given the task to identify potential new antibiotics.

Several molecular hits later, the model pointed towards a molecule that did not look like any other known antibiotic. Yet, the model stubbornly claimed it would be good. The molecule, halicin (named after HAL) has already been tested in mice and proved effective against several pathogens.

Combine this with one of the uses of AI in aging research. In the post where I discuss those, I wrote:

By combing through databases of existing drugs, AI might identify some that intervene in the decline associated with aging.

And guess what, the time has (partially) come.

A new study now uses a convolutional neural network (a type of machine learning system) to:

… identify senescent cells and a quantitative scoring system to evaluate the state of endothelial cells by senescence probability output from pre-trained CNN optimized for the classification of cellular senescence…

Let’s unpack that a little.

The researchers started from human umbilical vein endothelial cells (cells on the surface of the umbilical vein). With these cells in hand — or rather in a petri-dish — they induced aging. They did this in three different ways: using reactive oxygen species (aka oxidative damage), a chemotherapy drug, and through forced replication, all three of these things are known to induce senescence.

With over 300,000 detailed photographs of the cells receiving one of these treatments, the researchers had plenty of training material for their neural network.

In time, the network got really good at correctly classifying practice images. (The F1 score, for example was 0.93. This is not exactly the same as ‘percent correct’, but what matters is that this is really good.)

Step one: complete.

Then, they tested the network on images it had never encountered before. Similar — marginally lower — score.

Step two: complete.

Up until this part of the story, the neural network’s output is either 1 (senescent) or 0 (non-senescent).

Ideally, we want a bit more.

We’d like to say this cell/tissue is getting more or less senescent after intervention x.

To get there, the researchers performed a more fine-grained analysis of the images, to see whether the ‘amount’ of visible senescence correlated with the amount of senescence-inducing stuff (the chemo drug or oxidative damage).

It did. The result was that the scientists could develop the Deep Learning-Based Senescence Scoring System by Morphology (Deep-SeSMo). That’s a very long name to say: a score that can tell us if a cell/tissue is becoming more/less senescent.

Step three: complete.

The scientists then tested Deep-SeSMo on two known senescence inhibitors: metformin and NMN. (But be careful with thinking they’re magic pills.)

Success. The Deep-SeSMo score reflected the effects of the drugs.

Step four: complete.

Time to let the neural network roam across a sea of molecular data. The researchers let their system explore a kinase inhibitor library (a big dataset full of molecules, basically).

It got hits. Here’s the top four:

• Terreic acid, a metabolite of Aspergillus terreus, a species of fungus. (Known to extend yeast life span).
• PD-98059, a selective inhibitor of mitogen-activated protein kinase.
• Daidzein, an isoflavone in soybean.
• Y-27632·2HCl, an inhibitor of the Rho-associated coiled-coil-forming kinase (ROCK), a member of the serine/threonine kinases. Or: a molecule known to regulate cell proliferation, apoptosis, migration, metabolism, and senescence.

Step four: complete

Now, before you run out to try and find this stuff, words of caution:

One, we don’t know the exact mechanism through which these substances might work, meaning we don’t know if they’re safe or if the side-effects are worse than the good effects.

Two, cellular senescence is only one aspect of aging. Aging is a very complex process that affects all your body’s tissues and many molecular pathways.

Three, no machine learning system is perfect:

Although the CNN showed high performance, there were still mispredictions.

In the overall performance, these mispredictions probably didn’t affect the overall senescense score too much. But beware of the word probably.

Overall:

A CNN-based approach contributes to the establishment of a non-biased method to identify morphological differences in research and drug screening. We developed a quantitative scoring system that evaluates cellular status by pre-trained CNN. Deep-SeSMo may be applicable for drug screening in other diseases and as a landmark system for drug discovery.

Next steps?

• Keep training the system.
• Let it explore more molecular libraries.
• Start testing the most promising candidates.

We’ll be waiting.