A peek inside the companies using AI to combat aging

Last Wednesday I attended the Buck Institute’s workshop on AI and Longevity, where speakers from several different organizations discussed how they were using machine learning to advance different aspects of medical discovery. Uses for AI in aging include creating new drugs, finding new uses for existing drugs, and discovering new biomarkers–all of which are important for improving the current drug development cycle, which is long and expensive.

The Buck Institute itself was established in 1999 and remains the only independent anti-aging research center in the world. Its goal is to fight the diseases of aging by approaching them holistically, as opposed to traditional organ-focused research. 

A primer on deep learning

All talks used a machine learning technique called neural networks. In broad strokes, neural networks can do three different things:

• Classification – Given a picture of a cat, label it “cat”
• Regression – Given a house’s location and square footage, predict its sale price
• Generation – Given the word “cat”, produce a picture of a cat

All neural networks are “trained” on sample data before being given problems to solve. For example, the classification network would be fed thousands of pictures labeled either “cat” or “not cat”. The regression network would be fed hundreds of homes with their statistics and sale prices (and the amount of training data necessary is shrinking all the time). The trained network is then fed the same type of data (a picture of a cat, housing statistics) and attempts to give the answer (the label, the sale price).¹

The long march to market

The driving motivation behind every talk was the difficulty of discovering new medications. It takes 5,000-10,000 starting compounds and 10 to 15 years to bring a single medication to market. The time cost is even higher for anti-aging treatments, which require a very long trial period to see results. This is especially bad because the clock on a drug patent starts at the beginning of the pipeline, not when the drug is released to market. If a drug takes more than 20 years to prove its effects–which an anti-aging treatment very well might–there is no way to recoup investment costs.

Many companies are using AI to attack the first stage of the drug pipeline: identifying promising compounds. This can be done in several different ways.

• Identifying new uses for existing drugs (classification). Atomwise wanted to determine if any existing, FDA-approved medications could be useful for fighting Ebola. Researchers identified a likely target for intervention–a particular point on a viral coat protein that needs to receive another protein in order to operate. They hoped that by blocking the receptor they could prevent the virus from entering cells. After training their model by asking how well a set of compounds fit the target receptor, they fed the algorithm new compounds and asked how well these fit. Out of 7000 compounds, they identified 17 promising compounds.
• Generating entirely new molecules for a specific purpose (generation). Numerate wanted to target Alzheimer’s disease using the ApoE4² corridor. Starting with a database of 10 million compounds, they used a process called scaffold hopping to generate molecular formulas for compounds that would fit the same receptor as ApoE4. This process took ~9 months and $1,000,000 to identify 10 patentable classes, 4 of which went on to pass in vivo studies.
• Identifying potential side effects without testing (classification). Numerate trained an AI using compounds and their toxicity as training data. The trained model was then fed novel molecules and predicted potential negative side effects. This allows pharmaceutical companies to avoid spending billions of dollars testing compounds that will never be usable, even if they are effective.
• Finding the holy grail for clinical trials (regression). BioAge Labs took a slightly different approach: instead of using deep learning to look for new drugs, they instead looked for biomarkers, easy-to-collect measurements that are highly predictive of aging outcomes. Biomarkers should allow you to evaluate the use of your intervention without waiting for the event you are trying to cause or avoid. For example, looking for treatments that reduce blood pressure is easier than looking for treatments that reduce heart attacks, because you’re able to get statistically significant results from a smaller sample size over a shorter period of time. This means that useless compounds are discarded sooner, and promising compounds are brought to the public faster.

While all of the companies that spoke at the workshop target aging, there was nothing aging-specific about their use of neural networks; these are very broad, very new tools that are being used opportunistically. However, everyone present at the Buck’s seminar on Wednesday was very hopeful that these techniques will soon lead to new discoveries in anti-aging medicine.

1. For an explanation of how neural networks do these things, see this video from 3blue1brown.
2. ApoE is a class of proteins involved in cholesterol transport. Individuals with two copies of the ApoE4 variant have 10-30 times the chance of developing Alzheimer’s disease before age 75. This is not universal; certain ethnic groups, such as Nigerians, have very low rates of Alzheimer’s despite carrying the E4 variant.This may be because their cholesterol is so low.