AI has found a lung drug that makes people 3 years "younger" after taking it for 4 weeks.

A drug originally developed to treat pulmonary fibrosis made patients appear 3 years "younger" after 4 weeks of administration.

This drug has been intertwined with AI from the very start: AI first identifies the exact target for pulmonary fibrosis, then generative AI designs the corresponding drug molecule.

This story is far from distant from us.

Insilico Medicine, the company that developed the drug, has its main office in Hong Kong, while its core team responsible for new drug discovery and R&D has long been based in Zhangjiang, Pudong, Shanghai. Several years ago, AI selected a target called TNIK from massive disease datasets. Today, this judgment has been turned into a real drug that has entered the late stage of clinical trials.

In July this year, Insilico Medicine announced that rentosertib has officially entered Phase III clinical trials, with plans to recruit approximately 320 pulmonary fibrosis patients across 47 centers.

While this pulmonary drug continues to advance toward its marketing approval, researchers unexpectedly found a surprising signal from the blood samples left over from the previous Phase IIa trial.

42 patients, six different sets of "aging clocks" all pointed to the same direction:

Their blood appeared much younger.

In this way, a pulmonary drug discovered with the participation of AI has suddenly been linked to "anti-aging".

01

The target selected by AI several years ago has reached Phase III

Before a drug is officially researched, the first step is to find a biological entity worthy of intervention. It could be a certain protein, receptor, or a signaling pathway closely related to the disease progression.

This entity is the "target" of the drug.

The problem is that there are tens of thousands of proteins in the human body, and a large number of genes, cells and signaling pathways are involved behind diseases. In the past, searching for new targets relied on long-term basic research, accumulating evidence layer by layer, and then slowly judging which direction was worth investing in.

This is like drawing the target before shooting an arrow. If there is a problem with the judgment of the target, the subsequent drug R&D will definitely go off track.

AI has played a huge role in the task of finding targets.

It can simultaneously read the genome, transcriptome, proteome, disease databases, papers and biological networks, and find links between proteins and diseases that are worthy of further verification from information far beyond what humans can process one by one.

In recent years, target discovery with AI participation has become one of the important directions of AI-driven pharmaceutical research. This year, *Nature Reviews Drug Discovery*, the top review journal in the field of drug R&D, even published a special review discussing AI-participated target identification, stating that it is playing an increasingly important role in this link, but also emphasizing that whether a target is truly valid ultimately depends on experiments and even clinical verification.

In this field, BenevolentAI and AstraZeneca started very early to screen new targets from complex disease data using knowledge graphs and machine learning for directions such as chronic kidney disease and pulmonary fibrosis; Recursion has also been using large-scale phenotypic data and machine learning to find new disease mechanisms. For example, the key discovery behind its REC-4881 is that its platform identified that "inhibiting MEK1/2 may treat familial adenomatous polyposis", which has now advanced to Phase II.

It is not new for AI to help humans find targets, but the real difficulty is to push the research forward.

The target selected by the algorithm must first be verified by experiments; after the verification is confirmed, someone needs to design a truly usable molecule around it, and then go through toxicology tests, animal experiments, Phase I and Phase II trials, to prove that it is not just "theoretically feasible".

Many targets discovered by AI get stuck at a certain link in the middle.

Phase I mainly focuses on safety, and Phase II starts to observe efficacy signals and dosages; when it comes to Phase III, hundreds or even thousands of patients are usually recruited to compare with placebo or standard treatment under conditions closer to real clinical practice. At this stage, those answers that seemed "reasonable" with only dozens of patients often need to be re-verified. The seemingly obvious efficacy may be diluted by a larger sample size, and adverse reactions that were not exposed in small samples may emerge for the first time.

Precisely because of this, Phase III is usually one of the most critical, highest-cost and highest-failure-risk stages before a drug is launched on the market.

The most noteworthy part of Insilico Medicine's progress this time is that it has pushed rentosertib to Phase III clinical trials. The company claims that this is the first case of "AI-discovered new target + generative AI-designed new molecule" entering the critical clinical stage.

This route can be traced back to around 2019. At that time, Insilico Medicine was still using known targets such as DDR1 (discoidin domain receptor 1) to prove that generative AI "can design molecules". That year, they used AI to generate candidate molecules for DDR1 in 21 days and completed in vivo and in vitro verification, but that was more like a capability demonstration: the question was raised by humans, and AI was only responsible for solving it.

Insilico Medicine was not satisfied with finding molecules around known targets, but also wanted its AI target discovery platform PandaOmics to directly find new therapeutic directions from disease data. Idiopathic pulmonary fibrosis became one of the test scenarios, and AI finally brought TNIK (TRAF2 and NCK-interacting kinase, a protein kinase involved in cell signal transduction) to the forefront.

Then at the end of 2020, the candidate molecule designed around TNIK was officially confirmed as a preclinical candidate drug, which means it has been screened out from a large number of alternative molecules and is ready to enter systematic animal experiments and subsequent clinical development.

It took less than 18 months from starting target discovery to producing this candidate drug. Next, it was the turn of human trials to test whether this solution is reliable or not.

In February 2022, this drug, later named rentosertib, entered Phase I clinical trials; in July 2023, the Phase IIa trial was officially launched in China, with 71 idiopathic pulmonary fibrosis patients receiving 12 weeks of treatment.

In July this year, this drug officially launched large-scale Phase III clinical verification.

However, this is not the first drug that has participated in by AI and advanced to Phase III.

Less than a month ago, Moderna and Merck & Co. just announced a case that is even closer to the final endpoint.

On August 19, the two companies announced that the individualized mRNA cancer therapy intismeran autogene (V940 / mRNA-4157) combined with Keytruda achieved positive results in the Phase III trial INTerpath-001 for melanoma. The trial recruited 1,137 patients with stage IIB-IV melanoma who had been completely surgically resected but still had a high risk of recurrence. The results showed that the combination therapy was superior to Keytruda alone in two key indicators: recurrence-free survival and distant metastasis-free survival. Moderna stated that this is the first time that an individualized neoantigen therapy has obtained positive Phase III results.

The role AI plays here is also very specific: each cancer patient's tumor may carry a large number of different mutations. Moderna will sequence the patient's tumor and blood, then let a set of AI algorithms screen out up to 34 "neoantigens" that are most likely to activate the immune system from these mutations, and finally customize an mRNA therapy according to each person's unique mutation combination.

It is not the same AI pharmaceutical route as rentosertib. The former faces the problem of "among the numerous mutations in one patient, which ones should be selected to make a vaccine", while the latter is more similar to "among the numerous proteins and pathways behind one disease, which one is worthy of being developed as a new drug target".

However, when the two projects are viewed together, we can see that the answers given by AI participation are moving out of computers and early experiments, and entering Phase III clinical trials that truly determine the fate of drugs.

02

A pulmonary drug makes blood "appear younger"

Just as rentosertib was about to face the big test of Phase III, researchers unexpectedly found a surprising signal from the blood samples left over from the previous Phase IIa trial:

42 patients, six different sets of "aging clocks" all finally captured signals moving in the "younger" direction.

The most consistent result came from the dosage group of 30mg twice a day: at the 4th week, five of the six clocks pointed to "younger". And in the 60mg once a day group, the four clocks trained on actual age gave predicted age reductions of approximately 2.7 to 3.5 years.

In other words, the "3 years younger after 4 weeks of taking the drug" mentioned earlier is not entirely a metaphor — at least from the perspective of proteomic indicators, "reversing the aging process" may indeed happen.

The "younger" mentioned here mainly refers to the overall state of thousands of proteins in the blood.

The research team re-analyzed the blood samples left from the Phase IIa clinical trial. Finally, 42 patients were included in this proteomic analysis, who left blood samples before taking the drug, as well as at the 2nd week, 4th week and 12th week.

Researchers measured thousands of proteins in the samples at one time, and then put these data into six independently developed "proteomic aging clocks" respectively.

The so-called aging clock can be understood as an age prediction model. As people get older, many proteins in the blood will show relatively stable changing patterns.

It is a bit like "guessing age by looking at the face" — for people of different ages, wrinkles, skin condition and facial structure usually show some regular changes. After the model has seen enough data from people of different ages, it can estimate how old a person appears to be based on these features.

The proteomic aging clock does a similar thing, except that it does not look at the face, but at the combination of hundreds and thousands of proteins in the blood.

Therefore, the real question the researchers asked is: after taking rentosertib, did the blood of these patients become more similar to that of younger people or older people at the protein level?

The answer, at least at the 4th week, very consistently points to the former.

Speaking of which, it has to be mentioned that this question was not actually thought up temporarily by the researchers afterwards.

On the one hand, idiopathic pulmonary fibrosis itself is a highly age-related disease, and the reason why TNIK was targeted by Insilico Medicine in the first place was not only because it is related to fibrosis. In early analysis, TNIK was also linked to multiple typical aging mechanisms.

On the other hand, Insilico Medicine has long been engaged in aging and longevity research, so during the Phase IIa clinical trial, the research team left consecutive blood samples in advance, hoping to see: will a drug originally used to treat lung disease accidentally touch on the "anti-aging" effect.

Interestingly, the "younger" signal discovered this time does not completely coincide with the improvement of lung function.

In the previous Phase IIa trial, the group with the most significant improvement in lung function was the 60mg once a day group; but the group with the most stable change in the aging clock was the 30mg twice a day group. The researchers also compared the change of FVC, which is forced vital capacity, with the change of predicted biological age, and found that there was no strong correlation between the two.

This means that the blood appearing younger cannot at least be simply explained as "just because the lung disease has improved a little".

The research team also compared the protein changes of these patients with the natural aging trajectory of more than 50,000 elderly people in the UK Biobank. In the 30mg twice a day group, the changing direction of some proteins was indeed opposite to the normal aging process.

A drug originally used to treat pulmonary fibrosis thus put forward a completely different hypothesis:

Could it be that what it changes is not just the lungs?

03

How far is "3 years younger" from real anti-aging?

As mentioned earlier, Insilico Medicine has long been engaged in aging and longevity research.

This is closely related to its founder Alex Zhavoronkov, who is a very typical "longevist" in this field. Zhavoronkov left the IT industry at the age of about 24 and devoted himself to aging research. The reason is very simple — if humans really want to go further into the universe one day, we ourselves must live longer first.

He has been on this road for more than 20 years.

In his statements in recent years, he has repeatedly emphasized that humans can already enter space, communicate with machines, and create virtual worlds, but it is still very difficult to make ourselves live a few more years.

Facing the "3 years younger" result this time, Zhavoronkov appeared quite restrained.

In an interview with *The Wall Street Journal*, Zhavoronkov directly poured cold water on the claim that "AI will soon double human lifespan". He believes that no drug has been proven to truly extend human lifespan through rigorous human clinical trials so far.

However, he also said that if the goal is reduced from "doubling lifespan" to "reversing part of the biological age", he believes it is not completely impossible in the next ten years.

And the real question this paper needs to answer is here: What exactly does it mean that the six aging clocks are all turning back together?

First of all, the sample size is very small. This proteomic analysis finally included only 42 pulmonary fibrosis patients. After being divided into four groups, there are only about 9 to 11 people in each group. Although the six clocks are consistent in direction, they still analyze the same batch of patients and the same set of proteomic data.

S