From calculating a single molecule to searching for unknown life systems: Sixty years of AI in medical research

From the quantitative structure-activity relationship in 1964 to AlphaFold, AI scientific research Agent, and then to Claude's discovery of ART: What AI has truly changed is not pressing the fast-forward button for the 10-year R&D cycle all at once, but that different time dimensions in medical research are being redistributed.

21 hours, approximately 950 AI Agents, 210 million Tokens. Claude noticed a previously uncharacterized system in phage DNA. But this article does not discuss "the next CRISPR". What I care more about is: over the past 60 years, which time dimensions in medical research are being taken over by AI; and which time dimensions still belong to life itself to this day.

In September 2026, Anthropic assigned Claude a task that did not sound astonishing: to find some reverse transcriptases worthy of further research in a huge DNA database.

In the next 21 hours, about 950 AI Agents participated in the search, consuming 210 million Tokens. They collected more than 200,000 reverse transcriptases, screened out 3,500 new candidate systems, and finally narrowed down to 20 objects most worthy of further analysis.

When one of the Agents continued to look forward along a segment of phage DNA, it noticed a string of unusually neat repetitive sequences arranged next to the reverse transcriptase.

It stopped.

It counted the number of repetitions, measured the spacing, compared it with known systems, and then checked whether anyone had reported it before.

In terms of the organization form of the repeat array, it is reminiscent of CRISPR.

Schematic diagram: Redrawn based on the public process of Anthropic. The key point is not that "ART is already equivalent to CRISPR", but that the Agent continues to trace from a large number of candidates to an abnormal repeat array in the established task; the main biological function of ART is still under research.

After further analysis and experiments, the Anthropic team identified this pattern as a previously uncharacterized system and named it ART — array-associated reverse transcriptases. The team's preliminary experiments also found that these repeat arrays express a set of short RNAs; but what biological function ART exactly performs still has no answer to this day, and all wet experiments are still completed by human scientists.

So it is too early to call ART "the next CRISPR" now.

However, if we only focus on whether ART can finally become a gene editing tool, we may miss the more noteworthy part of this event: about 950 AIs spent 21 hours entering a DNA world that no scientist could read one by one, and began to participate in judging where it is worth stopping and taking a second look within a research problem defined by humans.

Which time dimensions in medical research are being taken over by AI?

The "medical research" I mentioned here refers to a slightly broader scope: from basic life sciences, drug discovery to clinical verification after entering the human body. Because the 10 years that a drug eventually takes is exactly the superposition of these different time dimensions.

Sixty years ago, a question was first raised: Can we calculate before conducting experiments?

In 1964, American chemist Corwin Hansch and Japanese scholar Toshio Fujita published a paper that later became a classic literature of quantitative structure-activity relationship, trying to use quantitative methods to connect chemical structure and biological activity.

That of course was not AI in the modern sense.

But the question behind it later appeared repeatedly in computational pharmaceutical research: Is it necessary to first synthesize a compound and then test it in experiments to see if it may have a certain biological activity? Or can we calculate it before conducting experiments?

One year later, Edward Feigenbaum and Joshua Lederberg of Stanford started the work later known as DENDRAL. The early documents left in 1965 directly referred to the initial question as a "machine induction research project"; later DENDRAL became a representative case in the history of expert systems, using computers to participate in inferring molecular structures from experimental information.

Schematic diagram: The 1964 quantitative structure-activity relationship and DENDRAL around 1965 respectively represent two early paths of "computabilization of research objects" and "computabilization of scientific research processes". This is a schematic of historical structure, and QSAR is not directly equated with AI in the modern sense.

More than 60 years later, it is very interesting to look at these two things together.

One side is trying to make the objects of life and pharmaceutical research computable; the other side has begun to try to hand over part of the process of scientists moving from evidence to judgment to machines.

AI did not break into medical laboratories suddenly in recent years. It has taken over a 60-year-long process of "computabilization of scientific research".

And one thing that keeps happening in this history is to move the trial and error that could only be completed in the expensive and slow real world to the computing world as early as possible.

The first thing machines took over was a lot of experiments that might be done in vain

Drug R&D has always faced a very simple difficulty: there are too many possibilities.

The number of possible compounds far exceeds the quantity that any laboratory can synthesize and test one by one.

Therefore, for decades, one of the core values of computing entering pharmaceutical R&D is not to cancel experiments, but to try to answer before actually starting work: among so many possibilities, which ones should we test first?

In 2020, a study searching for novel antibacterial compounds demonstrated this change very intuitively.

Researchers first used a deep learning model to find the candidate later named halicin in the Drug Repurposing Hub; in the same study, the model was applied to more than 107 million ZINC15 molecules. Finally, only 23 high-priority predictions were actually taken to the laboratory for testing, 8 of which showed antibacterial activity.

More than 100 million, and 23.

It is not that the experiment is suddenly 100 million times faster, but that 100 million possibilities no longer require 100 million experiments.

Bacteria do not grow faster. The time for drugs to act on cells has not changed. What is compressed is a large number of searches before the experiment.

In many cases, AI does not make the real world run faster. It just allows researchers to avoid many dead ends that they would not have known were dead ends until they walked into them.

After AlphaFold, the end point of some people has become the starting point of others

Around 2020, this change took a big step forward.

Protein structure is one of the basic information in life sciences. Knowing the amino acid sequence of a protein does not mean we already know how it works. The three-dimensional structure formed by protein folding is often an important entry point for understanding functions, disease mechanisms and drug effects.

But for a long time, obtaining structures was very expensive.

The AlphaFold 2 paper in 2021 wrote that at that time, there were only about 100,000 unique protein structures obtained through experiments, while the known protein sequences had reached the scale of billions; determining a protein structure might take months or even years of work.

AlphaFold did not make structural experiments lose their value. Proteins are not static sculptures, and model predictions are not equal to all conformations and interactions in the real environment.

What it changed is another thing: a large number of studies can first obtain a high-value structure prediction, and then decide where to ask the following questions.

Schematic diagram: AlphaFold did not cancel structural experiments, but allowed a large number of studies to start from high-value predictions. In 2022, the AlphaFold Database expanded to more than 200 million predicted structures.

By 2022, the AlphaFold Protein Structure Database expanded from less than 1 million predicted structures to more than 200 million, covering almost all catalogued proteins at that time.

This can hardly be understood as a certain experiment being "20% faster". It is more like the emergence of a map.

In the past, when a researcher went to a certain place, he might have to survey the terrain by himself first. Later, with the map, subsequent people can start from the next section of the road.

The deepest acceleration of AI is not necessarily to do the same thing faster, but to allow later people not to start from the same place.

A piece of knowledge that was originally the end point of a research project may later become the default starting point for another research project.

But the human body can still say "no"

When the story is written here, it is easy to have an intuition: the faster the calculation, the more accurate the model, and the drug R&D will naturally get faster and faster.

The reality is not that neat.

In 2020, Sumitomo Dainippon Pharma of Japan and Exscientia of the United Kingdom announced that a candidate drug named DSP-1181 entered Phase I clinical trial.

Sumitomo Dainippon Pharma said in the official announcement at that time that DSP-1181 completed the exploratory research phase in less than 12 months, while the average time of traditional methods cited in the announcement was about 4.5 years.

Less than a year, versus 4.5 years.

If the story stops here, it is almost a perfect poster of "AI accelerating pharmaceutical manufacturing".

But later, the development of DSP-1181 was terminated. Sumitomo Dainippon Pharma later confirmed that it did not meet the criteria set for this mechanism of action in Phase I clinical trials.

The machine sent the drug to the human body faster, but the answer given by the human body may still be "no".

This is not to prove that AI pharmaceutical R&D has failed. It just reminds us: finding a candidate drug that seems worthy of entering clinical trials and proving that it can treat real patients are two different problems.

As early as 2012, Jack Scannell et al. described a later famous anomaly: in the past few decades, computing, molecular biology, high-throughput screening and R&D management have continued to progress, but calculated according to inflation-adjusted R&D investment, the number of new drugs corresponding to every 1 billion US dollars has roughly halved every 9 years since 1950, with a cumulative decline of about 80 times. They called this phenomenon "Eroom's Law" — which is Moore spelled backwards.

Fourteen years later, in August 2026, *Nature Reviews Drug Discovery* re-examined AI pharmaceutical R&D, and the conclusion is still quite restrained: at this stage, the evidence of clinically relevant impact of AI in drug discovery is still limited; the authors believe that the industry needs to shift its attention from "how accurate the model itself is" to "whether the model really improves R&D decision-making".

Drug R&D has never been a whole block of time that can be fast-forwarded all at once.

By 2026, the compressed scope has begun to include scientists' reading and attention

The changes that took place this year have gone a step further.

Co-Scientist, published in *Nature*, no longer just accepts a specific molecule and then predicts a property. Researchers first give the research goal, and multiple Agents then search the literature, generate hypotheses, criticize each other, rank and continue to iterate, and some biomedical hypotheses finally enter experimental verification.

But I prefer another case.

Robin, also published in *Nature*, was used to study dry age-related macular degeneration. It connects literature retrieval, disease mechanism sorting, candidate hypotheses, drug screening, experimental suggestions and experimental result analysis together.

The paper gives a very striking set of numbers: Robin analyzed 551 papers in one process, taking 30 minutes. According to existing academic reading research, the authors estimate that if humans complete the corresponding scale of reading and information processing, it will take about 294 hours; the entire cognitive workflow is estimated to have been reduced from 359-424 manual hours to less than 2 hours.

This is not a strict man-machine competition. 294 hours is an estimate based on existing reading surveys, and the real experiments are still completed by humans.

But even if we greatly discount this number, the change still exists: a large amount of time originally spent on reading, searching, sorting, comparing and organizing hypotheses in scientific research has begun to enter the scope of parallel processing by machines.

In the past, when a scientist faced hundreds of papers, he was first limited by his own time and attention. Now, the new problem has gradually become: the machine has given so many possible directions, which one is really worth spending experimental resources to verify?

Anthropic also encountered this problem. They said that one research campaign may produce hundreds to thousands of candidate reports, so that "which hypotheses are worth experimenting" has itself become a research object.

From calculating a molecule to deciding where to take a second look

At this point, looking back at ART at the beginning of the article, the meaning is a little different.

More than 60 years ago, people first asked: Before the experiment, can we calculate the activity of a molecule?

Later, machines can screen more and more molecules at the same time, predict hundreds of millions of protein structures, sort out hundreds of papers in half an hour, and allow a large number of Agents to move forward along different search paths.

In this ART experiment, the research question was still raised by humans: to find interesting novel reverse transcriptases.

Claude did not decide on its own "what life sciences should study most now". But within the boundaries of this task, it began to choose its own search path: which RT families are worthy of further investigation, which gene neighborhoods are worthy of opening, and which anomalies in a segment of DNA are worth stopping for.

The significance of this step does not need to be proved by "whether ART is the next CRISPR".

The main function of ART is still unknown. Seeing an anomaly does not mean understanding the mechanism. One Agent found it this time, which does not mean that the same method can stably produce important discoveries every time.

It is worthy of being included in this 60-year history because the position of machines participating in scientific research has moved a little further: from answering a well-defined question, to participating in deciding where to look next within this question.

There is always more data than attention. There is always more literature than can be read. There are always more possible experiments than the laboratory can complete.

The ability of a researcher is not only to answer questions, but also to know where it is worth stopping when facing a huge and chaotic unknown world.

What AI changes may first be the lifetime of scientists

Therefore, I increasingly dislike a common question: Can AI shorten the 10-year new drug R&D cycle to 5 years, 2 years, or even a few months?

This way of thinking regards pharmaceutical R&D as a clock.

In fact, there are many completely different time dimensions here.

It takes time to search 100 million molecules. It takes time to obtain a kind of structural knowledge that was expensive in the past. It takes time to read hundreds of papers and form a research hypothesis. It takes time for cells to grow. It takes time for animal experiments. After a drug enters the human body, it also takes time to observe curative effects and side effects.

These time dimensions will not become faster in the same proportion because of the same technology.

Over the past 60 years, one thing that machines have continued to do is to move part of the trial and error that could only be completed in the real world into the digital world in advance.

For anything that can be moved in, the speed may change by an order of magnitude.

100 million molecules can be calculated first. Hundreds of millions of proteins can have their structure predicted first. Hundreds of papers can be sorted out by machines first.

But cells will not grow 100 times faster because of this. Mice will not grow up 100 times faster because of GPU upgrades. The human body will not tell us the curative effect and side effects five years in advance just because Tokens become cheaper.

AI is rearranging the time structure of pharmaceutical R&D.

Some links that used to be expensive and slow suddenly become cheap and fast, so those links that cannot be compressed at the same speed will be revealed more and more clearly.

AI has not eliminated the bottleneck of medical research. It just keeps pushing the bottleneck backward.

One change has already happened: the world we can study has become larger.

Sixty years ago, researchers began to try to let computers help them avoid some experiments that might be invalid. Sixty years later, machines can simultaneously enter many paths that a person is too late to explore one by one.

It still cannot give the final answer for life.

The human body still has its own time, the disease still has its own time, and medicine ultimately has to give the answer to the