AI System Beats Scientist at Finding Repurposed Liver Fibrosis Drugs

An AI research assistant called Co-Scientist outperformed a Stanford geneticist at finding drugs that could be repurposed to treat liver fibrosis, according to a study published in Advanced Science.

Gary Peltz, a geneticist at Stanford University School of Medicine, ran a head-to-head test. He asked Co-Scientist to propose three drug candidates for fibrosis and explain its reasoning. Separately, he picked two candidates himself, drawing on his own reading of the liver fibrosis literature and choosing drugs with a notable presence in it.

Then he put all five drugs through the same test: his lab's fibrosis testbed built from live human liver cells.

The results were lopsided. Peltz's two picks showed no benefit against fibrosis. Of Co-Scientist's three picks, two blocked fibrosis and promoted the regeneration of liver cells.

The stakes are high. Liver fibrosis is a scarring process that can progress to cirrhosis, which causes more than 1.4 million deaths each year. Any tool that accelerates the discovery of medicines able to slow, stop, or reverse fibrosis could matter far beyond hepatology, because fibrotic scarring also affects other organs.

The needle in the haystack

One detail from the experiment stands out. One of Co-Scientist's successful candidates had been linked to liver fibrosis in only a handful of papers — what Peltz's team describes as a needle in the haystack of scientific literature.

That result cuts to the core of why drug repurposing is hard. The literature on existing medicines is vast. Human experts gravitate toward the drugs they see cited most often, and in this experiment that instinct failed. The two drugs with the strongest presence in the fibrosis literature did nothing in the testbed. The AI, by contrast, surfaced a barely documented candidate that worked.

The standout pick was vorinostat, a cancer drug. In Peltz's experiments, vorinostat blocked 91% of a damage response that can drive liver scarring. That is a concrete, quantified result from live human liver cells, not a computational prediction left on paper.

A different theory of the disease

Co-Scientist's reasoning pointed toward drugs that reshape gene activity rather than drugs that target a single fibrosis pathway. That distinction may prove to be the most consequential finding in the study.

Most anti-fibrotic drug development has focused on blocking specific molecular pathways implicated in scarring. Co-Scientist's successful suggestions instead shared a mechanism class: epigenetic reshaping of gene activity, with vorinostat — a drug already approved for cancer — the leading example.

Peltz argues that such drugs deserve serious consideration as treatments for liver fibrosis. If he is right, the study could help launch a new generation of anti-fibrotic medicines built on a broader mechanism than the field has traditionally pursued.

The repurposing angle carries practical weight of its own. Vorinostat is already an approved cancer drug, with an existing safety record and manufacturing chain. Repurposed medicines can move toward clinical testing faster than novel compounds, and for a disease that kills more than 1.4 million people a year via cirrhosis, that timeline matters.

A controlled comparison, not a hype exercise

The design of the experiment deserves attention. Peltz did not simply ask an AI for ideas and publish them. He generated his own candidates through the traditional route — reading the literature and selecting the drugs most prominently associated with the disease. He then subjected both sets of candidates to identical testing in a live human liver cell testbed.

That structure makes the comparison meaningful. The same lab, the same assay, the same disease model, and two different methods for choosing candidates. The expert's method produced zero hits. The AI's method produced two out of three.

The sample size is small — three AI picks and two human picks. One experiment cannot establish that Co-Scientist is broadly superior to expert judgment in drug discovery. But it demonstrates something specific: an AI system reasoning over the existing pharmacological literature found a clinically relevant signal that a domain expert, using the same literature, missed.

Why this matters now

Co-Scientist belongs to a growing class of AI tools built to support scientific reasoning rather than simply answer questions. Peltz's team designed the study explicitly to test whether such a system could support efforts to identify repurposable drugs from the vast literature of existing medicines. The Advanced Science paper is one of the clearer published answers so far, because it pairs AI-generated hypotheses with wet-lab validation.

For the pharmaceutical industry, the implication is direct. Literature-scale reasoning is a task where human bandwidth fails — no researcher can hold a "handful of papers" buried across decades of publication in mind while also tracking the heavily cited candidates. AI systems can search that space without the citation-frequency bias that led Peltz's own picks astray.

For researchers, the implication is subtler. The winning mechanism — broad reshaping of gene activity — emerged from the AI's reasoning, not from the field's consensus view of fibrosis. AI hypothesis generation did not just accelerate the search; it redirected it toward a different theory of what an anti-fibrotic drug should do.

What comes next

Peltz's stated hope is concrete: drugs that reshape gene activity should be seriously considered as liver fibrosis treatments, and this line of work could ultimately help launch a new generation of anti-fibrotic medicines. The next steps for any repurposed candidate are the standard ones — further preclinical validation and, if vorinostat or its relatives hold up, clinical testing in fibrosis patients.

The broader test is whether Co-Scientist's success here replicates in other labs and other diseases. If AI-proposed, literature-mined candidates keep surviving contact with live human cells, the bottleneck in repurposing may shift from finding candidates to validating them — a far better problem to have for a field that currently loses 1.4 million patients a year to cirrhosis.