When Generative AI Starts Doing Science: Inside the Models Designing New Molecules
By reviewsideal Staff | September 27, 2026 | 8 min read
Generative AI made its name writing emails and drawing pictures. Now it is being pointed at biology's hardest search problems: finding new proteins, new enzymes, and new drug candidates hiding in vast molecular datasets. This week gave us one of the clearest examples yet, and a useful reminder of how much still separates a promising lead from a real discovery.
From Predicting Structures to Proposing Molecules
AI's role in molecular science has moved through distinct phases. The first wave, led by systems like AlphaFold, focused on prediction: given a protein's sequence, work out its three-dimensional shape. The second wave is generative: instead of analyzing molecules that already exist, models propose new ones, from protein designs to drug-like compounds, then rank candidates for lab testing. The third wave, now emerging, is agentic: AI systems that explore datasets on their own, form hypotheses, and hand a short list of candidates to human scientists to test.
The reason this matters commercially is simple. Traditional discovery is slow and expensive because the search space is enormous and most candidates fail. A model that can screen millions of possibilities in hours and surface a few worth testing changes the economics of the front end of research, even if the slow, costly clinical and validation stages remain unchanged.
Case Study: 950 Agents, 21 Hours, One Unusual Enzyme System
On September 23, Anthropic announced a new life sciences research group and lab, along with its first result. According to the company, roughly 950 Claude agents searched a DNA sequence database for about 21 hours, using around 210 million tokens. The agents gathered more than 200,000 reverse transcriptases (enzymes that copy RNA into DNA), flagged about 3,500 candidate systems, and narrowed those to 20 for detailed reports.
One agent then noticed something in the raw DNA next to an unusual reverse transcriptase family: a long array of evenly spaced repeats that had apparently not been recognized before. Anthropic calls the combination array-associated reverse transcriptases, or ART. The pattern resembles CRISPR arrays, and the company notes that a small number of similar programmable systems, capable of cutting, copying, or inserting DNA, are already being developed as tools. Initial experiments found that the repeat array produces short RNAs.
CRISPR pioneer Feng Zhang of MIT and the Broad Institute reviewed the work and called the identification of these RNA-repeat arrays "genuinely intriguing," while saying it merits further investigation.
Read this before repeating the headline: the system's biological function is still unknown, the accompanying preprint has not been peer reviewed, and it has not been shown to be a usable gene-editing tool. The CRISPR comparison refers to the repeat-array pattern, not to a demonstrated capability. Even the company's CEO has said its significance is not yet clear.
Discovery Is Not the Same as Design
It's worth being precise about what this result is. The agents did not invent a new molecule; they found an unrecognized pattern in existing natural sequences. That is discovery by mining, and it differs from de novo design, where a model generates a molecule that never existed in nature. Both are valuable, and they often work together: mining reveals what nature already does well, and design adapts those ideas for new purposes. Anthropic says this kind of genome mining could take an expert scientist weeks or months by hand, which is where the time savings show up.
The Money Is Following the Science
Investment is moving in step. Anthropic paid $400 million for Coefficient Bio in April, and OpenAI launched GPT-Rosalind, a model aimed at life sciences research, the same month. In the same week as the enzyme announcement, biotech startup Enveda secured $311 million to bring more nature-derived, AI-discovered drug candidates into clinical trials. The pattern across all three: general-purpose AI labs and specialized biotech firms are converging on the same target, using AI to find promising biology faster and then leaning on wet-lab experiments to prove it.
Three Modes of AI in Molecular Science
- Predictive: forecasting the structure or behavior of known molecules
- Generative: proposing entirely new proteins or drug-like compounds
- Agentic: autonomously mining data, forming hypotheses, and shortlisting candidates for the lab
The Bottleneck Hasn't Moved
Every serious team in this field says the same thing: computation is getting cheap, but proof is not. A model-generated candidate is a hypothesis, and turning it into a validated result still requires human scientists, physical experiments, and time. Anthropic's own setup reflects this: human scientists do all the lab work in a Bay Area facility restricted to lower biosafety levels, with no pathogens that infect humans. The realistic near-term gain is not "AI cures disease" but shorter, better-targeted searches that put more promising candidates in front of researchers sooner.
The Bottom Line
Generative and agentic AI are becoming genuine research tools in molecular biology, not just demos. The enzyme result is an early lead rather than a breakthrough, and its most honest framing is the one its own authors use: interesting, unexplained, and awaiting experiments. The bigger story is the workflow it demonstrates. When hundreds of agents can scan a genomic database overnight and hand scientists a short list worth testing, the pace of the first step in discovery changes, and everything after it gets to start from a better place.
Want to follow AI's push into biology and medicine? Follow reviewsideal as the results from these labs get tested.