Image created via OpenAI model
As 10x Genomics plugs Anthropic’s Claude into its cloud analysis stack, Michael Schnall-Levin, the company’s chief technology officer, is less interested in the novelty of “natural-language pipelines” than in solving a stubborn problem he calls the last mile of single-cell analysis.
For years, 10x has invested heavily in software and visualization tools to automate early-stage analysis. “For many years, our intent has been to democratize these tools,” Schnall-Levin said. “Even before LLMs became as prominent as they are now, we invested a lot in software and visualization to make analysis easier, and that has had real impact.”
The problem is that research questions and experimental designs are highly idiosyncratic. “Experimental setups are extremely varied,” he said. “People ask very different questions and design experiments specifically around those questions. It’s very hard for one piece of software to go all the way from raw data to ‘here is your exact answer’ for every possible design.”
Michael Schnall Levin, Ph.D.
Even with strong software that automates and streamlines the early steps, someone still has to take the outputs, write analysis code, and ombine and compare data from different samples in a way that is specific to their study
A simple example is a 100-person study with 50 treated and 50 controls. “You want to know which cell types or genes change the most in response to the treatment,” Schnall-Levin said. “Our software can give you building blocks, but you can’t just type in that question and get a final answer.”
Traditionally, a bioinformatician would write Python or R scripts to combine datasets, run the right contrasts and summarize the results. That is achievable work, but it requires a skill set that is in short supply, so non-computational researchers end up waiting in line.
“That’s where this can be impactful,” he said. “The last mile goes from ‘non-trivial work, requiring scarce expertise’ to something a much broader set of people can do themselves, without relying on someone else.”
What LLMs actually add
What makes this more than just a nicer GUI is the combination of tool calling and language understanding. “One of the strengths of LLMs is that they can do both: handle structured analysis outputs and read the literature, and then pull everything together in one powerful tool,” Schnall-Levin said.
Anthropic has already previewed this direction with its own MCP tools for PubMed. Schnall-Levin is interested in workflows where Claude can move seamlessly from a 10x analysis into the literature: “Show me all the differentially expressed genes from this analysis.” Or: “Now tell me interesting things about those genes based on all the papers that mention them.”
He sees similar potential in breaking down language and geography barriers in scientific publishing. “When I was in school there was a joke that Russian researchers had already made every discovery 10 years earlier, but it was all published in Russian,” he said. LLMs, which can read Russian, Chinese, English, French and more, are well-suited to mining that long tail of unstructured science.
Clinical data is another frontier. Medical records are often written differently by different doctors, using myriad coding systems, missing some information. “It’s messy and unstructured, but there’s a lot of signal there,” Schnall-Levin said. “Combining that with molecular data is extremely impactful, and I think LLMs could be very important in synthesizing and pulling that information together.”
For now, Claude’s role in the 10x collaboration is mostly orchestration and summarization rather than full-blown hypothesis generation. But the same primitives, tool calling plus language understanding, could eventually support richer, semi-automated workflows that keep scientists in the loop while offloading mechanical tasks.
Will this mean fewer researchers, or more?
Zooming out, Schnall-Levin is skeptical of confident predictions about long-term labor impacts, but his near-term view is more optimistic than fatalistic.
“In the near- to medium-term, I can see the argument for more researchers,” he said. “I don’t see this primarily as replacing research; I see it as accelerating it. Some jobs may get replaced, especially lower-value research tasks, but the counter-trend is that the ROI on research goes up. If it’s easier and cheaper to get to insights, you should see more research, not less.”
He also sees a clear stratification effect as labs adopt AI tools unevenly. “People who embrace these tools, learn how to use them, and actually get value from them could become much more impactful,” he said. “People who don’t embrace these tools will probably be at a disadvantage over time.”
The pattern is familiar from the genomics era. “This is similar to what we saw in biomedical research over the last 20 years or so,” he said. “With the emergence of genomics and large datasets, being computationally savvy became important in a way it wasn’t before. You didn’t necessarily have to write all the code yourself, but having interest in and comfort with computational tools generally helped people. People who embraced that trend often fared better — again, not universally, but in many areas. I think AI will be similar.”
The harder question is what happens if you zoom the lens out to the entire economy. “What this all means for different labor markets is tricky,” he said. “I don’t think anybody can confidently predict the long-term details beyond a few years.”
Why Anthropic?
On the partner choice, the 10x–Anthropic alignment is partly philosophical and partly personal. “First, as a company, they’re very science-forward,” Schnall-Levin said. He met Anthropic CEO Dario Amodei years ago in a graduate fellowship program, when Amodei was working in computational neuroscience and biophysics before moving fully into AI. “Between him and others on their leadership team, they have a deep interest in this area, so there’s good philosophical alignment.”
A more immediate catalyst was the move of Jonah Cool. formerly at the Chan Zuckerberg Initiative, where he worked with 10x on the Billion Cell project, to Anthropic. “When he went to Anthropic, a number of us here quickly reconnected with him, and that’s how the initial conversations started,” Schnall-Levin said.
“What’s nice is that there’s a lot of alignment between us and Anthropic around using these tools for science, and they’re quite forward-leaning in that area,” he added. Amodei’s background also helps. “He understands this world in a way that may be more foreign to some other AI folks.”
Other AI labs are also eyeing science, from OpenAI’s work with national labs to various efforts inside Google. But Anthropic has explicitly named biology and science as strategic focus areas. For 10x, that made the decision to expose its cloud through Claude via MCP less about chasing a trend and more about extending a long-running bet on single-cell data as the raw material for the next wave of biological discovery.
