The "death by Claude" meme asks whether your SaaS product is just a markdown file waiting to be replaced by a Claude skill. It is funny until you work in an industry where the answer has serious consequences.
I am Chief Data Science Officer at Ignota Labs. Our AI-driven approach to drug rescue is underpinned by our computational tools that tackle the complexity of human toxicity. We are, by any reasonable definition, bullish on what AI can do for drug discovery. And I think the industry is having the wrong conversation about being replaced by AI. Too much reassurance from people who have not stress-tested these tools. Too much panic from people who have not used them. Here is what I actually think.
GPT-Rosalind is OpenAI's frontier model built specifically for life sciences — reasoning over molecules, proteins, genes, pathways, and disease-relevant biology across multi-step workflows: literature review, sequence-to-function interpretation, experimental planning, data analysis. Anthropic is positioning Claude as infrastructure for the entire scientific process, connected to PubMed, Benchling, and 10x Genomics, with domain-specific agent skills and explicitly targeting everything from early discovery through to regulatory submission. On Protocol QA, a benchmark measuring understanding of laboratory protocols, Claude Sonnet 4.5 scores above the human baseline. These are not general-purpose tools being stretched into biology. They are frontier models aimed directly at the scientific stack.
Closer than most scientists want to admit. Further than the hype suggests. The gap between those two positions is where people will get hurt.
At Ignota we have run GPT-Rosalind and Claude Sonnet 4.5 models on real drug discovery problems. The outputs appear confident, fluent, scientifically literate. And then a scientist looks at them. The model missed a known confound. It made a wrong assumption about the biology. It ignored what the assay can actually measure. It was not vaguely wrong — it was specifically, confidently, articulately wrong. The kind of wrong that is genuinely difficult to detect unless you have spent years in the problem.
In drug discovery, a wrong hypothesis at the bench does not cost compute time. It costs weeks, reagents, and potentially an entire programme direction. The organisations treating AI as a straight replacement for scientific expertise are making a bet that will be expensive to lose.
This one is less debatable. Tasks that took hours take minutes. One person with good AI tooling can do what three people did before. The real question is not whether individual data scientists will be fired tomorrow, — it is whether organisations will hire three where they previously hired ten. At Ignota Labs, we have been riding this wave, but without it, would we be as far ahead as we are? Would we need to have hired 3x the amount of staff? If I truly reflect, almost certainly.. The data scientists who survive will be the ones who can do what the model cannot: know when the output is wrong and make the call that matters.
This is the most immediate disruption. AI frontier models can now perform tasks that previously required expensive purpose-built bio and cheminformatics tools. The companies charging significant licence fees to do things a frontier model can now do are being hollowed out from below. If your core value proposition is doing something Claude can do with a well-designed skill and a PubMed connector, the timeline on your business model is shorter than your current ARR suggests. The moat has to be proprietary data, deep integration, or regulatory validation. Not the workflow itself.
At a recent life sciences investment panel, a panellist asked me directly: what would you do if Claude could replicate your entire tech stack just by being asked to copy Ignota Labs?
It is the right question. And the honest answer is: someone probably could replicate high-level elements of what we do on the surface. The workflows, the report generation, the literature synthesis, the basic ADMET flagging. A determined competitor with good prompt engineering and access to the same public models could get surprisingly close to an Ignota Labs lookalike. The interface of what we do.
What they could not replicate is why we trust it, where we do not, and what we have learned the hard way. This includes the proprietary data, the validated benchmarks and the failure modes we have already encountered and corrected for. Claude does not have this accumulated knowledge; it is not a prompt. The accumulated knowledge of where our models misbehave and why. Claude does not know that. We do. That is not a prompt.
The question assumes the tech stack is the moat. For any serious AI drug discovery company it should not be. The moat is the data, the validation, the scientific credibility, and the track record. If your entire defensible position is the workflow, this question is the one you should be losing sleep over.
The floor has risen. The ceiling requirement has not moved. Someone still has to know when the model is confidently wrong. In drug discovery that person is the difference between a pipeline decision and an expensive mistake.
Use the tools. But be precise about where the human stays in the loop. The organisations that understand that distinction will move faster. The ones that do not will find out the hard way, probably at the point in the programme where it is most costly to find out.
Author: Layla Hosseini-Gerami
