Claude Science

I built one of the connected tools included in this launch (the Biomni HPC [1]), and I have spent an inordinate amount of my life working on this problem. (I also worked at Anthropic, but not on this product.) As other comments have pointed out, this is for data science – but it's capable of more than making plots and writing papers [2]. It has integrations with many databases and computational tools, including a researcher's institutional cluster. That alone is valuable. I founded a startup after struggling with this problem at a bio startup; integrating these tools and databases is hard and time consuming. If the only outcome of this product is that great APIs are built for LLMs, it will be a massive positive impact. Many databases used in computational genomics are still only accessible through FTP! LLMs are particularly good at navigating these tools and databases. It's often very specialized, but straightforward, work that benefits from in-context skills. Seeing an early glimpse of my former customers – bioinformaticians – using LLMs to solve this problem is what led me to join Anthropic in 2024. Also, this pattern isn't fundamentally constrained to data science: you can also integrate with a wet lab or a CRO for some kinds of science. This is what I'm spending my time on now. This type of science doesn't solve everything, but it's useful in some niches. For example, progress on many rare diseases is bottlenecked by researcher attention rather than a fundamental breakthrough. [1] https://x.com/phylo_bio/article/2029233694775624096 [2] In comparison, OpenAI's science product – Prism – was effectively a LaTeX editor they acquired with Crixet.

I watched the announcement and gave it a spin as I'm a heavy user of cowork/code. So far I'm super impressed. I used it to analyze my whole genome sequencing data I have as my son has a rare genetic condition. I used it to answer a question I'd asked a few bioinformaticians to help me with but never got a satisfactory answer, it solved it in about a minute - whether his n-of-1 de novo, heterozygous single nucleotide mutation was likely passed down from mom or dad. It performed a read-backed phasing analysis on the data, identified a nearby SNP with overlapping coverage where mom was homozygous and dad was heterozygous. Identified my variant on his mutated allele so looks like it came from me.. It also crosschecked my data against AMCG Secondary Finding genes and ClinVar likely pathogenic/pathogenic variants and came back with identical results to my Natera Horizon carrier screening results. I'd previously tried and failed to do this all with some ChatGPT guidance and subsequently hired a couple of bioinformatician post-docs at top tier universities via Upwork who had failed to give me satisfactory results. And this is just getting started!

I'm a scientist, (biophysicist). Over time I have become a bioinformatician and a python dev. I wrote articles and applications, and it always was a struggle. But now I can speed up, make it all go much faster. But I often feel like my mental models can't keep up. Recently the AI has generated a comprehensive data model (in Django) and I find myself retracing its steps with long discussions and explanations (with/from the LLM) and searching for documentation. With scientific assignments I find myself searching literature on my own, read whole papers as I used to. Checking the LLM constantly but adapting to it and I don't like it, don't like how it steers me, just let me search, let me wander the scientific landscape on my own, let me read the words of the authors with opposing views. Then let me make 20 plots and only use 1, let me wrestle with the data. Let me make wrong visuals that by chance communicate something important about the data. Because otherwise I feel uncomfortable, I need to understand, that is what I do. I can reason about so many things because my internal world model is comprehensive and mostly correct. That has taken 44 years so far. Hard work from time to time, but I've mostly enjoyed it. I still don't know what to make of these models, I use them everyday, but sometimes I wonder if I was not just as fast with Stack Overflow, because what I crave is understanding, not "some finished app". Yes, I rarely finish things fully (that's how I feel), but in research I've often been told they like my ability to move very fast and creatively in phase one, the development is left to others anyway... I crave an understanding of what these tools mean to me exactly. This comment is part of that. HN is part of that.

The most interesting thing here is that Claude Science runs a local server and a web-based UI that connects to that server from your browser. This is very different from Claude Code and Cowork, where the UI is more tightly coupled to the host machine (which makes things like computer use possible). I think I recognize the strategy: most pharma environments connected to interesting data are tightly locked down, to the point where you can't just connect your Macbook to the source data. Similarly, access to large genomic biobank datasets like UK Biobank or NIH's All of Us program is granted only through a Trusted Research Environment (TRE), a remote data analysis platform usually quite restricted on internet access, etc. You can't easily run desktop apps, but these environments do usually support running JupyterLab or VS Code, tunneling the user interface through to the end user. (Source: I previously ran the team that built the All of Us TRE.) Claude Science looks a lot more like something one could imagine spinning up in one of those highly-constrained data environments (with the "server" running within the TRE and the UI proxied to the end user's browser) than the does-everything Claude mega-app. That will be critical for traction within pharma R&D environments. I will say that for moderately-computational scientists, who are daily driving RStudio, JupyterLab, or maybe VS Code, Claude Science will be quite an unfamiliar shaped product. I'll be curious to see whether something like this gains adoption (1) in place of, (2) alongside, or (3) eventually wrapping around the more traditional data science workbench tools out there.

Tried this to see how it goes in my particular field - computational design of RNAi-based biopesticides. One-shotted a design for targeting the DvSnf7 transcript of western corn rootworm. It took a fairly naive approach (maybe how a 1st year PhD student would go about it), but got the job done. Also noted caveats with its approach (e.g. using mammalian design rules, limited off-target screening). Not bad really. But also not great. When its flaws were pointed out, the AI determined that it could have taken a more informed approach. Then Opus 4.8's safety system flagged the session.