Laptop screen says "back at it, lucho".

Claude for Life Sciences: Both Sides

Recently, there has been a big push by tech companies such as OpenAI and Anthropic to specialize their commercial large language models for bioinformatics. Their bold claims raise a lot of questions: is AI going to render bioinformaticians obsolete? Is this all marketing or is it actually possible? Is this going to revolutionize the drug industry? The arguments for and against the adoption of these LLMs are explored in this month's blog post

BEGINNER FRIENDLY

Sohum Bhardwaj

10/30/20255 min read

Intro

Recently, there has been a big push by tech companies such as OpenAI and Anthropic to specialize their commercial large language models for bioinformatics. OpenAI has partnered with ThermoFisher after having launched its "AI for sciences" package this September. This October, Anthropic has followed suit with their "Claude for Life Sciences" system.

Of course, many researchers are working on developing LLMs for specific tasks like protein engineering and DNA interpretation. However, these big tech companies are going for a completely different approach

In the words of Anthropic themselves,
"Our goal is to make Claude capable of supporting the entire process, from early discovery through to translation and commercialization"

These bold claims raise a lot of questions: Is AI going to render bioinformaticians obsolete? Is this all marketing or is it actually possible? Is this going to revolutionize the drug industry? Can Claude be my therapist?
okay, maybe not that last one

This blog post

01 Intro

02 The Case for Claude

03 The Case Against Claude

04 The Jury's Verdict

The Case For Claude

Automating Tedious Literature Reviews

Claude for Life Sciences is quite ambitious. We'll go up from the most achievable goals to the ones that garner scrutiny from bioinformaticians.

Anthropic highlights Claude's ability to connect with established 3rd party services in the bioinformatics field, one of those being PubMed! If you have ever used PubMed for a literature review, I'm sure you can see where this is going. Claude can read a csv file and then search the PubMed database for genes of interest!

There are many tools that allow scientists to use AI for literature reviews. However, in my opinion, this is a step up. Anthropic is abstracting away the research papers. Seamlessly, you should be able to ask Claude questions about a specific gene or cluster of genes and obtain results from several research papers. It can even detect holes in the literature around certain genes!

I have had the experience of sifting through PubMed articles when I did the Waksman summer experience program at Rutgers. While researching a single gene can be done relatively quickly, simply searching up the name and browsing the abstracts of several papers, researching ten or a hundred would take quite a while. This is why I am excited to see the potential of Anthropic's system.

Integrating with 3rd party Software

While Anthropic has announced partnerships with several 3rd party bioinformatics software apps, two stand out: 10x Genomics and Benchling. Benchling is a platform that helps bioinformaticians manage projects and 10x genomics offers single cell analysis tools.

Many have seen Claude's potential to streamline documentation, potentially by creating records in benchling. Anthropic also advertises its ability to reference documents in Benchling to answer queries.

10x genomics offers a suite of tools for spatial and single cell sequence analysis. However, they require command line expertise that many biologists do not have. It is great that Claude is offering the ability for biologists to work with these tools without having to learn the ins and outs of the terminal.

Many of the companies that work with these tools want to be partnered with a tech giant. 10x Genomics, for instance, saw a 9% increase in stock following the Claude for Life Sciences announcement. I fear that this may influence the way that both companies advertise the capabilities of this new system.

Generating Experimental Ideas

Anthropic brags that Sonnet 4.5, their latest model, has scored an 0.83 on a benchmark that measures knowledge of laboratory protocols. The human baseline is 0.79, but benchmarks are often skewed. The real test of laboratory knowledge will be when Claude is responding to users themselves.

Given this knowledge of laboratory proceedings, Anthropic claims that Claude will be able to generate hypotheses and conclusions from scientific research. This tool has the potential to offer a different perspective, one backed by the entire world's scientific knowledge, yet it could also be a burden to verify that Claude has based its conclusions on real data.

Automating Data Analysis

This is the most hyped up part of Claude's repertoire; Anthropic advertises Claude's ability to perform scRNA-seq quality control right on its website! You can read more about what the process entails here. We will get into the debates over this in the next section, don't worry!

Claude also offers the ability to generate slide decks for presentation. This feature can be extremely useful considering Claude has partnered with BioRender to be able to access a suite of diagrams and images of cells. The trailer also showcases the ability to make small changes through querying the chatbot.

The Case Against Claude

Reproducibility is one of the biggest issues in scientific research. It is necessary for professionals to be able to reproduce the same results across multiple analyses to validate their findings. AI is notorious for producing different results with each query. While this behavior is great for generating compelling text, it is problematic when researchers want to cross-check their work. It is a common criticism of AI that it will exacerbate these problems in research.

As Dean Lee, a computational biologist, puts it:
"It’s going to be a reproducibility nightmare. 'But Dean, my top 10 genes are not the same as your top 10 genes, why is that?' Well, I wouldn’t know unless you can tell me what your Claude-produced files/code mean"

Reproducibility

LLMs are trained to sound authoritative and knowledgeable. This can lead to them being confidently wrong when generating analyses or reporting on genes. These hallucinations can happen more often then you think, try asking an LLM a specific question on a topic you are highly knowledgeable about. I guarantee you will be surprised.

This can be an issue when asking LLMs for technical help. Writing the wrong thing into a terminal can completely delete your project or ruin the data in an insidious manner. And even if no harm is sustained, the loss in time from chasing a dead end is frustrating.

Due to this reason, many anticipate that the need for computational biologists will increase. Instead of performing analyses, in an alternate world, they might see themselves reviewing and verifying the analysis performed by LLMs like Claude.

Confidently Incorrect

While Anthropic officially presents an option to opt out of data collection, there have been many infamous times when companies have harvested data without the user knowing. This worry should be amplified since we are going to be trusting Anthropic with private likely personal clinical data.

Potential Ethics Crisis

The Jury's Verdict

Cautious Optimism

Claude certainly presents huge potential for time savings. Specifically in the case of literature reviews and slide deck generation. Claude's system has the ability to integrate AI seamlessly into the process of research enhancing productivity.

By building pipelines we can already streamline the process of quality control and data analysis. The use of AI in these places will likely cause more headaches than it's worth.

Anthropic has admitted that AI bypass the process of clinical trials, which take up most of the time and money during drug discovery. However its tools still have the potential to improve the efficiency of research by eliminating a lot of time spent searching for information.

As for whether AI will replace bioinformaticians, it seems like the opposite may happen. If more biologists start to use AI without experience in the underlying principles of data analysis, there will be a greater for bioinformaticians to review their work for errors.