Claude cracked a lab’s raw chemistry data in under 25 minutes, here’s how
Follow Us
Any molecule created by a chemist will have to answer two fundamental questions: whether it is the right molecule that they are after and what is the purity level of their creation. Addressing these questions usually takes an hour out of the time of a seasoned chemist for each sample; sometimes even more than that. Anthropic has proven that Claude can address these issues within 25 minutes, using nothing but instrument files and a two-sentence prompt.
SurveyAlso read: OpenAI Astra: Critical cybersecurity threshold explained, what exactly happened?
In the problem mentioned above, there are two methods which are the backbone of synthetic chemistry – NMR (nuclear magnetic resonance) spectroscopy and LC-MS (liquid chromatography-mass spectrometry). NMR allows determining the molecular structure by analysing the signals generated by the hydrogen atoms in the molecule. LC-MS is a method which breaks the molecule down into parts and measures their weight, which is what purity is based on. The instrument analysis takes minutes. The challenge has always been to go from the instrument raw output file (which is in a format unique to the instrument vendor) to something a human can understand.
As part of its normal practice, Anthropic provided Claude with the raw files for a quality control sample from a contract lab. It did not supply software from the vendor, nor anyone to stand by the side of the chemist, but simply the raw files and one single-line command that any chemist could give. Claude provided back complete NMR results in 23 minutes and LC-MS results in 19 minutes, simultaneously.
Not only were Claude’s analyses fast, but they produced results consistent with those produced by the lab itself. The purity percentage reported by Claude was 96.4% versus the lab’s 96.33%. Its hydrogen counts from the NMR peaks were within 0.08 of those of the lab. For the LC-MS file, which used a vendor format unknown to the chemist, Claude deciphered the code and validated his decoding of the file data by re-producing the total number of scans performed by the instrument.
The real value is not just in the speed but in the judgement. While classifying four wide peaks on an NMR as possible hydrogens bonded to nitrogen or oxygen, it made the suggestion of following up with an additional test by introducing deuterium oxide into the sample to determine the peaks’ disappearance. It turned out to be exactly what the lab performed three days after making the initial measurement. Having received the follow-up data, Claude identified and corrected its own mistake, since it had initially reported the disappearance of all four peaks, while only two had really disappeared.
Also read: ChatGPT for Teens: OpenAI’s solution for India’s huge student population using AI
As for the timing, it should be noted that the completed lab report was delivered to the lab four days after making the first measurement, which is not an uncommon delay, considering how labs process samples one-by-one. Claude provided the analysis along with the report within 25 minutes.
This is not about Claude learning how to conduct experiments that have never been conducted before by anyone else. This is about Claude being given the boring, routine work of a chemist that takes time to complete but hardly ever involves deep PhD-level insights, and doing it at lightning speed. This is not the most spectacular advance in the field of AI but definitely the most practical one; it will mean a huge difference between four days of waiting for lab results and having them in your hands by lunchtime.
Anthropic now offers a similar workflow via Claude Science, where you can give a model a raw NMR or LC-MS file and ask it to verify the identity and purity of a compound.
Also read: A day with the spectacular Hisense 116UX RGB Mini LED TV that costs Rs 30 lakh
