Anthropic’s AI Claims a New Enzyme, but Experts Question Its Impact

From AI to Biochemistry: How Claude Made the Leap

A team at Anthropic announced that its language model, Claude, had identified a previously unknown enzyme that could accelerate a key chemical reaction. The claim was made in late September, sparking excitement and skepticism across the scientific community. The announcement came after months of internal research and a series of public demonstrations that showed Claude’s ability to generate plausible biochemical hypotheses.

The discovery centers on a protein that could catalyze the conversion of a common industrial substrate into a more valuable product. According to Anthropic, the enzyme was generated by prompting Claude with a series of chemistry prompts and then refining the output through iterative testing. The company reported that the enzyme’s activity was comparable to that of a known catalyst, but it was not clear whether the protein was truly novel or simply a reassembly of existing motifs. Critics point out that the model’s output may be a recombination of known sequences, and that experimental validation is still pending.

Anthropic’s approach involved feeding Claude large datasets of protein structures and reaction mechanisms. The model then proposed a sequence that could fold into a functional enzyme. The team used a high-throughput screening system to test dozens of candidates, eventually isolating one that showed measurable catalytic activity. The process, according to the company, took only a few weeks, a fraction of the time required for traditional protein engineering. This rapid turnaround demonstrates the potential of generative AI to accelerate discovery in life sciences.

However, the methodology raises questions about reproducibility. The team did not publish the full dataset or the prompts used, limiting external verification. Moreover, the reported activity was measured under laboratory conditions that may not reflect industrial settings. Researchers emphasize that a single promising result does not equate to a breakthrough; the enzyme must be scalable, stable, and economically viable before it can be considered transformative.

Originality Under Scrutiny: What Scientists Say

When asked about the claim, several biochemists expressed caution. One researcher noted that AI models often produce „plausible but not novel” sequences, especially when trained on vast amounts of existing data. Another expert pointed out that the enzyme’s sequence shares 85% identity with a known family, suggesting it may be a variant rather than a new discovery. A third scientist highlighted the need for peer-reviewed publication and independent replication before accepting the result as a genuine breakthrough.

Anthropic has responded by inviting collaboration with academic labs to test the enzyme under diverse conditions. The company also stated that it will release the full experimental protocol in the coming months. While the scientific community remains intrigued, the consensus is that the announcement is a promising step, but far from a definitive leap in biochemistry.

The implications of this development are significant. If validated, the enzyme could reduce the cost of producing a key pharmaceutical intermediate, lowering barriers for small manufacturers. It could also open new avenues for sustainable chemical production, aligning with global efforts to reduce carbon footprints. Yet, the road from laboratory to market is long, involving regulatory approval, scale-up challenges, and economic analysis.

Frequently Asked Questions

In the meantime, the episode highlights the growing intersection between AI and experimental science. It also underscores the importance of transparency and rigorous validation in AI‑generated claims. As more organizations explore generative models for hypothesis generation, the scientific community will need robust frameworks to assess novelty and impact.

Q1: What does it mean that Claude „discovered” an enzyme? A1: The model generated a protein sequence that, when synthesized, showed catalytic activity. The claim is that the sequence was not previously documented.

Q2: How reliable is AI‑generated biochemical data? A2: AI can propose plausible sequences, but experimental validation is essential. The accuracy depends on the quality of training data and the model’s architecture.

Q3: Will this enzyme replace existing industrial catalysts? A3: It is too early to say. The enzyme must be tested for stability, scalability, and cost-effectiveness before it can compete with established catalysts.