OpenAI's GPT-Rosalind: An AI for Biology Under Control

OpenAI Unveils GPT-Rosalind, an AI Dedicated to Biology

Last Thursday, OpenAI introduced a new artificial intelligence model, GPT-Rosalind, designed to support scientific research in the field of biology. This model pays tribute to Rosalind Franklin, a 20th-century chemist recognized for her significant contributions. GPT-Rosalind is built on the latest technologies developed by OpenAI and is currently available in a testing version via ChatGPT, Codex, and the OpenAI API. However, this access is limited to certain validated users as part of a secure program.

Advanced Technical Capabilities

GPT-Rosalind represents a major advancement in the development of AI for scientific environments. Unlike previous models, this one is designed to delve into technical areas such as genomics, protein engineering, and chemistry. It does not merely skim these subjects but explores them in depth, cross-referencing data, drawing conclusions, and imagining credible biological pathways. It is also capable of organizing complete experimental protocols, tasks that previously took researchers years to accomplish.

To assess its performance, OpenAI subjected GPT-Rosalind to several recognized tests in the industry. On BixBench, a benchmark for evaluating performance in bioinformatics and real data analysis, GPT-Rosalind ranked at the top among models that published their results. On LABBench2, another more targeted test, it surpassed GPT-5.4 on six out of eleven tasks. The standout performance was particularly noted on CloningQA, a demanding exercise where it had to design reagents for molecular cloning protocols from scratch.

Collaboration with Dyno Therapeutics

The most telling test came from the field. In collaboration with Dyno Therapeutics, the model was presented with novel RNA sequences, free from noise data. Its mission was to predict and generate proteins related to their function. As a result, in the Codex environment, its proposals exceeded 95% of human experts for prediction tasks. For sequence generation, it reached the 84th percentile.

Restricted and Secure Access

With a model capable of handling sensitive concepts such as biological structures, OpenAI cannot afford to be generous. The company is thus relying on a structured program designed to prevent any risky usage. GPT-Rosalind is initially being released as a research version, reserved for a select group of companies in the United States. This launch is based on three well-defined pillars: a focus on collective interest, strict rules, and carefully filtered access.

Interested organizations cannot simply sign up. They undergo a thorough verification phase. The goal is to ensure that the work conducted is serious and useful, with a discernible positive impact. Only validated users can access it, in monitored and well-structured environments. However, they must also play by the rules. They are required to implement strict mechanisms to prevent any misuse and accept specific conditions related to this early access.

On the security front, the model has been designed with enhanced protections suitable for professional use. A small surprise, however, comes in terms of budget. During this testing phase, using the model does not incur any consumption of traditional credits or tokens. Researchers can thus experiment without immediate financial pressure.