Microsoft launches biological world model Quine, multiple giants double down on AI4S

English text is being prepared.

Recently, Microsoft Research launched an experimental AI research system called Quine, aimed at building a multimodal world model of biology and constructing an interactive harness framework that connects models, scientific tools, literature, and researchers, pushing AI from assisted analysis further toward scientific discovery.

According to the introduction, Quine is not trained on a single type of biological data; instead, it simultaneously covers multiple modalities such as genomics, proteins, chemistry, RNA, cell states, and biological imaging, enabling information across different modalities to mutually assist prediction.

In the specific scientific research workflow, after researchers pose a question, Quine can combine scientific literature and research tools to perform computational analysis, generate research plans and experimental designs, and hand the prediction results over to real experiments for validation. The new data generated by the experiments is then returned to the model and the researchers, used for model training and to drive the next round of question analysis, forming a loop of "question posing—model reasoning—experimental validation—result feedback."

Microsoft has partnered with the Broad Institute, jointly established with MIT and Harvard University, to use Quine in cancer biology and other research. A case disclosed by Microsoft shows that Quine predicted and ranked the potential of thousands of compounds to alter the state of pancreatic cancer cells, after which the top-ranked candidate compounds entered wet-lab validation, and the highest-ranked compound produced the largest expected change.

Microsoft says that from narrowing the compound search space to identifying candidates for laboratory validation, the entire process took only one weekend, saving months of experimental work and substantial research costs. However, Quine is still in an early research stage, and its results remain far from actual clinical application.

In fact, Microsoft is not the only tech giant to have recently stepped up its bet on AI4S.

On September 23, Anthropic announced the establishment of a life sciences research team and laboratory, and disclosed early results from Claude's participation in biological research. After analyzing more than 200,000 reverse transcriptase (RT) sequences, Claude selected 3,500 candidate systems and further narrowed them down to 20, ultimately discovering a new enzyme system, ART, that has been preliminarily validated by wet-lab experiments.

In addition, on September 16, Novo Nordisk announced a collaboration with Anthropic to use Anthropic's models and Claude Science for scientific reasoning and related workflows in drug R&D, in order to solve specific drug discovery problems.

OpenAI, for its part, launched GPT-Rosalind for life sciences research in April this year. The model is specifically oriented toward biology, drug R&D, and translational medicine, covering tasks such as medicinal chemistry, genomics, and quantitative biology, and emphasizes tool calling and real-world scientific research workflows. On September 11, OpenAI announced that GPT-Rosalind had ended its research preview and was opening access to eligible institutions.

NVIDIA chose to enter from the angles of computing power, models, and experimental infrastructure. In January this year, NVIDIA and Eli Lilly announced the joint establishment of an AI innovation laboratory, with the two sides planning to jointly invest up to $1 billion over five years in infrastructure and research, focusing on solving complex problems in drug discovery and exploring the application of robotics and physical AI in drug discovery and production.

On September 9, Eli Lilly announced that it would launch an artificial intelligence and machine learning platform called TuneLab, which will open to biotechnology companies AI drug discovery models trained on its years of research data.

Google's Isomorphic Labs, meanwhile, continues to bet on AI-driven drug design. Relying on technology accumulations such as AlphaFold, the company uses AI models for the design of small molecules and other types of drugs, and has carried out cooperation with pharmaceutical companies such as Novartis, Eli Lilly, and Johnson & Johnson.

English text is being prepared.

Founder Securities believes that large models have already landed in pharmaceutical companies, and the business models of computational R&D platforms such as CADD have also been proven, but the scale of publicly disclosed recurring revenue and the final number of marketed drugs remain relatively limited. Key indicators worth tracking going forward include whether cooperation between large models and pharmaceutical companies moves from the testing stage to actual use and fee disclosure, whether the contract value of professional R&D platforms continues to grow, and whether projects with AI participation in R&D can further enter clinical trials and form commercialized assets.

Everyone is connecting to DeepSeek and WorkBuddy—what will ultimately enable brokerages' AI to pull ahead?