AI
Microsoft launches Quine, an AI system bridging computation and wet-lab biology
A new multimodal AI research platform integrates genomics, proteins, and imaging data to accelerate discovery by working in a closed loop between computational modeling and real-world laboratory experiments.
Microsoft Research Blog
Microsoft Research has unveiled Quine, an AI-powered discovery platform designed to address the limitations of traditional biological research by combining computational prediction with real-world experimentation. The system integrates a unified world model spanning genomics, proteins, chemistry, cell state, and imaging alongside orchestration tools that connect researchers, literature, and laboratory equipment in a continuous feedback loop.
Why biology needs a new approach#
Experimental biology faces inherent constraints: laboratory work moves slowly, and the most pressing biological questions often involve interactions and design spaces too vast to explore through conventional testing alone. While automation has improved throughput, nature remains fundamentally complex and resistant to first-principles derivation. Recent advances in foundation models and reasoning systems offer a potential solution—technologies that can integrate information across domains and reason iteratively through multistep problems, similar to how they have transformed language processing.
This observation prompted a central question driving Quine's development: could an AI system learn a coherent model of biology capable of predicting how biological states change in response to interventions? Such a platform wouldn't eliminate the need for experiments; instead, it would enable researchers to computationally explore and rank promising directions before allocating limited laboratory resources.
How Quine bridges modalities#
Unlike systems that orchestrate separate specialist models for different data types, Quine trains a single world model jointly across multiple biological modalities and scales. By learning shared representations across genomic sequences, protein structures, chemical compounds, cellular states, and microscopy images simultaneously, the platform allows evidence from one layer—say, protein interactions—to inform predictions in another, like cell behavior. This multimodal integration reflects the interconnected nature of biology itself, where understanding any single level requires context from adjacent scales.
The system generates computational predictions about how compounds or interventions might shift cellular states or drive therapeutic outcomes. These computational hypotheses feed directly into the experimental workflow, where laboratory teams validate top-ranked candidates through wet-lab assays. Results then return to the platform as training signal, refining the model and accelerating the next research cycle.
