AI Scientists Discover New Drug in 2.5 Months Autonomously

Scientists have used AI to discover a drug for treating rare diseases in just 2.5 months, completing the entire process from hypothesis generation to validation autonomously. This achievement represents a significant challenge in combining AI with physical experiments, as it requires AI to understand which experiments are feasible, assess their reliability, identify obscure biases or contaminants, and effectively leverage various models (such as images and sequencing data) to update its world knowledge. Currently, laboratory equipment often lacks the dexterity required to perform many intricate experimental operations, which has become a major bottleneck in automating scientific research. Samuel Rodriguez, a researcher, even likened this challenge to "the door to the stars," emphasizing that substantial investment (he suggested at least $10 million) is needed to push forward with AI-driven scientific automation. To address these issues, FutureHouse has established a robust infrastructure for evaluating AI systems' precision and reliability, expanding human labor assessment capabilities. They developed a platform called LAB-Bench, which can handle a wide range of scientific tasks. Rodriguez noted that large language models generally do not perform well under zero-shot conditions, requiring an environment that can simulate the core processes of scientific research and provide high-quality reward signals to train AI effectively. This may also necessitate breaking down traditional barriers of expertise to accommodate highly complex and open environments. The Robin system, developed by FutureHouse, highlights the ongoing nature of this spatial development. For instance, while Robin can generate experimental designs, future iterations will aim to provide more detailed, precise, and executable experimental plans, thereby significantly reducing the amount of human intervention required in the laboratory. In data analysis, the current system relies heavily on expert-designed processes to ensure reliable and high-quality results. However, FutureHouse's vision is to endow Finch, their AI's core intelligence, with greater autonomy. Finch should be capable of independently generating data analysis prompts or at least adjusting existing prompts based on different data models, thus achieving a more efficient and autonomous scientific discovery process. Ultimately, although Robin currently uses large models for speculative judgment in hypothesis generation, FutureHouse believes that a key future focus will be on better integrating AI's hypothesis generation and evaluation processes with those of human top-tier scientists. The goal is to achieve a steady and reliable production of high-quality scientific hypotheses, leveraging AI's ability to detect subtle biases that human experts might overlook. References: 1. https://arxiv.org/pdf/2505.13400 2. https://www.linkedin.com/company/futurehouse/ 3. https://techcrunch.com/2025/05/06/futurehouse-previews-an-ai-tool-for-data-driven-biology-discovery/ 4. https://www.theinformation.com/articles/startup-building-ai-scientist?rc=qjiy7u 5. https://www.sam-rodriques.com/post/what-does-it-take-to-build-an-ai-scientist 6. https://x.com/SGRodriques/status/1925024623948902801 Editor’s Note: This summary maintains the key points of the original text while enhancing clarity and readability. It outlines the challenges and advancements in AI-driven scientific automation, particularly in the context of drug discovery for rare diseases.