Accelerating scientific discovery with AI

Scientific progress, the engine of human advancement, appears to be slowing. Over the past five decades, researchers have noted a troubling trend: it now demands more time, greater funding, and larger teams to achieve discoveries that once came faster and cheaper. This decline in scientific productivity is largely attributed to the increasing complexity and specialization of research, compelling scientists to spend excessive time on literature reviews, intricate experiment designs, and data analysis.

Enter FutureHouse, a philanthropically funded research laboratory spearheading a transformative solution. Founded by MIT alums Sam Rodriques PhD ’19 and Andrew White, FutureHouse is deploying an innovative AI platform designed to automate many of the crucial steps in scientific research. Their vision is bold: to arm every scientist with sophisticated AI agents, effectively dismantling the biggest bottlenecks in science and accelerating solutions to humanity’s most urgent challenges.

“Natural language is the real language of science,” explains Rodriques, emphasizing the distinct approach of FutureHouse. While other models focus on the ‘language’ of DNA or proteins, Rodriques asserts that true discoveries, hypotheses, and reasoning are best articulated and advanced through natural language. This philosophy underpins the development of their platform.

The genesis of FutureHouse traces back to Rodriques’ PhD research at MIT with Professor Ed Boyden. He recognized that even with an abundance of information about complex systems like the brain, the sheer volume of literature made comprehensive understanding and theory formation nearly impossible. This insight led him to explore new models for innovation, culminating in his PhD thesis chapter in 2019 advocating for novel large-scale research collaborations.

The advent of generative AI models like Chat-GPT 3.5 in November 2022 and Chat-GPT 4 proved to be a pivotal moment. Andrew White, a computational chemist at the University of Rochester, had already developed the first large language agent for science using early access to Chat-GPT 4. Recognizing their shared vision, Rodriques and White joined forces to establish FutureHouse.

Initially, FutureHouse focused on creating distinct AI tools. Their early successes included PaperQA, released in September 2024 (now rebranded as Crow), which swiftly became a leading AI agent for retrieving and summarizing scientific literature. Simultaneously, they launched Has Anyone (now Owl), a tool enabling scientists to quickly ascertain if specific experiments or hypotheses had already been explored. These initial tools were direct responses to common bottlenecks faced by scientists.

The official launch of the FutureHouse platform on May 1 of this year introduced a suite of powerful, rebranded, and new AI agents. Falcon, an advanced agent, can compile and review significantly more sources than Crow. Phoenix offers specialized tools to assist researchers in planning complex chemistry experiments. Finch is specifically designed to automate data-driven discovery within biology. In June, FutureHouse further demonstrated its commitment to open science by releasing ether0, a 24B open-weights reasoning model tailored for chemistry.

A notable demonstration on May 20 showcased the platform’s integrated multi-agent workflow. This workflow successfully automated key steps of the scientific process to identify a new therapeutic candidate for dry age-related macular degeneration (dAMD), a leading cause of irreversible blindness globally. Rodriques emphasizes the synergistic nature of these tools: “Soon, the literature search agents will be integrated with the data analysis agent, the hypothesis generation agent, an experiment planning agent, and they will all be engineered to work together seamlessly.”

Today, the FutureHouse agents are accessible to anyone via platform.futurehouse.org, sparking considerable excitement within the scientific community. Early success stories are already emerging: one FutureHouse scientist leveraged the agents to identify a gene associated with polycystic ovary syndrome and propose a novel treatment hypothesis. A researcher at the Lawrence Berkeley National Laboratory utilized Crow to develop an AI assistant for Alzheimer’s disease research on the PubMed database. Additionally, scientists at another research institution reported that FutureHouse’s agents outperformed general AI tools in conducting systematic reviews of Parkinson’s disease-relevant genes.

Rodriques advises scientists to view these agents less as a sophisticated search engine and more as an intelligent assistant scientist. “People who are looking for really faithful literature reviews tend to get more out of our agents,” he notes, contrasting them with general large language models that might lean towards speculation.

Looking ahead, FutureHouse aims to integrate raw data from research papers to enhance reproducibility testing and conclusion verification. In the long run, the goal is to imbue their agents with tacit knowledge for more sophisticated analyses and grant them the ability to utilize computational tools for exploring hypotheses. This progressive integration of specialized scientific tools with advanced AI models will be critical in sustaining and accelerating the march of scientific progress.