This AI lab wants to automate scientific discovery

FutureHouse, a research lab co-founded by Sam Rodriques PhD ’19 and Andrew White, is developing an AI platform designed to automate many of the most critical steps in the scientific process. The goal is to address a well-documented problem: scientific productivity is declining.

Automating science to reverse declining productivity

Over the last few decades, researchers have observed that scientific discovery is becoming slower and more resource-intensive. Larger teams and higher budgets are now required to produce results that previously came with less effort. One explanation is that science has become too complex for individuals to manage efficiently, particularly given the need to analyze data, review literature, and design increasingly intricate experiments.

FutureHouse aims to solve this challenge by building a set of AI agents that perform specific scientific tasks. These include information retrieval, literature synthesis, experimental design, chemical synthesis planning, and data analysis. The lab’s hypothesis is that if these core steps can be automated or accelerated, the entire research pipeline will become more efficient.

Natural language as the foundation of discovery

Co-founder Sam Rodriques emphasizes that science ultimately relies on natural language. While other research groups focus on building machine learning models that understand DNA or protein sequences, FutureHouse takes a different approach. It views natural language as the universal interface for hypothesis generation, communication, and reasoning.

According to Rodriques, a major driver of the lab’s mission came from his time as a PhD student at MIT. He realized that even if the scientific community had already discovered how the brain works, the knowledge would remain hidden due to the sheer volume of unread literature. Scientists, he concluded, lack the time and tools to synthesize the vast research corpus into actionable insights.

Counterpoint data shows the global smartwatch market is now shrinking

Forming the idea and building early tools

After completing his PhD, Rodriques worked at the Francis Crick Institute but continued exploring structural challenges in science. He was particularly interested in how to scale scientific productivity with new organizational models and emerging technologies. In 2022, with the launch of ChatGPT 3.5 and access to GPT-4, he saw the potential for language models to support scientific reasoning.

At the same time, computational chemist Andrew White had developed a large language agent for scientific applications. Their collaboration resulted in the formation of FutureHouse. Their first tools included PaperQA, an AI assistant for scientific literature, and Has Anyone, which checks whether specific experiments or hypotheses have already been explored. These tools were released in late 2024.

Platform launch and tool integration

On May 1, 2025, FutureHouse officially launched its platform and rebranded several tools:

• Crow (formerly PaperQA) for literature summarization
• Owl (formerly Has Anyone) for hypothesis tracking
• Falcon for reviewing and comparing sources
• Phoenix for planning chemistry experiments
• Finch for automating data-driven biological research

The company also released ether0, a 24-billion parameter open-weights reasoning model for chemistry, in June 2025. FutureHouse demonstrated a multi-agent pipeline that identified a new therapeutic candidate for dry age-related macular degeneration, one of the leading causes of irreversible blindness.

Linking agents into a cohesive system

Rodriques stresses that the agents are designed to work as a coordinated system rather than standalone tools. Literature search agents will eventually integrate with data analysis, hypothesis generation, and experimental planning agents. The vision is to create a seamless scientific assistant that can reason across multiple domains.

Real-world usage and results

Scientists at various institutions have started using FutureHouse’s platform. One researcher used the tools to identify a gene potentially linked to polycystic ovary syndrome and to form a new treatment hypothesis. At Lawrence Berkeley National Laboratory, another scientist developed an assistant for PubMed searches related to Alzheimer’s disease.

In comparative studies, institutions reported that FutureHouse agents outperformed general-purpose AI tools in systematic reviews of genes associated with Parkinson’s disease. The tools were particularly effective when used as collaborators rather than search engines.

Rodriques believes the agents will soon be able to analyze raw data from scientific papers to test result reproducibility. In the long term, FutureHouse plans to embed agents with tacit knowledge and allow them to interact with other computational models, including those used in biology and physics.

The goal is to create agents that can use domain-specific tools, reason across disciplines, and accelerate discovery without requiring users to manage every technical detail. Rodriques frames the next phase of FutureHouse as an infrastructure effort—developing bridges between AI agents and the specialized tools already used in research labs.

A new model for accelerating science

With the declining rate of scientific breakthroughs and growing complexity in research workflows, FutureHouse offers a model for reversing the trend. By designing language-first, task-specific agents that integrate into multi-step research processes, the lab is working toward a system where scientific discovery becomes faster, more reproducible, and more accessible.

FutureHouse’s platform is now publicly available at platform.futurehouse.org. Scientists are encouraged to explore the tools and incorporate them into their own research workflows.

Featured image credit