Numerous scholars have adopted a comprehensive perspective on scientific advancement over the past five decades and arrived at the same concerning assessment: Scientific output is diminishing. It now requires more time, increased funding, and larger teams to achieve discoveries that previously occurred more swiftly and economically. While numerous explanations have been proposed for this deceleration, one is that as research grows increasingly intricate and specialized, scientists must dedicate more time to reviewing publications, designing intricate experiments, and interpreting data.
Currently, the philanthropically supported research facility FutureHouse is striving to expedite scientific inquiry through an AI platform intended to automate several crucial steps on the journey to scientific advancement. This platform consists of a collection of AI agents tailored for tasks such as information retrieval, data synthesis, chemical synthesis design, and data interpretation.
FutureHouse’s founders, Sam Rodriques PhD ’19 and Andrew White, are convinced that by providing every scientist access to their AI agents, they can overcome the most significant obstacles in science and contribute to addressing some of humanity’s most urgent challenges.
“Natural language is the true language of science,” Rodriques asserts. “Others are developing foundational models for biology, where machine learning models articulate the language of DNA or proteins, which is indeed powerful. However, discoveries are not encoded in DNA or proteins. The sole method we have for representing discoveries, formulating hypotheses, and reasoning is through natural language.”
Identifying Major Issues
During his PhD studies at MIT, Rodriques aimed to grasp the complexities of the brain within Professor Ed Boyden’s lab.
“The entire concept behind FutureHouse originated from an impression I had during my PhD at MIT that even if we possessed all the information necessary to understand how the brain functions, we wouldn’t truly know it because no one has the time to read all the literature,” Rodriques clarifies. “Even if they could read everything, they wouldn’t be able to integrate it into a cohesive theory. That was a fundamental element of the FutureHouse framework.”
Rodriques articulated the necessity for new forms of large research collaborations as the concluding chapter of his PhD dissertation in 2019. After graduation, although he spent some time leading a lab at the Francis Crick Institute in London, he found himself drawn towards broader challenges in science that no individual lab could address.
“I was intrigued by how to automate or scale scientific efforts and what kinds of new organizational frameworks or technologies would enable higher scientific productivity,” Rodriques states.
Upon the launch of Chat-GPT 3.5 in November 2022, Rodriques recognized a pathway towards advanced models capable of generating scientific insights independently. Around that time, he also met Andrew White, a computational chemist at the University of Rochester who had received early access to Chat-GPT 4. White had developed the first large language agent for scientific purposes, leading the researchers to collaborate and establish FutureHouse.
The founders initially aimed to create distinct AI instruments for tasks such as literature searches, data analysis, and hypothesis formulation. They commenced with data collection, ultimately unveiling PaperQA in September 2024, which Rodriques describes as the leading AI agent for retrieving and summarizing information in scientific literature. Simultaneously, they launched Has Anyone, a tool enabling scientists to ascertain whether specific experiments have been conducted or particular hypotheses explored.
“We were just brainstorming, ‘What types of questions do we, as scientists, frequently ask?’” Rodriques recalls.
When FutureHouse officially introduced its platform on May 1 of this year, it rebranded several of its tools. Paper QA is now known as Crow, and Has Anyone is referred to as Owl. Falcon is an agent capable of compiling and reviewing more sources than Crow, while a new agent named Phoenix can employ specialized tools to assist researchers in planning chemistry experiments. Additionally, Finch is an agent designed to automate data-driven discoveries in biology.
On May 20, the company showcased a multi-agent scientific discovery workflow to automate key elements of the scientific process and pinpoint a new therapeutic candidate for dry age-related macular degeneration (dAMD), a primary cause of irreversible blindness globally. In June, FutureHouse introduced ether0, a 24B open-weight reasoning model for chemistry.
“You really have to conceptualize these agents as components of a larger system,” Rodriques explains. “Soon, the literature search agents will be integrated with the data analysis agent, the hypothesis generation agent, and an experiment planning agent, all designed to function together seamlessly.”
Agents for Everyone
Currently, anyone can access FutureHouse’s agents at platform.futurehouse.org. The launch of the company’s platform sparked enthusiasm within the industry, and reports have begun to emerge about scientists utilizing the agents to expedite their research.
One of FutureHouse’s scientists utilized the agents to identify a gene potentially linked to polycystic ovary syndrome and proposed a new treatment hypothesis for the condition. Another researcher at the Lawrence Berkeley National Laboratory leveraged Crow to create an AI assistant capable of searching the PubMed research database for information relevant to Alzheimer’s disease.
Scientists at another research institution employed the agents to conduct systematic reviews of genes pertinent to Parkinson’s disease, discovering that FutureHouse’s agents outperformed general agents.
Rodriques notes that scientists who approach the agents as smart assistant scientists rather than simply as tools like Google Scholar benefit most from the platform.
“Individuals seeking speculative insights tend to gain more from deep research with Chat-GPT, while those in search of faithful literature reviews usually get greater value from our agents,” Rodriques elaborates.
Rodriques also envisions that FutureHouse will soon reach a stage where its agents can utilize raw data from research papers to assess the reproducibility of results and validate conclusions.
In the long term, to sustain the momentum of scientific advancement, Rodriques states that FutureHouse is focused on equipping its agents with tacit knowledge to conduct more sophisticated analyses while also enabling them to utilize computational tools to explore hypotheses.
“There have been numerous advancements concerning foundational models for science and language models for proteins and DNA, so we now need to provide our agents access to those models along with the various tools typically employed in scientific endeavors,” Rodriques states. “Developing the infrastructure for agents to utilize more specialized scientific tools will be crucial.”