Need a research hypothesis? Ask AI

MIT researchers have created a way for humans and artificial intelligence to collaborate to autonomously generate and evaluate promising research hypotheses across fields. In a new paper, they describe how this framework can be used to create evidence-driven hypotheses to address unmet research needs in the field of bioinspired materials.

Published today in advanced materialsThe study was co-authored by Alireza Ghafarollahi, a postdoctoral researcher in the Laboratory of Atomic and Molecular Mechanics (LAMM), and Markus Buehler, the Jerry McAfee Professor of Engineering in MIT’s Department of Civil and Environmental Engineering and the Department of Mechanical Engineering and chair of MIT’s Department of Engineering. . Mu.

The framework, which the researchers call SciAgents, is composed of multiple artificial intelligence agents, each with specific functions and data access rights. They leverage “graph reasoning” methods, in which artificial intelligence models use knowledge graphs to organize and define different sciences. relationships between concepts. Multi-agent approaches mimic the way biological systems organize themselves into groups of basic building blocks. Buhler points out that this “divide and conquer” principle is a prominent paradigm at many levels of biology, from materials to insect colonies to civilizations – all examples showing that total intelligence is far greater than the sum of individual abilities.

“By using multiple AI agents, we try to simulate the process by which a community of scientists makes discoveries,” Buhler said. “At MIT, we do this by having a bunch of people from different backgrounds working together and bumping into each other in coffee shops or in the endless hallways of MIT. But it’s very serendipitous and slow. We The pursuit is to simulate the process of discovery through exploration. Whether artificial intelligence systems can be creative and make discoveries.

Great idea for automation

Recent developments have shown that large language models (LLMs) have demonstrated impressive capabilities in answering questions, summarizing information, and performing simple tasks. But they are quite limited in their ability to generate new ideas from scratch. MIT researchers wanted to design a system that would enable artificial intelligence models to perform more complex multi-step processes, not just recalling information learned during training but also infer and create new knowledge.

The basis of their approach is an ontological knowledge graph, which organizes and establishes connections between different scientific concepts. To create the graphs, the researchers fed a set of scientific papers into a generative artificial intelligence model. In previous work, Buehler used an area of ​​mathematics called category theory to help AI models abstract scientific concepts into graphs, rooted in defining relationships between components, in a way that can be determined by a process called graph reasoning. Other models are analyzed. This focuses AI models on developing a more principled way to understand concepts; it also enables them to better generalize across domains.

“This is important as we create science-focused AI models because scientific theories are often rooted in generalizable principles, not just knowledge recall,” Buhler said. “By focusing AI models on ‘thinking’ in this way, we can go beyond traditional approaches and explore more creative uses of AI.”

In the latest paper, the researchers used about 1,000 scientific studies on biological materials, but Buehler said more or fewer research papers in any field could be used to generate the knowledge graph.

After building the diagram, the researchers developed an artificial intelligence system for scientific discovery, with multiple models specialized to play specific roles in the system. Most components are built on OpenAI’s ChatGPT-4 family of models and use a technique called contextual learning, in which hints provide contextual information about the model’s role in the system while allowing it to learn from the data provided. study.

The agents in the framework interact with each other to jointly solve a complex problem that they cannot do alone. Their first task is to develop a research hypothesis. LLM interaction starts after defining a subgraph from the knowledge graph, which can happen randomly or by manually entering a pair of keywords discussed in the paper.

In this framework, a language model that the researchers named “ontology” is tasked with defining the scientific terms in the paper and examining the connections between them, enriching the knowledge graph. A model named “Scientist 1” then develops a research plan based on factors such as its ability to discover unexpected properties and novelty. The proposal includes a discussion of potential findings, implications of the study, and speculation on underlying mechanisms. The “Scientist 2” model expands on this idea, proposes specific experimental and simulation methods, and makes other improvements. Finally, the “Critics” model highlights its strengths and weaknesses and suggests further improvements.

“It’s about building a team of experts who think differently,” Buhler said. “They have to think in different ways and have different abilities. The Critic agent is deliberately programmed to criticize other people, so you don’t get everyone to agree and say this is a good idea. You have an agent saying, ‘Here’s A weakness that can you explain better? This makes the output very different from the single model.

Other agents in the system are able to scour the existing literature, which provides the system with a way to not only assess feasibility but also create and evaluate the novelty of each idea.

Make the system more powerful

To validate their approach, Buehler and Ghafarollahi built a knowledge graph based on the words “silk” and “energy-intensive.” Using this framework, the Scientist 1 model proposes combining silk with dandelion-based pigments to create biomaterials with enhanced optical and mechanical properties. The model predicts that the material will be much stronger than traditional silk materials and require less energy to process.

Scientist 2 then made suggestions, such as using specific molecular dynamics simulation tools to explore how the proposed material interacts, adding that a good application for the material would be biomimetic adhesives. The Critic model then highlighted several advantages of the proposed material and areas for improvement, such as its scalability, long-term stability, and the environmental impact of solvent use. To address these issues, reviewers recommend conducting process validation pilot studies and rigorous analysis of material durability.

The researchers also conducted additional experiments using randomly selected keywords, generating a variety of original hypotheses, including more efficient biomimetic microfluidic chips, enhanced mechanical properties of collagen scaffolds, and interactions between graphene and amyloid fibrils. Interact to create bioelectronic devices.

The system is able to come up with these new, rigorous ideas based on the paths of the knowledge graph,” Ghafarollahi said. “In terms of novelty and applicability, these materials appear to be robust and novel. In future work, we will generate thousands or tens of thousands of new research ideas, which we can then sort through to try to better understand How these materials are produced and how they can be further improved.

Going forward, the researchers hope to incorporate new tools for retrieving information and running simulations into their framework. They can also easily replace basic models in the framework with more advanced models, allowing the system to adapt to the latest innovations in artificial intelligence.

“Because of the way these agents interact, improvements in a model, even slightly, can have a huge impact on the overall behavior and output of the system,” Buhler said.

Since releasing a preprint containing open-source details of their approach, the researchers have been contacted by hundreds of people interested in using the framework in different fields of science and even in areas such as finance and cybersecurity.

“There are a lot of things you can do without going to the lab,” Buehler said. “Basically, you want to go to the lab at the end of the process. Labs are expensive and take a long time, so you need a system that can drill down to your best ideas, formulate your best hypotheses, and accurately predict emergencies. We The vision is to make it easy to use so you can use the app to introduce other ideas or drag in a dataset to really challenge the model to make new discoveries.