Educational Technology and Change Journal

By Jim Shimabukuro (assisted by Gemini)
Editor

Introduction: Bryan Walsh, in “We’re running out of good ideas. AI might be how we find new ones” (Vox, 13 Dec. 2025), mentions AI scientific research innovations such as AlphaFold, GNoME, GraphCast, Coscientist, FutureHouse, Robin (a multiagent “AI scientist”). I asked Gemini to expand on them. -js

AlphaFold

AlphaFold is a groundbreaking artificial intelligence system developed by Google DeepMind, in partnership with the European Bioinformatics Institute (EMBL-EBI), that effectively solved the “protein folding problem,” a grand challenge in biology that had puzzled scientists for half a century. The core function of AlphaFold is to accurately predict the three-dimensional (3D) structure of a protein based solely on its one-dimensional amino acid sequence.

It achieves this remarkable feat using deep learning techniques, specifically a deep neural network that was trained on the publicly available Protein Data Bank (PDB) of experimentally determined protein structures. Prior to AlphaFold’s success, determining a protein’s structure required complex and time-consuming laboratory methods like X-ray crystallography or cryo-electron microscopy, often taking years per structure.

The significance of AlphaFold is vast, initiating a revolution across structural biology and pharmaceutical research. Proteins are the essential molecular machines of life, and their function is fundamentally determined by their 3D shape. By providing highly accurate structures within minutes, AlphaFold drastically accelerates the pace of understanding biological systems, including the mechanisms of disease. For instance, knowing the precise structure of a viral protein allows researchers to rationally design drugs that can block its function.

The subsequent release of the AlphaFold Protein Structure Database, containing over 200 million predicted structures for nearly all cataloged proteins, has made this foundational knowledge freely accessible worldwide, transforming drug discovery, enzyme engineering, and basic life science research. The developers, including Demis Hassabis and John Jumper, have been recognized for this work, which has successfully transitioned AI from a purely computational tool into a practical, indispensable engine for scientific discovery at an unprecedented scale.

Source: AlphaFold Protein Structure Database, Google DeepMind and EMBL-EBI, ongoing resource. “AlphaFold DB provides open access to over 200 million protein structure predictions to accelerate scientific research.”

---

GNoME

GNoME, which stands for Graph Networks for Materials Exploration, is a sophisticated deep learning tool created by Google DeepMind’s Materials Discovery team, led by researchers like Ekin Dogus Cubuk. Its purpose is to rapidly and accurately predict the stability of novel inorganic crystal structures, a process that is traditionally the most significant bottleneck in materials science.

GNoME operates using graph neural networks (GNNs), an architectural choice particularly well-suited for modeling the connections and relationships between atoms in a crystal lattice. The system employs an active learning loop, where it generates candidate structures, predicts their stability using computational methods like Density Functional Theory (DFT), and then incorporates those results back into its training data to iteratively improve its performance and accuracy.

The material world relies on the properties of inorganic crystals, which are crucial for technologies ranging from superconductors and solar panels to next-generation batteries and computer chips. Historically, human effort and traditional computational methods had identified about 48,000 stable materials. In a single breakthrough, GNoME generated predictions for 2.2 million new crystals, identifying 380,000 as stable and highly promising candidates for experimental synthesis.

This expansion of the known universe of stable materials by nearly tenfold is why GNoME is considered a transformative force. Its high prediction accuracy, reaching around 80% precision for material stability, provides scientists with a massive, validated dataset to explore. Furthermore, the GNoME predictions have been leveraged by autonomous robotic laboratories, such as Berkeley Lab’s A-Lab, to physically synthesize new materials with minimal human input, establishing a critical “AI-to-lab” pipeline that dramatically accelerates the path from theoretical discovery to practical application.

Source: Millions of new materials discovered with deep learning, Amil Merchant and Ekin Dogus Cubuk, Google DeepMind Blog, November 29, 2023. “GNoME’s discovery of 2.2 million materials would be equivalent to about 800 years’ worth of knowledge and demonstrates an unprecedented scale and level of accuracy in predictions.”

---

GraphCast

GraphCast is a state-of-the-art artificial intelligence model developed by Google DeepMind that is engineered to provide fast and highly accurate global medium-range weather forecasts. Unlike traditional Numerical Weather Prediction (NWP) systems, which rely on solving complex, resource-intensive physics-based equations, GraphCast utilizes a machine learning approach centered on Graph Neural Networks (GNNs). The system was trained on decades of historical weather data, allowing it to learn the complex, underlying cause-and-effect relationships that govern the Earth’s weather evolution, effectively replacing brute-force calculation with data-driven pattern recognition.

The primary function of GraphCast is to predict weather conditions up to 10 days in advance with a level of accuracy that often surpasses the industry gold-standard NWP models. Its key advantage, however, is speed. A ten-day forecast that can take traditional supercomputers over an hour to generate, GraphCast can produce in approximately 45 seconds. This immense acceleration in computational efficiency represents a significant step forward, translating to a time-saving factor that is crucial for forecasting agencies and global industries.

The importance of GraphCast lies in its potential to save lives and mitigate economic damage. Its improved accuracy in predicting high-impact extreme weather events, such as the path of hurricanes, allows authorities to issue earlier and more reliable warnings. By providing faster and more precise weather intelligence, GraphCast supports critical decision-making across aviation, agriculture, energy grid management, and disaster response, proving that AI can serve as a powerful complement to, and sometimes an improvement upon, established scientific methodologies.

Source: GraphCast: AI model for faster and more accurate global weather forecasting, Remi Lam on behalf of the GraphCast team, Google DeepMind Blog, November 14, 2023. “GraphCast takes a significant step forward in AI for weather prediction, offering more accurate and efficient forecasts, and opening paths to support decision-making critical to the needs of our industries and societies.”

---

Coscientist

Coscientist is a term used for AI systems designed to act as advanced partners or collaborators in the scientific process, significantly augmenting human capability. The most recent version, often referred to as the AI co-scientist, was developed by Google and is a multi-agent system built upon the advanced capabilities of the Gemini 2.0 large language model. Its purpose is to automate the most intellectually demanding stages of research: the generation, debate, and iterative refinement of novel research hypotheses and proposals. This system functions as a digital laboratory team, employing specialized agents—such as Generation, Reflection, Ranking, and Evolution agents—that continuously work together in a tournament-style framework to achieve a research goal provided by a human scientist in natural language.

The power of Coscientist is its capacity to synthesize vast amounts of existing scientific literature and data to propose verifiable new knowledge. In biomedical research, this system has demonstrated its importance by identifying potential drug candidates for complex diseases like liver fibrosis, even proposing therapeutic targets that showed anti-fibrotic activity in human organoid models. Moreover, it showcased its core capability by independently deducing a complex, previously unknown mechanism of bacterial gene transfer that had required years of dedicated human experimentation to uncover.

While another system also named Coscientist, developed at Carnegie Mellon University, focuses on using LLMs to translate natural language into code for controlling robotic chemical laboratories, the Google-developed “AI co-scientist” highlights the system’s ability to automate the intellectual, hypothesis-driven core of the scientific method. This shift allows human researchers to delegate the initial stages of deep inquiry, freeing them to focus on the experimental validation and high-level strategy necessary to accelerate the timeline for scientific breakthroughs.

Source: Google’s AI co-scientist just solved a biological mystery that took humans a decade, Eric W. Dolan, Psypost, 1 Nov. 2025. “By generating novel and experimentally verifiable hypotheses, tools like the AI co-scientist have the potential to supercharge human intuition and accelerate the pace of scientific and biomedical breakthroughs.”

---

FutureHouse

FutureHouse is a non-profit AI research lab, co-founded by Sam Rodriques, Andrew White, and Jessica Wu, with the explicit mission to build an “AI Scientist” capable of scaling scientific discovery beyond human limitations. The organization addresses the overwhelming challenge faced by modern researchers who must navigate massive, ever-growing scientific literature to formulate viable hypotheses. FutureHouse’s core contribution is an open-access platform featuring specialized large language model (LLM) agents, such as Crow, Falcon, Owl, and Phoenix, each tailored to distinct stages of the research workflow, initially in biology and chemistry.

The system functions primarily as a powerful cognitive accelerator for literature review and strategic planning. Its agents leverage LLMs and retrieval-augmented generation (RAG) to scour vast corpuses of scientific papers and databases, synthesizing answers and analyses, and critically evaluating the quality and credibility of sources, much like an expert researcher would. For example, a researcher can task the Falcon agent with performing a comprehensive literature review on a complex disease, a task that might take a human weeks, and receive a synthesized report with full citations in minutes.

By creating a collaborative and chainable agent architecture, FutureHouse enables scientists to link these tools together for complex, end-to-end investigations, going from identifying a research gap to proposing an experimental design. The importance of FutureHouse lies in its commitment to democratizing advanced research. By making its platform and agents accessible, the organization is lowering the barrier to entry for complex research, providing academic institutions and smaller labs with tools that allow them to maintain continuous, automated literature surveillance and high-volume analytics, thus fostering innovation globally.

Source: FutureHouse AI Agents: A Guide to Its Research Platform, Adrien Laurent, IntuitionLabs, 10 Dec. 2025. “It is very difficult for scientists to maintain their own agent deployments, so we provide an API… to facilitate researcher workflows.”

---

Robin (a multiagent “AI scientist”)

Robin is a multi-agent AI system developed by FutureHouse researchers and represents one of the most advanced implementations of the “AI scientist” concept. Its development marks a critical shift because it is the first system capable of fully automating the entire intellectual cycle of scientific discovery in a continuous workflow, establishing a true “lab-in-the-loop” framework. Robin integrates literature search agents (like Crow and Falcon from the FutureHouse platform) with specialized data analysis agents (such as Finch), allowing it to generate hypotheses, propose experiments, interpret the resulting experimental data, and then refine its initial hypotheses for the next round of testing, all autonomously.

The multi-agent architecture enables Robin to systematically pursue novel discoveries by decomposing the scientific process into manageable, collaborative sub-tasks. The profound importance of Robin was demonstrated in its application to dry age-related macular degeneration (dAMD), a leading cause of blindness with limited treatment options. Starting with the disease’s pathophysiology, Robin autonomously proposed enhancing retinal pigment epithelium phagocytosis as a therapeutic strategy.

Crucially, the system then identified and validated ripasudil, a clinically used drug previously unassociated with dAMD, as a promising therapeutic candidate. By generating, testing, and mechanistically validating this novel drug repurposing opportunity without human intervention in the core intellectual steps, Robin proves that AI systems can significantly reduce the time and resources required for therapeutic development. It effectively streamlines a process that traditionally involves years of laborious, manual intellectual and experimental work, setting a new standard for accelerated scientific inquiry.

Source1: Robin: A multi-agent system for automating scientific discovery, Ali Essam Ghareeb et al., alphaXiv Preprint, May 19, 2025. “As the first AI system to autonomously discover and validate a novel therapeutic candidate within an iterative lab-in-the-loop framework, Robin establishes a new paradigm for AI-driven scientific discovery” (source2).

[End]

Filed under: Uncategorized |