#10 — Orion automates biomedical image, data analysis

What happens when an AI can automate the analysis of biomedical images and then write its own research reports? That’s a capability now being demonstrated, and it's our lead story on Research Agent. Here is a look at the system. [LEAD] Researchers presented Orion, a computer-using AI agent for biomedical image analysis Orion is an AI agent designed to automate the computational layer of laboratory work in biomedical image analysis and interpretation. It aims to address the limitations of labor-intensive computational workflows in laboratory discovery by automating complex processes. This agent represents a step towards advanced lab automation. SOURCE MATERIAL: (Orion: Towards Lab Automation with Computer-Using Agents, 2026-06-16): Laboratory discovery increasingly depends on computational workflows that connect experimental data to analysis, interpretation and follow-up hypotheses. Yet these workflows remain constrained by labor-intensive use of specialized software, visual inspection through graphical user interfaces, and integration of knowledge across multiple sources. Here, we present Orion, a computer-using AI agent for biomedical image analysis and interpretation that moves towards lab automation by automating this computational layer of laboratory work. Orion combines large language models with terminal execution, GUI control and adaptive multi-step reasoning in a shared computing environment. It can inspect vi [LEAD] Orion combines large language models with terminal execution, GUI control, and adaptive multi-step reasoning The Orion AI agent integrates large language models with various computational functionalities within a shared computing environment. These functionalities include terminal execution, control over graphical user interfaces, and adaptive multi-step reasoning. This architecture allows Orion to inspect visual data, operate standard scientific software, and mine web resources. SOURCE MATERIAL: (Orion: Towards Lab Automation with Computer-Using Agents, 2026-06-16): Laboratory discovery increasingly depends on computational workflows that connect experimental data to analysis, interpretation and follow-up hypotheses. Yet these workflows remain constrained by labor-intensive use of specialized software, visual inspection through graphical user interfaces, and integration of knowledge across multiple sources. Here, we present Orion, a computer-using AI agent for biomedical image analysis and interpretation that moves towards lab automation by automating this computational layer of laboratory work. Orion combines large language models with terminal execution, GUI control and adaptive multi-step reasoning in a shared computing environment. It can inspect vi [LEAD] Orion achieved over 90% accuracy on biomedical database and literature retrieval tasks In benchmark tests, the Orion AI agent demonstrated high accuracy in specific tasks related to information retrieval. It achieved over 90% accuracy when performing biomedical database and literature retrieval. Orion also successfully learned to utilize popular tools like CellProfiler and QuPath for quantitative analysis of cellular and tissue images. SOURCE MATERIAL: (Orion: Towards Lab Automation with Computer-Using Agents, 2026-06-16): Laboratory discovery increasingly depends on computational workflows that connect experimental data to analysis, interpretation and follow-up hypotheses. Yet these workflows remain constrained by labor-intensive use of specialized software, visual inspection through graphical user interfaces, and integration of knowledge across multiple sources. Here, we present Orion, a computer-using AI agent for biomedical image analysis and interpretation that moves towards lab automation by automating this computational layer of laboratory work. Orion combines large language models with terminal execution, GUI control and adaptive multi-step reasoning in a shared computing environment. It can inspect vi [LEAD] Orion generated 52 research reports and 22 plausible mechanistic hypotheses in 100 hours of autonomous exploration During 100 hours of autonomous exploration of a large-scale perturbation imaging dataset, Orion produced 52 research reports. Human scientists subsequently reviewed these reports and identified 22 plausible mechanistic hypotheses. This outcome shows Orion's capacity to facilitate autonomous discovery within experimental imaging data, providing a scalable route from data to hypotheses. SOURCE MATERIAL: (Orion: Towards Lab Automation with Computer-Using Agents, 2026-06-16): Laboratory discovery increasingly depends on computational workflows that connect experimental data to analysis, interpretation and follow-up hypotheses. Yet these workflows remain constrained by labor-intensive use of specialized software, visual inspection through graphical user interfaces, and integration of knowledge across multiple sources. Here, we present Orion, a computer-using AI agent for biomedical image analysis and interpretation that moves towards lab automation by automating this computational layer of laboratory work. Orion combines large language models with terminal execution, GUI control and adaptive multi-step reasoning in a shared computing environment. It can inspect vi [LEAD] Prometheus, PhysicsX, and Mistral collectively raised $15.8 billion. The companies Prometheus, PhysicsX, and Mistral collectively secured $15.8 billion in funding. This significant capital injection coincided with the week these organizations adopted similar 'Physics AI'-related naming conventions. SOURCE MATERIAL: (Three companies named "Physics AI" the same week and raised ..., 2026-06-15): Three companies named "Physics AI" the same week and raised $15.8 billion between them. Prometheus, PhysicsX, and Mistral each independently reached for the same new term in the same week -- and collectively raised $15.8B. Plus: the AI IPO wave nobody noticed. [LEAD] Prometheus raised $12 billion Prometheus, a physical AI startup co-founded by Jeff Bezos and Vik Bajaj, secured $12 billion in a funding round. This investment valued the company at $41 billion. The funding highlights continued investor interest in physical AI companies. SOURCE MATERIAL: (The Week's 5 Biggest Funding Deals – AI, robotics, space ..., 2026-06-16): The week was again one of the strongest weeks for late-stage startup funding, led by mega-rounds in physical AI, robotics, satellite intelligence, data security, and enterprise IT automation. The biggest announcement came from Prometheus, the physical AI startup co-founded by Jeff Bezos and Vik Bajaj, which raised $12 billion at a $41 billion valuation. The weeks highlight continued investor appetite for companies building mission-critical AI infrastructure, automation platforms, sovereign intelligence systems, and enterprise security solutions. [LEAD] OpenAI disproved the Erdos unit distance conjecture autonomously in 2026 The paper cites a hypothetical future event where OpenAI autonomously disproves the Erdos unit distance conjecture. This projected event is used as an illustration of advanced AI qualitative reasoning. It demonstrates AI's capacity to identify inadequate frameworks and derive solutions from new conceptual spaces. SOURCE MATERIAL: (A Three-Layer Framework for AI in Scientific Discovery, 2026-06-11): Current discussions of AI in scientific discovery are often dominated by two visible capabilities: search over existing knowledge and execution through optimization, simulation, and automation. Both are important, but neither fully captures the central act of discovery: the formation and evolution of models. This paper proposes a three-layer view of AI in discovery. Layer 1 is search and retrieval by large language models. Layer 2, as the main innovation of this paper, is model formation through qualitative reasoning: the capacity to recognize when a current framework is structurally inadequate and to understand the problem within a broader representational space, not through trial and error [NOTABLE] Florence Healthcare announced Model Context Protocol access for its application suite. Florence Healthcare made Model Context Protocol (MCP) access available for its application suite. This enables AI agents to interact with its eleven specialized site workflow tools across a large network of research sites and protocols. The integration is designed to automate administrative tasks and streamline clinical research operations. SOURCE MATERIAL: (Florence Healthcare Unlocks Agentic AI Across Global Site Network at DIA 2026, 2026-06-16): Florence Healthcare announced the release of Model Context Protocol (MCP) access for its application suite at the 2026 DIA Global Annual Meeting, enabling AI agents to interact with its eleven specialized site workflow tools across 65,000 research sites and 30,000 protocols. This expansion shifts clinical trial operations to an agent-ready infrastructure, allowing sponsor and site AI agents to securely query, analyze, and act upon "site reality" in real-time. The integration aims to automate administrative tasks, predict enrollment risks, and streamline clinical research, ultimately returning thousands of hours to coordinators for patient care. [NOTABLE] Databricks released Omnigent as an open-source meta-harness Databricks introduced Omnigent, a meta-harness designed to integrate and manage various AI agents. This system aims to address the limitations of individual agent harnesses by providing composition capabilities, advanced policy control, and live collaboration features. Omnigent is being open-sourced under the Apache 2.0 license to foster a new layer for working with AI agents. SOURCE MATERIAL: (Introducing Omnigent: A Meta-Harness to Combine, Control and ..., 2026-06-13): Databricks developed Omnigent, a meta-harness designed to integrate and manage various AI agents like Claude Code, Codex, and Pi. This system aims to overcome the limitations of individual agent harnesses by enabling composition, advanced policy control, and live collaboration among teammates. Omnigent is being open-sourced under Apache 2.0 to foster a new layer for working with agents. [NOTABLE] Researchers developed a method for interpreting AI models in materials discovery Researchers from Japan created a new method to interpret artificial intelligence models used in materials discovery. This approach addresses the 'black box' nature of these models by extracting key features from them. The method groups materials with similar characteristics based on atomic structural data and optical absorption spectra, providing insights for efficient materials design. SOURCE MATERIAL: (AI in Materials: Revealing Model Prediction Insights | Mirage News, 2026-06-17): Researchers from Japan have developed a method to interpret artificial intelligence (AI) models used in materials discovery, addressing the "black box" nature of these models. This approach extracts key features from AI models trained on atomic structural data and optical absorption spectra, then groups materials with similar characteristics. The method, which can be extended to other material properties, provides physical and chemical insights for more efficient materials design. [NOTABLE] Singapore's AI-for-Science program received $120 million in funding. Singapore's AI-for-Science (AI4S) program, which includes the Materials Data Foundry, secured $120 million in funding. The program focuses on accelerating scientific discoveries through the application of artificial intelligence across various fields. SOURCE MATERIAL: (Singapore using AI to hasten the discovery of recipes for next-gen semiconductors, clean hydrogen, 2026-06-16): Singapore has established a new materials laboratory between the National University of Singapore and the University of Toronto, aiming to use artificial intelligence to accelerate the discovery and manufacturing of next-generation semiconductors and clean hydrogen. This $10 million Materials Data Foundry is part of Singapore's AI-for-Science (AI4S) programme, backed by $120 million in funding to speed up scientific discoveries. The initiative seeks to bridge the gap between lab discoveries and real-world manufacturing by rapidly running thousands of experiments. [NOTABLE] China introduced the AI-powered drug discovery platform, GalaxyVS. China launched GalaxyVS, an artificial intelligence-powered platform designed for drug discovery. This platform significantly reduces the initial phase of identifying candidate molecules from years to seconds. This introduction supports China's broader objective of achieving a pharmaceutical revolution. SOURCE MATERIAL: (China aims to win the race for future drugs manufactured in seconds | News, 2026-06-16): China is striving to achieve a pharmaceutical revolution by combining supercomputing, artificial intelligence, industrial control, and state support. The country has introduced GalaxyVS, an AI-powered drug discovery platform that can reduce the initial phase of finding candidate molecules from years to seconds. This advancement positions China as a leading global center for bio-pharmaceutical innovation, moving beyond its previous role as a manufacturer of generic drugs. [NOTABLE] Professor Paul Atzberger's group developed methods to integrate geometric understanding into AI systems The group led by Professor Paul Atzberger at UC Santa Barbara is developing new methods to incorporate geometric common-sense knowledge into AI systems. Their approach uses mathematics and differential geometry to enable AI to understand data as interconnected forms rather than discrete points. This work, delivered as an open-source Python package, aims to create more robust AI applications for various fields. SOURCE MATERIAL: (Giving AI geometric awareness allows it to better understand the ..., 2026-06-16): Professor Paul Atzberger's group at UC Santa Barbara is developing new methods to integrate common-sense knowledge, specifically geometric understanding, into AI systems to improve their comprehension of the world. By leveraging concepts from mathematics and differential geometry, their AI algorithms perceive data beyond mere collections of points, recognizing interconnected parts with edges, curved surfaces, and other forms. This approach, which they’ve implemented in an open-source Python package, results in more robust AI applications across various industries and research fields. [NOTABLE] Carnegie Mellon University's NSF AI Institute for Societal Decision Making and Meta's AI for Good program are partnering to develop dynamic situation reports for natural disaster management. Carnegie Mellon University's NSF AI Institute for Societal Decision Making (NSF AI-SDM) and Meta's AI for Good program have initiated a collaboration. Their objective is to create dynamic situation reports to aid first responders in managing natural disasters. The partnership will integrate aggregated mobility and connectivity data from Meta, alongside satellite imagery and open-source AI models, to produce actionable visualizations and insights. SOURCE MATERIAL: (Carnegie Mellon and Meta Partner To Develop AI Tools for Emergency Response, 2026-06-16): Carnegie Mellon University's NSF AI Institute for Societal Decision Making (NSF AI-SDM) and Meta's AI for Good program are partnering to develop dynamic situation reports to help first responders manage natural disasters. This collaboration will use aggregated mobility and connectivity data from Meta, combined with other sources like satellite imagery and open-source AI models, to create clear visualizations and actionable insights for emergency managers. The tools will be evaluated during natural disasters in 2026 and eventually disseminated via the Humanitarian Data Exchange and the NSF AI-SDM website. [NOTABLE] Fundraise Insider's Q1 2026 report revealed that AI startups secured funding faster and disproportionately compared to other sectors. Fundraise Insider's Q1 2026 report analyzed startup funding trends across various market segments. The report indicated that AI startups achieved Series A and B funding rounds more rapidly than companies in other sectors. While AI companies constituted 36.4% of funded entities, they received 57% of all disclosed capital, highlighting a significant concentration of investment. The report also noted debt financing as a common method for these companies to raise capital. SOURCE MATERIAL: (Q1 2026 Startup Funding Report: AI Takes 57% of All Capital, 2026-06-16): Fundraise Insider's Q1 2026 report reveals AI startups are achieving Series A and B funding faster than other market segments, with debt financing also emerging as a common method for companies to raise capital. The report, covering 1,729 companies, highlights that AI companies, while making up 36.4% of funded companies, absorbed 57% of all disclosed capital. This indicates a funding market split into two economies, with a "capital event of historic scale" in AI. [NOTABLE] U.S. companies secured 88% of global AI-related startup funding. U.S. companies acquired nearly 88% of global startup funding specifically allocated to AI-related projects. The majority of this funding was directed towards OpenAI and Anthropic. SOURCE MATERIAL: (The AI Startup Funding Boom Is Not A Global Phenomenon, 2026-06-15): The flood of AI-focused funding has pushed global startup investment to record levels this year, yet most countries have not seen these gains. U.S. companies have secured nearly 80% of global seed- through growth-stage financing in 2026, a significant increase from previous years. The U.S. share of AI-related startup funding is even higher, at nearly 88%, with most of that going to OpenAI and Anthropic. [NOTABLE] A paper proposed a three-layer view of AI in scientific discovery The paper introduces a three-layer model describing AI's role in scientific discovery. This model categorizes AI functions into search and retrieval, qualitative reasoning for model formation, and execution and optimization. The authors emphasize that model formation (Layer 2) is critical yet currently underdeveloped. SOURCE MATERIAL: (A Three-Layer Framework for AI in Scientific Discovery, 2026-06-11): Current discussions of AI in scientific discovery are often dominated by two visible capabilities: search over existing knowledge and execution through optimization, simulation, and automation. Both are important, but neither fully captures the central act of discovery: the formation and evolution of models. This paper proposes a three-layer view of AI in discovery. Layer 1 is search and retrieval by large language models. Layer 2, as the main innovation of this paper, is model formation through qualitative reasoning: the capacity to recognize when a current framework is structurally inadequate and to understand the problem within a broader representational space, not through trial and error [NOTABLE] A machine-learning approach forecasts scientific breakthroughs. Researchers introduced an explainable machine-learning model designed to predict scientific breakthroughs. This approach models the evolution of OpenAlex concept networks to forecast the emergence and intensification of links between research concepts. The two-stage LightGBM model uses 59 semantic and topological features to predict both concept pair formation and their future weight. SOURCE MATERIAL: (Explainable Forecasting of Scientific Breakthroughs from Concept Network Dynamics, 2026-06-02): We introduce an explainable machine-learning approach that forecasts the structural precursors of scientific breakthroughs -- the emergence and intensification of links between research concepts -- by modelling how OpenAlex concept networks evolve over time. Using 59 semantic and topological features, a two-stage LightGBM model jointly predicts the formation and the future weight of concept pairs, adding a regression stage that quantifies expected intensity to prior link-existence forecasts. Relative to the state of the art, the approach improves accuracy and explainability at once: comparative validation across four technology and biomedical domains yields ROC-AUC in [0.954, 0.967] at all h [NOTABLE (FOLLOW-UP)] A machine-learning approach improves breakthrough forecasting accuracy and explainability. The new machine-learning approach demonstrates improved accuracy and explainability compared to prior models. Validation across technology and biomedical domains showed ROC-AUC scores between 0.954 and 0.967, surpassing previous models' 0.90. Its classification performance achieved an AUC of approximately 0.95, with stable regression (RMSLE 0.45-0.6 over 1-5 years), relying on structural, auditable features. SOURCE MATERIAL: (Explainable Forecasting of Scientific Breakthroughs from Concept Network Dynamics, 2026-06-02): We introduce an explainable machine-learning approach that forecasts the structural precursors of scientific breakthroughs -- the emergence and intensification of links between research concepts -- by modelling how OpenAlex concept networks evolve over time. Using 59 semantic and topological features, a two-stage LightGBM model jointly predicts the formation and the future weight of concept pairs, adding a regression stage that quantifies expected intensity to prior link-existence forecasts. Relative to the state of the art, the approach improves accuracy and explainability at once: comparative validation across four technology and biomedical domains yields ROC-AUC in [0.954, 0.967] at all h [NOTABLE] A three-layer decision architecture turns forecasts into research strategy. A three-layer decision architecture was outlined to translate the scientific breakthrough forecasts into evidence-based research strategy and policy. This architecture comprises detection, expert translation, and institutional integration layers. The system is designed to be anchored in open data and explainable features for informed decision-making. SOURCE MATERIAL: (Explainable Forecasting of Scientific Breakthroughs from Concept Network Dynamics, 2026-06-02): We introduce an explainable machine-learning approach that forecasts the structural precursors of scientific breakthroughs -- the emergence and intensification of links between research concepts -- by modelling how OpenAlex concept networks evolve over time. Using 59 semantic and topological features, a two-stage LightGBM model jointly predicts the formation and the future weight of concept pairs, adding a regression stage that quantifies expected intensity to prior link-existence forecasts. Relative to the state of the art, the approach improves accuracy and explainability at once: comparative validation across four technology and biomedical domains yields ROC-AUC in [0.954, 0.967] at all h That is all for this episode. We will be back with more developments in AI for scientific research. Until next time, on Research Agent.