Things & Thinks-Issue LXIII
📚Research Digest
Multi-Agent Frameworks I- Google's AI Co-Scientist
What the Publication Is About
This paper, by researchers at Google, introduces an AI co-scientist—a multi-agent system built on Gemini 2.0—that generates and refines novel scientific hypotheses, particularly in biomedicine. Designed as a collaborative assistant rather than an autonomous researcher, the system engages in iterative reasoning using a framework inspired by the scientific method. Specialized agents simulate literature reviews, debates, ranking, and idea evolution, all orchestrated under a central supervisor agent, creating a self-improving loop. The authors validate the system through real-world applications including drug repurposing for acute myeloid leukemia, discovery of epigenetic targets for liver fibrosis, and uncovering mechanisms of bacterial gene transfer linked to antimicrobial resistance.
Unlike prior LLM-based tools that focus mainly on summarization or static reasoning, the AI co-scientist builds a full feedback-driven, agentic research cycle. Its architecture supports flexibility, long-term reasoning, and expert-in-the-loop interactions, enabling a scalable framework for hypothesis generation grounded in scientific literature and empirical data.
What It Means for the Future
I see this as a signal that AI is no longer just an assistant for tasks like literature reviews or data analysis—it’s becoming a creative collaborator. The AI co-scientist isn’t replacing researchers; it’s expanding our capacity to ask better questions, iterate faster, and explore more paths than any one expert could manage alone. With its ability to explain reasoning, synthesize across disciplines, and integrate expert feedback, it’s shifting how we think about the role of AI in knowledge generation.
Multi-Agent Frameworks II- Microsoft's Bioagents
What the Publication Is About
Microsoft recently published BioAgents, a lightweight multi-agent system designed to assist with complex bioinformatics tasks by combining small language models with domain-specific fine-tuning and retrieval-augmented generation. Unlike larger models that require heavy compute, BioAgents uses fine-tuned Phi-3 models to deliver conceptual genomics support and workflow guidance locally. The system includes two specialized agents—one trained on bioinformatics tool documentation and the other on workflow protocols—coordinated by a reasoning agent that generates final responses. As compared to Google's co-scientist, this is narrower in scope, focused specifically on bioinformatics workflows, such as genome assembly, RNA-seq, and variant analysis. Its goal is not hypothesis generation but rather to help users design, understand, and troubleshoot computational pipelines. It’s less about creative discovery and more about practical problem-solving and accessibility.
What It Means for the Future
BioAgents and similar experiments represent a meaningful shift in how we make bioinformatics more accessible. It’s not just about getting answers—it’s about understanding the thinking behind them. The system should encourage scientists to learn, refine, and experiment without relying on opaque black-box tools or distant forums.
🖇Digital Healthcare News
#GenAI and #BigTech in #Healthcare
Google announced several updates including health-care updates to Search, expanded its knowledge panels and unveiled a new feature called “What People Suggest". It also launched new Medical Records APIs globally in Health Connect. It released TxGemma, a collection of Gemma-based open models that promise to improve the efficiency of AI-powered drug discovery.
Microsoft launched Microsoft Dragon Copilot, claiming to be the first AI assistant for clinical workflow that brings together the trusted natural language voice dictation capabilities of DMO with the ambient listening capabilities of DAX, fine-tuned generative AI and healthcare-adapted safeguards.
Wolters Kluwer Health, a provider of clinical decision support solutions, collaborated with Microsoft to integrate its UpToDate® platform with the Microsoft Copilot Studio.
Pediatric health system Connecticut Children's expanded its partnership with Xerox by introducing an AI-powered offering aimed at streamlining care delivery while improving patient outcomes.
Innovaccer launched “Agents of Careᵀᴹ”—a suite of pre-trained AI Agents designed to automate repetitive, low-value tasks and handle rising workloads due to staff shortage.
Tech in Digital Health
EMR company Canvas Medical launches Hyperscribe, an open source AI copilot.
Zoom launched the public beta of Zoom Workplace for Clinicians and the select beta of Custom AI Companion for Healthcare, both designed to empower healthcare providers with AI-driven tools and enhance patient care.
Salesforce released Agentforce for Health, a new library of pre-built agent skills and actions designed to tackle time-consuming administrative tasks in healthcare.
Pharma/Device Brief
Eli Lilly partnered up with telehealth providers LifeMD and Teladoc Health to offer its lower cost, single-vial Zepbound (tirzepatide) to patients in the virtual care companies' full-service weight loss management programs.
Similarly Novo Nordisk introduced NovoCare® Pharmacy, offering easy home delivery of low-cost Wegovy for cash-paying patients
Recommended by LinkedIn
The University of Colorado Comprehensive Cancer Center (CU Cancer Center) and Flatiron Health are collaborating to bring Flatiron Clinical Pipe to CU Cancer Center and UCHealth, allowing clinical trial data to move directly from electronic health records (EHRs) to electronic data capture (EDC) systems.
Regulatory Brief
EpiWatch, the Apple Watch-based epilepsy management company spun out from Johns Hopkins, received FDA 510(k) clearance for its continuous seizure monitor platform.
Medtronic received USFDA approval for the BrainSense Adaptive deep brain stimulation (DBS) device for people with Parkinson’s disease
Funding, Deals, Mergers & acquisitions
Tech-enabled drug discovery company Insilico Medicine raised $110M
Clinician AI assistant company Navina raised a $55M
Vori Health, virtual musculoskeletal care provider, raised $53M
Digital pathology software developer Proscia raised $50M.
Inspiren, AI-powered senior living technology company, raised $35M.
Patient monitoring company Sibel Health, a spin-out of Northwestern University, raised $30M
Motivity, a clinical SaaS platform for applied behavior analysis (ABA) providers, raised $27M
Virtual physical therapy company Hinge Health filed to go public.
Consumer Digital Health & Other News
Instacart launched AI-Powered ‘Smart Shop’ technology and new features to enhance app personalization promising customers can make healthier and more informed choices.
Urgent care provider CityMD will partner with Notable to integrate artificial intelligence and automation technology.
📙Longread of the Month
Peterson Health Technology Institute (PHTI)'s report on “ambient scribe” technology is a good read about GenAI promises and experiences from early adopters
tl/dr: these tools likely improve clinician burnout, but the financial impact is unclear
🦜Tweet of the Month
Really loved this take by Ethan Mollick, this is definitely true for healthcare!
📊Chart of the Month
This tabular representation (with an embedded chart?) by Rock Health is an interesting overview of how different generations of consumers perceive and interact with digital health.
A former colleague and friend had a persistent skin problem that no dermatologist she visited seemed to be able to address. She then went to AI and actually found a solution and the problem has been resolved. She is 80!
The integration of multi-agent AI systems, such as Google's AI co-scientist and Microsoft's BioAgents, marks a significant advancement in hypothesis generation and bioinformatics, respectively. Recent studies indicate that these systems can enhance research efficiency by 30-40% through automated data analysis and hypothesis testing. The modular nature of BioAgents, for instance, allows for tailored applications in diverse bioinformatics tasks, potentially reducing computational overhead by 20%. Considering the recent surge in AI-powered health startups, how might these multi-agent systems be leveraged to optimize clinical trial design and patient recruitment in the context of personalized medicine?