Weekly Tech+Bio Highlights #47: Microsoft AI's Diagnostic Multi-Agent Orchestrator
Also: New Foundation Model for Zero-Shot Antibody Design & The Role of AI in Rare Disease Tech
Hi! This is BiopharmaTrend’s weekly newsletter, Where Tech Meets Bio, where we explore technologies, breakthroughs, and cutting-edge companies.
If this newsletter is in your inbox, it’s because you subscribed, or someone thought you might enjoy it. In either case, you can subscribe directly by clicking this button:
We just published a feature for our paid subscribers exploring the fast-evolving world of brain-computer interfaces—from the history of EEG and early implants to Neuralink’s recent trials, non-invasive wearables, and a growing wave of AI-enhanced neurotech.
If you’re interested BCI’s clinical progress, emerging startups, national roadmaps, and ethical flashpoints—take a look!
🤖 AI x Bio
(AI applications in drug discovery, biotech, and healthcare)
🔹 Programmable antibody design with generative AI—Chai Discovery releases Chai-2, an all-atom generative model achieving 16-20% hit rates in zero-shot antibody design across 52 novel antigens.
🔹 Deniz Kavi of Tamarind Bio announces IntFold, a new structure prediction model matching AlphaFold 3 in accuracy on FoldBench, with improved performance on antibody-antigen and protein-ligand tasks, aimed at advancing structure-guided drug discovery.
🔹 Cell atlas and AI curation tools—Miraomics, Pythia Biosciences, and LatchBio release a 30 million cell atlas covering 150 diseases and 200 tissue types, alongside an AI-powered curation framework designed to automate and accelerate molecular data structuring for drug discovery and research.
🔹 Five-year AI partnership in oncology—Owkin and Newcastle upon Tyne Hospitals NHS Foundation Trust launch a five-year collaboration to apply agentic AI in medical research.
🔹 XtalPi and Pfizer deepen AI drug discovery collaboration—expanding their partnership to develop a high-throughput AI- and physics-based platform for small molecule drug discovery.
🔹 AI for EU drug regulation—Sweden’s Medical Products Agency launches REGULUS, a generative AI tool for regulatory support.
🔹 On-prem LLMs in clinical practice—Maxime Griot and collaborators report on deploying a GDPR-compliant chatbot using open-source LLMs within the Epic EHR at a European hospital.
🔹 AI accelerates sperm detection in infertility care—Columbia University Fertility Center used AI to identify viable sperm in a man with azoospermia within hours.
🔹 Can AI build a virtual cell?—Ewen Callaway reports on major efforts by Chan Zuckerberg Initiative, Arc Institute, and DeepMind to create AI models simulating cell behavior.
🔹 PathAI receives 510(k) clearance for its digital pathology platform used in primary diagnosis.
🔹 TOBY, Inc. receives FDA Breakthrough Device Designation for its AI-powered urine test for early bladder cancer detection.
💰 Money Flows
(Funding rounds, IPOs, and M&A for startups and smaller companies)
🔹 Medtech VC funding hit $4.1B in Q1 2025—its strongest showing since early 2022—according to Dr. Luka Nićin of Pace Ventures, with large rounds from Neko Health ($260M) and EverBridge Medical ($139M) helping drive a modest recovery despite continued exit challenges.
🔹 AI-enhanced protein sequencing—Portal Biotech raises $35M Series A co-led by NATO Innovation Fund and Earlybird to commercialize AI-powered nanopore technology for full-length single-molecule protein sequencing.
🔹 XtalPi backs AI-driven anti-aging startup—XtalPi leads pre-seed funding in MIT-led Foundry BioSciences to co-develop gen-AI protein design tools for anti-aging therapies.
🔹 Longevity startup targets aging with plasma exchange—Circulate Health raises $12M in seed funding led by Khosla Ventures to expand therapeutic plasma exchange clinics, following trial results showing reduced biological age by 2.6 years.
🔹 Australia-based Venstra Medical received $1M from MTPConnect to advance its miniature heart pump for cardiogenic shock.
🔹 AbbVie acquires Capstan Therapeutics for up to $2.1B to access its targeted LNP platform for in vivo mRNA-based CAR-T therapy.
🔹 Gene control for advanced therapies—Laverock Therapeutics raises over €23.3M in expanded seed funding (led by Calculus Capital, also Eli Lilly and others) to advance its programmable gene control platform for oncology and genetic medicine applications.
🔹 AI-native clinical OS—Tandem Health raises $50M in Series A funding led by Kinnevik to develop an AI-native operating system for clinical workflows, expanding from AI-generated documentation to broader tools for coding, care coordination, and decision support across European health systems.
🔹 Argenx inks $1.5B deal for AI-designed macrocycles—Argenx partners with Unnatural Products to develop AI-guided macrocyclic peptides against undruggable targets.
🔹 French startup RainPath AI raises €2.5M (Teampact.ventures, THE QUEST, ADVANS Lab, SHARPSTONE, Bpifrance, and business angels) to advance its AI-driven virtual staining technology for faster, reagent-free biopsy analysis.
⚙️ Other Tech
(Innovations across quantum computing, BCIs, gene editing, and more)
🔹 Tomb fungus yields leukemia drug lead—UPenn researchers discover cancer-fighting molecules in Aspergillus flavus, a fungus linked to King Tut's tomb, showing leukemia-specific activity and paving the way for future animal and human trials.
🔹 Inside the British lab growing a biological computer—Financial Times piece explores how Cortical Labs and bit.bio built a hybrid biocomputer ($35k CL1) with 200,000 live neurons trained to play Pong and respond to stimuli, raising hopes for brain-based computing—and ethical questions about consciousness and memory in living circuits.
🔹 Tobacco plants engineered to produce cancer drug precursor—scientists inserted 17 yew tree genes into tobacco to sustainably make baccatin III, a key step toward low-cost Taxol production.
🏛️ Bioeconomy & Society
(News on centers, regulatory updates, and broader biotech ecosystem developments)
🔹 FDA removes REMS restrictions on all approved CAR-T therapies for blood cancers, enabling broader access by allowing administration in community centers and reducing monitoring burdens, potentially doubling treatment uptake.
🔹 Fred Hutch’s Aaron Ring argues AI won’t revolutionize drug discovery for Big Pharma but is transforming early-stage development for small labs and startups.
This newsletter reaches over 8.9K industry professionals from leading organizations across the globe. Interested in sponsoring a spot?
Contact us at info@biopharmatrend.com
Towards Autonomous Medical AI
Microsoft AI built “MAI Diagnostic Orchestrator” (MAI-DxO), a model-agnostic orchestrator to simulate a virtual panel of physicians with different diagnostic styles, working together to solve complex cases. The system can ask follow-up questions, order tests, and refine its hypotheses while operating under a set budget to avoid excessive testing and simulate real-world trade-offs.
To test it, the team built a sequential diagnostic benchmark from 304 complex, narrative-style cases published in the New England Journal of Medicine, restructuring them into stepwise challenges that simulate real clinical reasoning—asking questions, ordering tests, reviewing results, and refining a differential.
The orchestrator wraps around foundation models like GPT, Llama, Claude, Gemini, Grok and DeepSeek, coordinating them as if they were a panel of virtual physicians. In benchmark tests, the highest-performing configuration (MAI-DxO + OpenAI’s o3) correctly diagnosed 85.5% of cases—compared to an average of 20% by a group of 21 experienced physicians from the US and UK.
Notably, the AI reached correct diagnoses using fewer and less costly tests than both the clinicians and the individual LLMs acting alone. The evaluation introduced virtual test costs to pressure models into cost-aware reasoning, surfacing trade-offs between diagnostic accuracy and resource use.
While the tool is not available for public or clinical use, the team is preparing a formal peer-reviewed publication and exploring public release of the Sequential Diagnosis Benchmark (SD Bench). The researchers acknowledge current limitations (like the lack of validation on routine primary care presentations) and note on the need for real-world trials and regulatory guardrails before broader deployment.
New All-Atom Foundation Model for Antibody Design
Chai Discovery has introduced Chai-2, a generative model designed for zero-shot antibody and binder discovery. It works directly from target structure and epitope, without relying on existing antibody templates or scaffold libraries. According to the company, the model produced successful binders for over half of 52 novel antigens, with hit rates in the ~16–20% range using just 20 designs per target.
The system handles a range of modalities like scFvs, nanobodies, and miniproteins, and was also tested on targets generally considered computationally difficult. It builds on Chai’s earlier single-sequence structure prediction work and expands into atomic-level modeling across diverse molecular formats.
Chai frames this as a step toward programmable biologics—generating candidates that go straight from in silico design to wet-lab testing, potentially bypassing screening libraries and reducing iteration cycles. The model isn’t yet optimized for full therapeutic profiling, but future versions are expected to integrate manufacturability and pharmacokinetic constraints.
Access is being offered selectively under the company’s Responsible Deployment framework. The company previously released Chai-1 in late 2024, which used single-sequence structure prediction (no MSAs) and scored 77% on PoseBusters. Chai raised ~$30M at a $150M valuation within six months of founding, with backers including OpenAI, Thrive Capital, and Dimension.
The Role of AI in Rare Disease Tech
In her latest rare disease tech column for BiopharmaTrend , Louise von Stechow explores how AI is helping decode the dark genome—the 98% of human DNA that doesn’t code for proteins but holds most disease-linked variation.
An estimated 80% of rare diseases are genetic, and about 90% of those variants lie in non-coding regions. Long-read sequencing, including the 2022 Telomere-to-Telomere project, has made these regions more accessible. Now, models like DeepMind’s AlphaGenome aim to predict variant effects across both coding and non-coding DNA at scale.
The piece closes with a case from CHOP, where genome sequencing helped guide a custom gene-editing treatment for an infant—offering a glimpse into what personalized rare disease medicine might look like when the dark genome comes into focus.