Weekly Tech+Bio Highlights #65: Can AI Infrastructure Become Biotech’s Next Big Business?
2025 Pharma Review, Lilly Exec Explains Rationale Behind TuneLab AI Model Sharing & Biweekly News Roundup
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:
🤖 AI x Bio
(AI applications in drug discovery, biotech, and healthcare)
🔹 RosettaFold 3 and RFdiffusion3 push frontier of molecular design — UW’s Institute for Protein Design, with Microsoft, launches RosettaFold 3, a unified model for atom-level structure prediction across proteins, nucleic acids, and small molecules, while RFdiffusion3, now open source, enables design of DNA binders and catalytic enzymes.
🔹 New benchmark brings AI closer to real-world oncology care — Researchers release MTBBench, a NeurIPS-accepted dataset that models longitudinal, multimodal cancer patient journeys to evaluate foundation models and AI agents in Molecular Tumor Board–style clinical decision-making.
🔹 AI platform hits discovery milestone in big-pharma collaboration — Terray Therapeutics reached a discovery milestone with BMS using its EMMI platform, which couples high-throughput chemistry with generative AI to find novel small molecules for hard-to-drug targets.
🔹 AI-designed antibody enters late-stage testing for severe asthma — Generate:Biomedicines will launch two global Phase 3 trials enrolling ~1,600 patients to evaluate a long-acting anti-TSLP antibody, engineered with AI for six-month dosing, aiming to reduce asthma exacerbations and treatment burden.
🔹 Degrader design — Insilico unveiled an AI-generated PKMYT1 PROTAC that both degrades and inhibits its target, showing selective pathway shutdown and anti-tumor activity in biomarker-defined models with oral exposure in animals; the molecule has advanced to pre-candidate validation.
🔹 AI-guided repurposing flags new GBM candidate — Insilico and Atossa used a multi-omics/AI pipeline to rank (Z)-endoxifen as a top glioblastoma opportunity, identifying 1,400+ shared dysregulated genes and validating stronger in-vitro activity than standard therapy, with in-vivo tolerability supporting further exploration.
🔹 Quantum–biotech tie-up for advanced therapies — IonQ and CCRM formed a strategic partnership to apply hybrid quantum/AI methods to bioprocess optimization and disease-modeling within CCRM’s global regenerative-medicine network, with initial projects launching in 2026.
🔹 AI-designed antibodies with high hit rates from tiny design sets — Galux reports that just 50 de novo designs per epitope yielded ~32% binders and multiple picomolar hits, showing a shift from library-based antibody discovery to precise, small-scale computational design validated directly in IgG format.
🔹 BenchSci unveils LEAP to predict novel biology and design experiments — an AI inference engine that generates evidence-backed hypotheses and experimental plans by reasoning across a predictive knowledge graph built from multimodal, structured biomedical data. BenchSci is also hosting its first hackathon with Google to prototype AI co-scientists for preclinical research.
🔹 Ranking pathogenicity of every variant in a genome — HMS researchers unveiled popEVE, a model that scores each genetic variant by disease likelihood, distinguishing benign vs. pathogenic changes and predicting age-of-onset severity.
🔹 RapidAI deepens AWS partnership for global imaging AI — The companies will co-develop scalable clinical-AI tools, combining cloud pipelines with domain-specific imaging models to speed diagnosis and workflow automation across hospitals in 100+ countries.
🔹 IQVIA taps AWS to power agentic AI for clinical trials — The collaboration will deploy a cloud-based agentic AI platform to automate trial execution, medical affairs, and analytics at scale.
🔹 New biotech startup proposes shift from language to biology in oncology AI — Standard Model, emerged from stealth in September 2025, outlines its multimodal foundation model that predicts patient state changes over time, aiming to replace text-based medical AI with causal, intervention-aware simulations of disease progression.
This newsletter reaches over 10K industry professionals from leading organizations across the globe. To reserve your sponsor slot in one of the upcoming issues, contact us at info@biopharmatrend.com
🚜 Market Movers
(News from established pharma and tech giants)
🔹 GE HealthCare and Mayo Clinic build digital twins for personalized radiotherapy — The partners launched GEMINI-RT to use imaging, AI, and patient-specific digital models to automate radiation planning, predict treatment outcomes, and monitor side effects.
🔹 Imaging giants roll out end-to-end AI systems at RSNA 2025 — GE HealthCare, Philips, and Siemens unveiled AI-integrated MRI, CT, PET/SPECT, and interventional imaging platforms.
🔹 AI-guided antibody optimization expands — LabGenius and Sanofi launched a second collaboration in which LabGenius’ ML platform will optimize compact antibody formats for multiple inflammation targets, following successful results from their prior partnership.
💰 Money Flows
(Funding rounds, IPOs, and M&A for startups and smaller companies)
🔹 AI rekindles VC appetite for biotech — PitchBook reports that AI-native biotechs command nearly double the valuations of peers, with $3.2B invested across 135 deals in the past year, as foundation models and in-silico discovery convince investors that AI can cut R&D risk, cost, and timelines.
🔹 Gene-writing moves toward clinic — Tessera inked a $150M deal with Regeneron to advance its gene-writing therapy for AAT deficiency, with clinical trials set to begin soon; competing approaches from CRISPR Therapeutics and Prime Medicine are also nearing the clinic.
🔹 Automated chemistry for AI-driven drug discovery — Excelsior Sciences raises $95M, including a $70M Series A led by Deerfield, Khosla, and Sofinnova, to scale its machine-executable chemistry platform that enables closed-loop small molecule discovery and reshoring of pharmaceutical manufacturing.
🔹 Largest US oncology trial network takes shape — Paradigm Health acquired Flatiron’s clinical research business, creating a network spanning 25+ academic centers and ~100 community oncology sites and forming a long-term partnership to embed AI-enabled trial infrastructure into routine cancer care and expand access to more representative studies.
🔹 Human-data over animal models — Revalia Bio won up to $26.7M from ARPA-H to build AI models trained on donated human organs and organ-on-chip systems, aiming to predict drug safety/efficacy more accurately and cut late-stage failures.
🔹 $31.7M ARPA-H award to replace animal testing — Deep Origin will develop AI-driven virtual human safety models using ex vivo assays and foundational ML tools to accelerate and humanize drug testing.
🔹 Quantum drug sims get faster — QSimulate raised seed funding to $11M+ and released QUELO v2.3, enabling 1000x-faster quantum-level modeling now extended to larger and peptide molecules, supporting work with multiple global pharma partners.
🔹 AI-designed industrial proteins draw fresh capital — Aether Biomachines raises $15M to scale its AI/robotics platform for engineering proteins that enable faster, stronger materials and industrial biocatalysts for metal extraction, carbon capture, and pollutant cleanup.
⚙️ Other Tech
(Innovations across quantum computing, BCIs, gene editing, and more)
🔹 Stem-cell cure without HIV-resistant cells — a man with HIV and leukemia is now virus-free for 7+ years after a stem-cell transplant from a donor without the CCR5-null HIV-resistant mutation, reinforcing that cures may not require CCR5-deleted cells and expanding potential donor pools.
🔹 Human myelin repair-in-a-dish — Roche scientists built a myelinating human brain organoid that models demyelination and remyelination within ~2 months, revealing a key role for microglia in clearing damaged myelin and supporting repair.
🔹 BioAge advances NLRP3 inhibitor targeting age-related inflammation — Early Phase 1 data show BGE-102 is well tolerated with strong anti-inflammatory effects and brain penetration.
🏛️ Bioeconomy & Society
(News on centers, regulatory updates, and broader biotech ecosystem developments)
🔹 We’ve just published a new BiopharmaTrend feature on why Europe’s path to biotech success looks so different from the U.S. model.
🔹 Six months into deploying its generative AI tool Elsa, the FDA is hindered by fragmented, inconsistent data standards that limit AI’s impact; experts call for mandatory digital protocols, shared vocabularies, and stronger incentives to modernize regulatory submissions.
🔹 FDA expands AI use with new agentic system — introducing agentic AI tools to support complex tasks from reviews to inspections, building on its earlier LLM rollout as it seeks to accelerate workflows amid staffing cuts and missed decision deadlines.
🔹 Hybrid models take the lead in virtual cell prediction — Arc Institute announces winners of its inaugural Virtual Cell Challenge, where over 5,000 participants built AI models to predict cellular responses to genetic perturbations, with top teams combining statistical and deep learning methods to outperform pure AI approaches.
🔹 Pentagon flags WuXi AppTec as a potential security risk — The DoD urged Congress to add WuXi to its 1260H list for alleged ties to China’s military, a move that could strain the company’s U.S. business despite its denials and prior avoidance of stricter Biosecure Act sanctions.
🔹 CDC to end all monkey research in major policy shift — The CDC is shutting down its nonhuman primate research program, impacting ~200 macaques used in HIV and infectious disease studies, amid broader federal moves to reduce animal testing in favor of alternative methods.
2025 Pharma Review: Patent Cliffs, Crowded Pipelines, Modality Risk
Bruce Booth’s latest Atlas Venture presentation offers a dense overview of the current pharma landscape in 2025, with an unusually explicit focus on the political context around regulation, manufacturing, and funding.
Patent cliffs as main gravity
Over the next five years, $40–45B in branded revenue is expected to go generic each year (roughly triple the recent average) forcing big pharma to replace 8–10 blockbusters annually. That gap is driving aggressive BD, earlier licensing, and more willingness to back preclinical assets that can move quickly to proof of concept.Crowded mechanisms and compressed timelines
Thirty-six drug targets now have 50+ active programs globally, and time from first- to third-in-class approval has shrunk from ~15 years to under 2. In hot areas like GLP-1s, KRAS, and ADCs, programs compete on speed and capital efficiency as much as on biology, which erodes value for anything that isn’t clearly first- or best-in-class.Modality hangover after the “platform decade”
New modalities—gene therapy, gene editing, molecular glues, and others—are generating real signals in patients, but safety events (including roughly two dozen deaths in systemic gene therapy/editing trials), CMC complexity, and uncertain commercial models are forcing a recalibration of expectations after years of platform optimism.
Lilly Exec Explains Rationale Behind TuneLab AI Model Sharing
In a recent Decoding Bio feature, Zahra Khwaja interviews Aliza Apple, VP of AI/ML at Eli Lilly, for a deeper look at the thinking behind TuneLab—Lilly’s federated platform for sharing in-house AI models with biotech partners. Launched in 2025, TuneLab allows partners to train Lilly’s internal models on their own data without transferring it, using a federated learning setup. This setup keeps proprietary data local while updating shared models globally.
Apple emphasizes that the models offered through TuneLab (covering ADMET and antibody developability) are the same ones used internally, trained on over 500,000 data points from two decades of experiments (and ~$1B worth of proprietary drug discovery data).
Apple explains that Lilly’s own data flows through the same infrastructure, including pre-IP assay data, and that the models offered are the exact ones used internally. Rather than compete on early discovery tools like ADMET prediction, the idea is to open up infrastructure in areas where most biotechs face the same pain points. The strategy reflects a shift toward shared tooling and ecosystem-wide iteration—something Apple compares to co-developing SaaS platforms. The conversation also touches on future directions: integrating clinical data, incorporating third-party models, and exploring agentic AI workflows.
Can AI Infrastructure Become Biotech’s Next Big Business?
In a recent Century of Biology Substack post, Elliot Hershberg asks whether AI infrastructure companies in biology (those selling tools rather than drugs) can become large, investable businesses. He contrasts AI-native drug developers like Xaira, Formation Bio, and Isomorphic Labs with infrastructure/tool plays like EvolutionaryScale, Chai Discovery, Cradle Bio, Achira, Latent Labs, and Axiom, noting that skepticism is shaped by the Schrödinger precedent and the belief that drugs capture more value than platforms.
Hershberg structures the argument as three stories:
First, “compute as a reagent,” using next-generation sequencing as an analogy: Illumina, 10x Genomics, and Natera show how a foundational technology can scale from zero to a ~$10B+ market via elastic demand and strong margins, and he suggests AI+compute could fill a similar role as a ubiquitous research and clinical reagent, with NVIDIA and pharma GPU buildouts as early signals.
Second, “from the wet lab to the GPU cluster,” he describes AI models increasingly standing in for experiments (AlphaFold, virtual cell models, Amgen’s viscosity prediction), raising the prospect that a growing share of R&D experiments shift from physical CRO-style work to computational runs sold as services.
The third story, “AI for program decisions,” focuses on Axiom as an example of a company building toxicity prediction models that sit directly on high-value decision points in drug development, more analogous to Cadence in semiconductors than to a generic software vendor.
Across these three lenses, Hershberg frames the opportunity as a set of quantitative questions—what fraction of R&D spend, experiments, and per-program budgets will flow into AI tools over the next decade, and what share of that can new startups capture—rather than as a foregone conclusion that infra businesses will succeed.
Read also:
Three Big Ideas in Aging Research That Could Shift the Therapeutic Landscape