Weekly Tech+Bio Highlights #40: Foundation Models & Drug‑Discovery Economics
Also: FDA Sets Deadline for Internal AI Rollout, Structuring Meaning for AI with Ontologies, & Emerging Startups
Hi! This is BiopharmaTrend’s weekly newsletter, Where Tech Meets Bio, where we explore technologies, breakthroughs, and cutting-edge companies.
In today’s highlights: Toronto startup launches with $11M and a robotic formulation platform—Insilico Medicine files for a third IPO, the UK trains a generative AI on records from 57 million patients—Recursion cuts four programs to focus on oncology and rare disease—Moderna to onboard its R&D org to Benchling after digital team cuts—Vertex ends AAV gene therapy research—Siddhartha Mukherjee’s startup cuts 95% of staff—UNITY Biotechnology, once valued over $700M, lays off all employees.
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:
Let’s get to this week’s topics!
🤖 AI x Bio
(AI applications in drug discovery, biotech, and healthcare)
🔹 Amid growing diagnostic demand, North Bristol NHS Trust has become the first European reference site for FUJIFILM’s digital pathology platform, agreeing to help develop and showcase tools for large-scale, AI-integrated slide analysis.
🔹 Lantern Pharma has received FDA clearance to initiate a Phase 1b/2 trial of LP-184 for advanced non-small cell lung cancer (NSCLC) patients with KEAP1 and/or STK11 mutations and low PD-L1 expression, using its AI-driven platform to target a biomarker-defined, treatment-resistant population.
🔹 Iktos has made its generative AI drug design platform available on AWS Marketplace, offering pharmaceutical teams cloud-based access to tools for exploring chemical space, prioritizing compounds, and accelerating early-stage design.
🔹 Nabla Bio, a Cambridge startup combining generative AI with physics-based modeling for antibody design, reports in a new paper that test-time scaling—iterative inference without retraining—boosted hit rates and enabled de novo creation of rare CXCR7 agonists, highlighting a low-cost, model-agnostic route to improve design success.
🔹 AI agents could take on 55% of biopharma work, according to a study by Accenture and Wharton reported in R&D World Online by Brian Buntz—potentially unlocking $180–240B in annual value. Unlike past automation, these agents adapt via natural language, handling tasks from data retrieval to dynamic lab workflows, while escalating anomalies to humans for oversight.
🔹 Tristan Cazenave reportedly solved all 100 puzzles in the Eterna RNA design challenge using his AI system Montparnasse (arxiv)—a first for the field. The algorithm refines its design strategy during inference, achieving a 60% success rate on the hardest cases and offering a new path for programmable RNA molecules in synthetic biology and medicine.
🔹 The UK has trained a generative AI model Foresight on the health records of 57 million patients, using 10 billion medical events to predict over 1,000 conditions including heart attacks and hospital admissions. Built on LLaMA 2 and NHS data from 2018–2023, it's the first known attempt to train an AI on a whole national health system; some raise concerns.
🔹 A new study from NCATS, MIT, and UNC Chapel Hill shows graph neural networks can predict synergistic drug combinations in pancreatic cancer with 83% accuracy, validating 307 pairs from 1.6 million candidates. As Joseph Pareti of BioPharmaTrend notes, GNNs are effective because they operate directly on molecules’ graph structures, preserving the spatial and connectivity relationships critical to biological activity.
🔹 Romanian startup ZayaAI, which develops laboratory information systems and telepathology tools for cancer diagnostics and workflow automation, has integrated Paige’s AI models into its platform to support pathologists with AI-assisted cancer detection across prostate, breast, and GI tissues.
🔹 Pathos AI, a precision oncology startup, has appointed Iker Huerga as CEO following news of a collaboration with Tempus and AstraZeneca to develop a foundation model based on a large multimodal oncology dataset.
🔹 San Francisco–based Devano AI, a small software startup focused on automating biomedical data workflows, has announced its first strategic partnership with clinical-stage biotech Verge Genomics. Devano’s AI agents automate curation of gene expression and literature data, reportedly cutting prep time per disease indication from 30 hours to minutes.
🚜 Market Movers
(News from established pharma and tech giants)
🔹 Recursion is cutting three clinical programs, pausing one, and shutting down a preclinical asset to focus on oncology and rare diseases. The AI-focused biotech now advances six active projects and aims to out-license dropped assets like its C. difficile candidate, amid mixed trial results and narrowing priorities.
🔹 After cutting 10% of its digital team and the departure of its CIO in early 2025, Moderna plans to onboard its broader R&D organization to Benchling’s platform to centralize experiment tracking, workflow automation, and AI integration.
🔹 Vertex has discontinued its AAV-based gene therapy research, joining a broader industry shift away from viral vectors amid delivery and safety concerns. The company will continue investing in genetic and cell therapies.
🔹 AI-driven biotech insitro has laid off 22% of its staff—about 60 employees—to extend its runway into 2027, citing a “tumultuous market environment”. The company last raised funding in a $400M Series C round in 2021.
🔹 Cell therapy startup Vor Bio, founded by oncologist and author Siddhartha Mukherjee, is winding down operations and laying off 95% of staff after a strategic review and continued funding challenges. The company now retains just eight employees to explore potential asset sales, licensing, or merger options.
🔹 UNITY Biotechnology, once valued over $700M (commentary by
) and backed by high-profile investors like Bezos Expeditions and Founders Fund, has laid off all staff—including CEO Anirvan Ghosh—after years of setbacks in senescence-targeting therapies. Its lead assets failed to meet clinical endpoints in both osteoarthritis and ophthalmology, and with just $16.9M in cash remaining, the company is now pursuing strategic alternatives.💰 Money Flows
(Funding rounds, IPOs, and M&A for startups and smaller companies)
🔹 Insilico Medicine has filed for a third Hong Kong IPO after raising $110M in a March Series E and securing a $100M loan, aiming to build on renewed biotech market interest. The company’s lead drug, rentosertib, is in Phase 2a for pulmonary fibrosis, with additional programs in cancer, IBD, and other areas.
🔹 Skipping costly synthesis steps, Inductive Bio raised $25M in Series A to grow its AI system trained on a “pre-competitive consortium” of anonymized ADMET assay data. By aggregating results from diverse drug programs, the platform predicts key liabilities like toxicity and permeability early in design, helping chemists eliminate flawed compounds before lab work begins.
🔹 Dutch biotech Khondrion has secured €5M to begin the first phase of a Phase 3 trial for sonlicromanol, a small-molecule therapy for the m.3243A>G mitochondrial mutation. The study, planned for late 2025, follows earlier trials suggesting long-term symptom benefits.
🔹 Seqera has raised $26M in Series B to expand its cloud-native platform for life science R&D, built on Nextflow. The platform enables reproducible, scalable workflows across hybrid infrastructure, supporting everything from RNA-seq to AI-powered pipelines.
🔹 Swiss biotech Haya Therapeutics has raised $65M in Series A to advance its long non-coding RNA-targeting heart failure candidate and expand its RNA-guided drug discovery platform, with backing from Eli Lilly and other investors including Sofinnova Partners, Earlybird Venture Capital, and Alexandria Venture Investments.
⚙️ Other Tech
(Innovations across quantum computing, BCIs, gene editing, and more)
🔹 In a shift from electron-only brain mapping, Google researchers and collaborators at ISTA have developed the first method to reconstruct dense neural circuits using light microscopy alone.
🔹 Scientists led by Jure Tica and Mark Isalan have successfully engineered a synthetic gene circuit in E. coli that forms regular, self-organizing stripe patterns—demonstrating an implementation of a Turing mechanism in living cells (published in Cell Systems).
🔹 Crown Bioscience unveils a 3D bone marrow model that mimics native cell–matrix interactions for better early-stage drug testing. Presented at AACR 2025, the platform supports immune co-culture, patient-derived samples, and high-content readouts, aiming to reduce reliance on animal models and more accurately predict therapeutic resistance.
🔹 Scientists at Fujita Health University have developed a tool for precise mitochondrial DNA manipulation, using engineered enzymes to adjust the ratio of normal to mutant mtDNA in patient-derived stem cells. The method enables controlled modeling of the m.3243A>G mutation, offering a new way to study mitochondrial disease mechanisms and inform future therapies.
🔹 LatchBio has introduced image overlays for spatial embeddings in its AnnData viewer, allowing researchers to align histology images with transcriptomic data directly in Python.
🔹 A Harvard–MGH team, including Annabel Sorby-Adams, shows that portable, low-field MRI can still capture key brain imaging data, potentially making white matter scans faster and more accessible. The results, backed by Hyperfine, will be shared at ISMRM 2025 (May 12).
🏛️ Bioeconomy & Society
(News on centers, regulatory updates, and broader biotech ecosystem developments)
🔹 Fyodor Urnov (Innovative Genomics Institute) calls for a regulatory overhaul to enable broader CRISPR treatment of Mendelian diseases, arguing in The CRISPR Journal that current practices make development too costly and narrow, despite clinic-grade tools and delivery methods. He proposes treating mutation sets as syndromes, simplifying guide RNA approvals, and “platformizing” CRISPR akin to a pizza recipe—changing toppings, not redoing the dough.
🔹 The FDA has completed its first generative AI pilot for scientific review and plans a full internal rollout across all centers by June 30, 2025. The system cuts multi-day review tasks down to minutes, aiming to reduce administrative burden and accelerate drug evaluations while maintaining secure integration with FDA data systems.
🔹 FT reports on executive order to cut U.S. drug prices by up to 80%, tying Medicare payments to international reference prices—reviving a "most favoured nation" policy first floated in 2020.
🚀 A New Kid on the Block
(Emerging startups with a focus on technology)
🔹 Backed with $11M, Toronto-based Intrepid Labs has emerged from stealth with an autonomous formulation platform that combines active-learning algorithms with robotic execution to explore drug delivery options without relying on historical data. Emerging from University of Toronto research, the system conducts multi-objective optimization in real time—balancing factors like release profile and manufacturability—allowing formulators to iterate through hundreds of designs in days rather than months.
🔹 Milan-based ReportAId has raised €2.2M in pre-seed to turn free-text medical reports into structured care plans via AI, working with hospitals like San Raffaele to streamline workflows and close continuity gaps.
This newsletter reaches over 8.6K industry professionals from leading organizations across the globe. Interested in sponsoring?
Contact us at info@biopharmatrend.com
Foundation Models & Drug‑Discovery Economics
Maggie Basta (VP at Scale Venture Partners) writes—today’s wave of biological foundation models may succeed where past software tools in drug discovery have largely failed. The key difference is a shift from offering standalone insights to addressing recurring bottlenecks—molecular modeling, toxicity prediction, and de novo design—through scalable, infrastructure-level solutions. These are pain points with established budgets and newly capable technical buyers, not speculative opportunities.
Historically, Basta notes, software tools in biotech struggled to reach meaningful scale. Efforts to “vendorize” research—selling platforms that surface novel insights—often floundered, as value ultimately accrued to downstream assets. Schrödinger’s path is said to be instructive: despite widely used modeling tools, the company only gained traction after launching its own drug pipeline. Earlier genomic tooling startups met similar ends, unable to translate software into outcomes without pivoting to therapeutics.
See also: 13 Foundation Models: Startups, Industry Updates and the Nobel Prize
What’s different now, according to Basta, isn’t just that foundation models like AlphaFold 2 have expanded the technical frontier, but that their open release forced the industry to adapt. Lacking a service-heavy sales model, these tools demanded technical fluency from buyers. This shift—pharma companies building in-house capabilities to adopt and integrate AI—could open the door for more software-native business models in the future.
She also distinguishes the new AI tools not by their novelty alone, but by their potential to automate expensive, repeatable steps; predicting protein structures in minutes instead of months, simulating molecular interactions with growing precision, and screening toxicity in silico. The challenge remains high, particularly as many models risk becoming commoditized through open-source competition, but the path to monetization is clearer when framed around displacing large R&D line items rather than promising singular discoveries.
In Basta’s view, opportunities fall into three main domains:
Molecular modeling—where startups now compete to move beyond static folding toward dynamic simulations, incorporating physics and quantum chemistry.
Toxicity prediction, a market still dominated by lab tests with low predictive value, but now open to disruption through AI and shifting FDA policies.
De novo design, where generative models aim to produce novel biologics or small molecules from scratch. While the latter remains asset-centric, the former two offer a clearer route for software-led value creation.
Ultimately, Basta writes, the risk is shifting. It is no longer primarily pharmacological or clinical, but rather technical. That aligns better with the expertise and risk appetite of software-focused investors. But she tempers the optimism: despite the excitement, these are still asset-based bets at their core. For those entering with software-level risk tolerance and biotech-scale return expectations, disappointment remains a real possibility.
Related: We’ve recently published a new report outlining a framework for what defines modern AI-driven drug discovery for 2025 and beyond.
FDA Sets Deadline for Internal AI Rollout
On May 8, FDA Commissioner Dr. Martin A. Makary announced an accelerated timeline to implement generative AI tools across the entire agency by June 30, 2025, following what he called a successful pilot in scientific review. The internal deployment aims to reduce manual workloads and accelerate regulatory review of new therapies, with all centers adopting a unified, secure platform integrated with FDA’s data systems.
The initiative is led by newly appointed Chief AI Officer Jeremy Walsh and Sridhar Mantha. According to internal feedback, reviewers reportedly completed tasks in minutes that previously took several days. Makary emphasized the need to “value scientists’ time” and move beyond prolonged panel discussions on AI.
The move follows two other major AI- and NAM-related policy developments in 2025: in April, the FDA outlined its intent to replace animal testing in monoclonal antibody evaluations with human-based models and AI simulations; in January, the agency released draft guidance on AI use in regulatory decisions, highlighting risk assessment, lifecycle maintenance, and transparency.
Post-June, FDA will continue refining AI use cases, expanding functionality, and tailoring the system to center-specific workflows, while maintaining compliance with existing policy. Public updates are expected later this summer.
Presently, NIH is aligning with this direction. This month, it launched the Office of Research Innovation, Validation, and Application (ORIVA) to scale non-animal research technologies across its portfolio.
Structuring Meaning for AI with Ontologies
Knowledge‑graph architect Tony Seale argues that building an ontology (a structured representation of concepts and their relationships within a domain) is more than arranging taxonomies and hierarchies, it is a form of data factorisation.
Like dimensionality reduction in linear algebra, it projects complex, high-dimensional information into a lower-dimensional conceptual space, enabling AI systems to interpret data through a more structured lens.
By defining and naming key ontological classes—“Employee”, “Travel Expense”, “Diagnosis”, and so on—developers impose a selective, domain-informed filter over sprawling, unstructured data. This act of naming is not merely descriptive but a cognitive decision: it identifies what matters, abstracts it, and compresses surface-level complexity into a smaller, semantically rich subspace.
Seale frames these ontological classes as domain-specific features—essentially handcrafted priors for large language models and other systems. This process constrains entropy not by discarding information but by organizing it around meaning. As a result, AI pipelines gain clear axes of interpretation, improving linkages across sources and aligning outcomes with domain logic.
The takeaway: ontology is not simply documentation or backend scaffolding. It is a form of conceptual compression—a high-leverage intervention that enables faster iteration, greater explainability, and a more coherent internal representation of knowledge. For knowledge-centric AI, it is the starting point.
Cover image: “Highway and Byways” (1929) by Paul Klee, Swiss‑German modernist and former Bauhaus master, depicting layered horizontal bands in muted pastel tones.
Read also:
13 Foundation Models: Startups, Industry Updates and the Nobel Prize