Weekly Tech+Bio Highlights #38: Predictive Models & Benchmarks
Also: Two-Hour Zebrafish Brain Benchmark, BCI Patent Landscape, Steering Complex Structure Predictions
Hi! This is
’s weekly newsletter, Where Tech Meets Bio, where we explore technologies, breakthroughs, and cutting-edge companies.Note: we’ve recently launched a new report outlining a framework for what defines modern AI-driven drug discovery for 2025 and beyond.
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)
🔹 Pfizer and AstraZeneca researchers developed a flow-matching AI model for ligand generation, conditioned on protein pocket geometry, achieving a 24-fold speedup over diffusion methods.
🔹 Gabriele Corso (MIT) announced Boltz-1x, an update to the Boltz-1 model for predicting 3D structures of biomolecules (full write-up later in the article).
🔹 NHS Wales pilots Paige’s AI-based pathology triage tool, trained on over 3 million digital slides, to flag suspicious cancer cases across multiple tissue types and prioritize them for diagnostic review.
🔹 Tempus introduces Loop, an AI-Powered target discovery and validation platform that integrates real-world patient data, human-derived biological models, and CRISPR screens to accelerate novel oncology target identification — additionally, Tempus has formed a new Advisory Board of leading U.S. physicians to guide its AI-driven precision medicine initiatives.
🔹 AstraZeneca, Tempus, and Pathos partner on a $200M project to build an oncology foundation model using multimodal de-identified data.
🔹 Immunai and Parker Institute collaborate to build a large real-world single-cell dataset for immunotherapy research, applying multi-omic profiling, AI-driven immune analysis, and 10x Genomics technology to uncover biomarkers, model immune response, and guide patient stratification.
🔹 Sonrai and Paige team up to make it easier for biopharma researchers to use AI for biomarker discovery and precision medicine, combining Paige’s foundation models with Sonrai’s no-code data platform.
🔹 Profluent launches ProGen3, a new family of billion-parameter AI models trained on 3.4B protein sequences to design antibodies and compact gene editors.
🔹 UC San Diego researchers found that elevated PHGDH levels may contribute to Alzheimer’s disease progression, with AlphaFold-based modeling providing a clue to a hidden DNA-binding role; a small molecule inhibitor later slowed disease features in mouse models.
🔹 Harvard researchers publish a cross-modality foundation model trained on 2M biomolecular complexes to predict molecular interactions across proteins, RNAs, DNA, lipids, and small molecules, outperforming specialized models and mapping disease pathways in 27 conditions.
🔹 Neurosnap unveils a hybrid AI model that designs high-affinity protein binders and optimizes solubility, stability, and immunogenicity, blending machine learning with physics and evolutionary constraints.
🔹 Duke researchers simulated a full lung cancer screening trial fully in silico, using virtual patients, scanner models, and AI readers.
🔹 Vidith Phillips from Johns Hopkins highlights a curated Notion list of open-source clinical, imaging, and biomedical AI models.
🔹 Tony Shen (Broad Institute/Northeastern) shares a new AI method to generate 3D molecules with synthesis pathways, stating that synthesizability remains a key bottleneck for AI-designed drugs.
🔹 Massachusetts General Hospital researchers found Hologic’s mammography AI caught 90% of known breast cancers and 32% of previously missed cases, performing better for aggressive tumors but missing some low-grade and lobular cancers, highlighting both strengths and detection gaps.
🚜 Market Movers
(News from established pharma and tech giants)
🔹 Vertex Pharmaceuticals' new research suggests that reducing opioid prescriptions by using nonopioid therapies could save $1.8B–$4.5B annually, prevent up to 260K cases of opioid use disorder and ~8,900 overdose deaths over 15 years, and avoid over 1M future cases of opioid use disorder.
🔹 Illumina announces a second round of layoffs in 2025, cutting 172 jobs in San Diego as part of a $100M cost-reduction plan. The layoffs follow an earlier cut of 96 positions and come amid trade tensions and China’s ban on Illumina’s sequencers.
🔬 Boehringer Ingelheim partners with Dutch startup Tessellate Bio to develop synthetic lethality-based therapies for hard-to-treat ALT-positive cancers, a segment lacking targeted treatments; the deal could exceed €500M in value.
🔹 Genentech partners with Flagship-backed Repertoire Immune Medicines, paying $35M upfront (plus up to $730M in milestones), to develop T-cell-targeted therapies for autoimmune diseases using Repertoire’s immune synapse-mapping platform.
🔹 Bayer forms multiyear partnership with ConcertAI to access its molecular database and AI tools, with the goal of supporting drug development and clinical trial design in oncology. The collaboration leverages data from the CancerLinQ network, which includes records from over 9 million patients.
💰 Money Flows
(Funding rounds, IPOs, and M&A for startups and smaller companies)
🔹 Alis Biosciences launches to recover ~$30B trapped in struggling public biotechs, using private equity-style structures to return cash to shareholders and manage residual IP, with plans to list publicly in the future.
🔹 Grove Biopharma raises $30M Series A, led by DCVC Bio with Eli Lilly and others, to advance a synthetic biology platform combining precision polymer chemistry, AI-driven computational chemistry, and protein engineering to develop cell-permeable molecules targeting intracellular protein-protein interactions.
🔹 Synthetic Design Lab emerges with $20M to develop next-generation antibody-drug conjugates (ADCs), aiming to boost targeted payload delivery more than 10x over current technologies. Founded by cancer immunotherapy veteran Daniel Chen (ex-Genentech), the company plans to enter clinical trials in 2026.
🔹 Biolinq raises $100M in Series C, led by Alpha Wave Ventures, to advance its needle-free glucose sensor patch toward FDA submission and commercialization for Type 2 diabetes management.
⚙️ Other Tech
(Innovations across quantum computing, BCIs, gene editing, and more)
🔹 In Cherry Biotech’s Organoids Digest, new organoid models replicate human sensory pathways, full liver zonation, and long-term functional growth of adult hepatocytes.
🔹 Microsoft releases a taxonomy of failure modes in agentic AI systems, identifying risks like memory poisoning, agent compromise, and multi-agent coordination failures, and recommending mitigations such as memory hardening, identity controls, and environment isolation.
🔹 Nuclera developed a faster method to express, purify, and stabilize membrane proteins like transporters and ion channels in 48 hours, using cell-free synthesis, digital microfluidics, and nanodisc-based screening — an approach supported by Innovate UK's engineering biology initiative.
🔹 Google Research, in collaboration with HHMI Janelia and Harvard, releases ZAPBench, a benchmark dataset capturing whole-brain neural activity at single-cell resolution in larval zebrafish, aimed at improving models that predict brain dynamics from structural and functional data.
🔹 BD launches a new hemodynamic monitor with AI-driven tools to predict blood pressure instability during surgeries, using a novel cerebral autoregulation index and noninvasive sensors to better personalize patient care.
🏛️ Bioeconomy & Society
(News on centers, regulatory updates, and broader biotech ecosystem developments)
🔹 In The Atlantic, Matteo Wong explores the reality behind Silicon Valley's promises that AI will “cure all disease”, finding that while models like AlphaFold and Google’s “co-scientist” are improving scientific workflows, the real-world impact remains incremental. AI has helped design drugs that passed early clinical trials and aided companies like Pfizer in identifying new cancer targets. But drug development timelines might shrink by only 3–5 years (from 20 years down to ~15), and even the best AI systems still struggle with hallucinations.
🔹 CRO Novotech reports over 800 global IPF trials launched since 2020, with Asia-Pacific leading in activity; trends include biomarker-driven patient stratification, small molecule and RNAi therapies, and early use of AI-designed drug candidates.
🔹 According to Asimov Press, China tripled its annual clinical trial volume between 2017 and 2023 by streamlining approvals, accepting foreign trial data, and incentivizing hospital participation.
🔹 BenchSci has partnered with Bentham Science to expand its experimental evidence discovery platform, adding access to over 130 journals covering cardiometabolic diseases, neuroscience, oncology, and immunology.
🚀 A New Kid on the Block
(Emerging startups with a focus on technology)
🔹 Axiom, announced by co-founder Brandon White, aims to replace animal testing in drug toxicity studies with AI models. The company raised $15M and built a dataset of over 100,000 molecules combining lab methods and structured clinical outcomes. Their AI model predicts drug-induced liver injury and is being evaluated through early pilots with pharma and biotech partners.
🔹 Flagship Pioneering launches Etiome focused on forecasting disease progression and developing early interventions, using AI to track biological changes over time. The effort, backed by $50M, aims to shift healthcare from reactive treatment to preemptive therapies for chronic and progressive diseases.
This newsletter reaches over 8.5K industry professionals from leading organizations across the globe. Interested in sponsoring?
Contact us at info@biopharmatrend.com
The Two-Hour Whole-Brain Benchmark
Last week we looked at C2S-Scale, a language-model toolkit for single-cell genomics. With HHMI Janelia and Harvard, Google Research now turns to whole-brain function with ZAPBench, an open dataset and benchmark built from calcium-imaging recordings of roughly 70,000 neurons in a six-day-old larval zebrafish.
Why zebrafish? The entire transparent brain fits under a light-sheet microscope, so nearly every neuron’s activity can be recorded while the fish views nine virtual-reality visual scenes.
The experiment fixes the fish in a jelly-like agarose gel, leaves its eyes free, and places it inside a panoramic screen. Nine visual conditions—shifting currents, alternating light and dark, looming objects and others—are projected while a light-sheet microscope sweeps a laser through the brain one slice at a time. The fish expresses the calcium indicator GCaMP, so active neurons flash green as calcium enters the cell.
Two hours of 3D volumes are recorded while tail electrodes capture muscle attempts to swim against the simulated currents.

ZAPBench ships two versions of the data: a 3D “movie” of the whole brain and a lighter spreadsheet-style file that lists each neuron’s activity over time. The test is straightforward—give a model a short clip, then see how well it can guess what the brain will do 30 seconds later.
The authors report that giving models longer stretches of prior activity improves prediction accuracy; they further emphasize another pattern:
…volumetric training (on the 3D video) out-performed the time-series models in some instances, which we suspect is thanks to the 3D models having access to the spatial relationships between cells. Also, interestingly, when the models make mistakes they tend to happen more frequently in specific areas of the brain, meaning some brain areas are harder to predict than others. We also saw that video models worked well with lower resolution data.
While apturing two uninterrupted hours of activity from nearly every neuron in a single fish, ZAPBench pairs it with a forthcoming full connectome of that same specimen—delivering what appears to be the first benchmark that links whole-brain function to matching structural wiring.
Electron-microscopy reconstruction of that same brain is underway. Once the synaptic map is released, models will be able to anchor their activity forecasts in ground-truth wiring.
All data and code are openly licensed, so any group can download and work with ZAPBench.
BCI Patent Landscape
PatentVest recently released a report looking at how intellectual property is shaping the brain-computer interface (BCI) field. With neurological conditions like paralysis, epilepsy, and treatment-resistant depression driving demand, BCIs are seen as a next-generation technology—not just restoring lost function, but potentially shifting care models from reactive to proactive.
The report frames U.S. brain-computer-interface (BCI) opportunity as “exceeding USD 400 billion”, with roughly USD 80 billion in near-term, high-acuity indications and about USD 320 billion tied to longer-range neurological and psychiatric use-cases. The same report counts more than 2,160 BCI-related patent families across 664 organisations. The report’s discussion mostly spotlights implantable BCIs.
Companies like Synchron, Neuralink, INBRAIN Neuroelectronics, Blackrock Neurotech, Precision Neuroscience, and Paradromics are moving early to secure technical and clinical positions—but also to build defensible IP portfolios. Synchron, in particular, is highlighted for combining a minimally invasive approach with partnerships and patent filings across 10 jurisdictions.
Beyond these, there’s a broader ecosystem across a range of companies, including Kernel, NeuroPace, Neurolutions, Cognixion, Neurable, Snap (via its acquisition of NextMind), Panasonic, Arctop, and CereGate.
Universities like Tianjin, Stanford, and the University of California system also continue to own a large share of foundational patents in areas like neural signal processing and interface design, shaping the base technology that newer entrants build from.
Steering Complex Structure Predictions
Another open-source update comes from the complex structure prediction space. Researchers behind Boltz-1 have introduced Boltz-1x, an enhanced model aimed at generating physically more accurate biomolecular complexes.
Previous models—including Boltz-1 itself, AlphaFold3, Chai-1—often produced structures that broke physical rules: distorted bond lengths, wrong stereochemistry, steric clashes, or overlapping chains. These artifacts limited the use of predicted structures in downstream tasks like molecular simulation or drug design.
Boltz-1x tries to address this with a new inference-time technique called Boltz-steering—a method that gently biases the model’s outputs toward more physically plausible structures as they are generated.
In evaluations, Boltz-1x maintains the structural prediction accuracy of Boltz-1 but shows improvements in the physical validity of generated poses. Early tests suggest that nearly all Boltz-1x structures pass key physical plausibility checks that earlier versions often failed.
The full model, codebase, and benchmarks are openly available under an MIT license.
Read also:
13 Foundation Models: Startups, Industry Updates and the Nobel Prize