Weekly Tech+Bio Highlights #33: The "Holy Grail" of Digital Biology
Utility of DNA Language Models, Digital Pathology Fundraisers, & Can AI Filter Out What Won’t Dissolve?
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:
Let’s get to this week’s topics!
🤖 AI x Bio
(AI applications in drug discovery, biotech, and healthcare)
🔹 Thought-controlled smart homes are no longer sci-fi. Synchron is partnering with NVIDIA to enhance its brain-computer interface (BCI) using the NVIDIA Holoscan platform, improving real-time neural processing and developing large-scale brain-language models, showcased with the unveiling of Chiral, a cognitive AI brain foundation model, at GTC 2025—where an ALS patient demonstrated hands-free control of lighting and appliances via Apple Vision Pro.
🔹 Google unveils TxGemma, a forthcoming suite of “open” AI models for drug discovery capable of interpreting both natural language and molecular structures, aiming to assist researchers in evaluating therapeutic properties—despite industry-wide caution following recent AI drug discovery stumbles.
🔹 Cells have social networks too—and AI is mapping them. Researchers have developed NicheCompass, an AI tool that analyzes how cells communicate within tissues using single-cell and spatial genomic data.
🔹 Roivant spinout VantAI has unveiled Neo-1, an AI model combining biomolecule structure prediction and molecular generation to design new drug candidates, including molecular glues and degraders—already in use with Janssen, Blueprint Medicines, and BMS.
🔹 Neuron23, a precision medicine company using genetics and AI for drug discovery, will present Phase 1 data showing its LRRK2 inhibitor reduced disease-related activity and was well-tolerated in 150 healthy volunteers, targeting both genetic and non-genetic Parkinson’s.
🔹 Techcyte, an AI-driven digital pathology company, launched Fusion, a platform combining anatomic and clinical pathology workflows, aiming to improve diagnostic accuracy and streamline lab operations by integrating AI, imaging systems, and patient data in a single, cloud-based system.
🔹 Hippocratic AI has launched a 4.2-trillion parameter suite of 22 healthcare LLMs that reportedly boosts clinical accuracy to 99.38%, enhances emotional intelligence and multilingual support, and integrates with major EMRs, improving patient satisfaction and engagement in over 1.85 million real-world calls.
🔹 Alpenglow Biosciences has introduced LUMI, a new interface for its 3Di Light-Sheet Microscope that streamlines high-resolution 3D tissue imaging with real-time feedback and AI-supported analysis, already used by five of the top 10 pharma companies to accelerate drug development and clinical diagnostics.
🔹 MIT and Harvard Medical School have developed VoxelPrompt, a vision-language agent that allows clinicians to analyze 3D medical images using natural language commands, unifying multiple radiology tasks and matching or exceeding specialized models in accuracy.
🚜 Market Movers
(News from established pharma and tech giants)
🔹 23andMe has filed for Chapter 11 bankruptcy after rejecting a takeover bid from CEO Anne Wojcicki, citing $277M in assets and $215M in liabilities—following strategic missteps, a major data breach, and struggles to pivot from genetic testing to drug discovery and telehealth.
🔹 Recursion has appointed former FDA Principal Deputy Commissioner Namandjé Bumpus and Mammoth Biosciences COO Elaine Sun to its board, bringing experience in regulatory strategy, drug discovery, and life sciences finance as the company expands its AI-driven drug development platform.
🔹 GE HealthCare has launched an AI-powered 3D ultrasound that enhances image quality, reduces scan time by 40%, and improves accuracy in detecting early-stage cancer in dense breast tissue.
🔹 Heart exams are getting an AI-powered upgrade as Fujifilm Healthcare partners with Us2.ai to equip its ultrasound system with automated echocardiogram analysis, cutting measurement and report creation time by 70% and helping address the global shortage of sonographers.
💰 Money Flows
(Funding rounds, IPOs, and M&A for startups and smaller companies)
🔹 European biotech startups are getting a major boost as Sofinnova Partners close its €165M Biovelocita II fund, surpassing targets with backing from Amgen, Bristol Myers Squibb, and Pfizer Ventures, expanding its biotech acceleration strategy beyond Italy to France, the UK, and Denmark.
🔹 Proscia has raised $50M, bringing total funding to $130M, to expand its AI-driven platform for pathology, supporting enhanced biomarker discovery and algorithm development as rising cancer cases and a shrinking pathology workforce drive demand for faster, more precise diagnostics.
🔹 Ampersand Biomedicines raised $65M, bringing total funding to $115M, to advance two immunology drug candidates for inflammation and cancer, using its machine learning-driven platform to improve drug delivery and reduce off-target effects as it expands partnerships with Pfizer and Flagship Pioneering.
🔹 ReactWise has raised $3.4M in pre-seed funding to develop AI-driven robotic labs to accelerate drug manufacturing by up to 30 times, using predictive models trained on high-throughput chemical reactions—already running 12 pilot studies with major pharma companies.
⚙️ Other Tech
(Innovations across quantum computing, BCIs, gene editing, and more)
🔹 Turning skin into brain cells—MIT engineers have developed a streamlined method to convert mouse skin cells directly into motor neurons using just three transcription factors, achieving a ~1,100% yield increase and showing early success in engrafting the neurons into mouse brains—offering hope for treating spinal cord injuries and diseases like ALS.
🔹 Brain chips that predict movement intentions without draining power—Researchers at San Diego State University have developed a low-power brain chip that monitors groups of neurons instead of single cells, automatically switching on when neural activity suggests the user wants to move.
🔹 A baby’s life saved by gene therapy. A world-first gene therapy by iECURE has successfully treated 14-month-old Tomas’s OTC deficiency, allowing his liver to eliminate toxic ammonia—freeing him from restrictive diets and medications and offering new hope for the rare disease community.
🔹 Inspired by the human brain’s wiring, Microsoft and Inait are partnering to develop AI models based on mammalian neural activity to create adaptive trading algorithms and responsive industrial robots capable of real-time learning.
🔹 Researchers have introduced NeuroBench, a universal benchmark for neuromorphic computing, providing a dual-track framework to evaluate both algorithms and hardware for tasks like vision, motor control, and chaotic function prediction.
🔹 Michael Levin (Tufts) and Rosalia Moreddu, in their pre-print in arXiv, have proposed cancer-inspired computing, a new paradigm that applies cancer’s adaptive strategies—like mutation, metastasis, and immune evasion—to develop more resilient and self-optimizing computing systems for fault tolerance and cybersecurity.
🏛️ Bioeconomy & Society
(News on centers, regulatory updates, and broader biotech ecosystem developments)
🔹 PicnicHealth report shows nearly 80% of life sciences leaders are considering virtual models for observational research to tackle data gaps and patient retention issues, with 61% already using AI-driven solutions for real-world evidence generation.
🔹 The end of the genetic paradigm of cancer—researchers are challenging the idea that cancer is purely a genetic disease, with a new essay in PLOS Biology arguing that inconsistencies in sequencing data suggest cancer may stem from disruptions in tissue organization and gene regulatory networks rather than just genetic mutations.
🚀 A New Kid on the Block
(Emerging startups with a focus on technology)
🔹 Exobiosphere, founded in 2024, has secured a contract from the Luxembourg government to develop the Orbital High-Throughput Screener, a space-based drug discovery platform using AI and automated screening to study disease modeling and drug interaction in microgravity.
This newsletter reaches over 8.2K industry professionals from leading organizations across the globe. Interested in sponsoring?
Contact us at info@biopharmatrend.com
The "Holy Grail" of Digital Biology
In a recent podcast on Ground Truths, Eric Topol explored the ambitious vision of creating a "virtual cell"—a comprehensive, AI-powered digital foundation model of cellular biology we wrote about in January 2025. Speaking with Charlotte Bunne (EPFL) and Steve Quake (Stanford University and Chan-Zuckerberg Initiative), the two of the fourty-two co-authors of a landmark paper published in Cell, Topol unpacked why the virtual cell is considered biology’s ultimate moonshot.
Quake articulated an ambitious goal for biology: to flip the ratio from being "90% experimental and 10% computational" to precisely the opposite within the next decade.
To realize this vision, the strategy involves building specialized computational architectures that integrate diverse types of biological data across multiple scales. At the core of this approach are “universal representations,” AI models designed to merge information from high-dimensional omics, multiplex imaging (which can capture up to 150 proteins in one experiment), and other variable data streams. Complementing these are “virtual instruments,” dynamic analytical tools that simulate the effects of perturbations on cells and help determine which new data should be collected to enhance predictive accuracy.
However, challenges remain—Topol notes the complexity inherent to cellular biology, which is far greater than earlier computational biology successes like protein structure prediction (AlphaFold). Unlike protein research, benefitting from centralized resources like Protein Data Bank, cellular biology data remain fragmented, dynamic, and often incomplete, although with dedicated GPU supercomputing clusters now available for non-profit research and new funding opportunities emerging, the field is gradually building the infrastructure needed to tackle these challenges.
Quake and Bunne argue that initiatives such as the Human Cell Atlas, aiming to map billions of individual cells, represent vast data resources that, when effectively harnessed, could drive major breakthroughs in building the virtual cell. Some efforts running in paralell include Noetik’s OCTO-VirtualCell platform, Recursion’s post-merger development of a world model for cell biology, and Google DeepMind’s work.
Utility of DNA Language Models
In his “Socratic dialogue”, machine learning engineer Abhishaike Mahajan takes a conversational approach to explore the utility of DNA language models—using the recent Evo 2 release by the Arc Institute as a prime example.
Evo 2, trained on a 9.3 trillion-nucleotide dataset spanning over 128,000 genomes from diverse life forms, can identify disease-causing genetic variants by assessing how "natural" a DNA sequence appears—essentially predicting whether a mutation might be harmful or benign based on evolutionary stability.
Rather than assuming that these tools will automatically revolutionize our understanding of genetic variants, Mahajan questions what they can realistically achieve. His discussion focuses on variant pathogenicity prediction, examining whether these models truly offer insights beyond traditional conservation scores, and hints at the broader potential—and limitations—of applying AI to decode the language of DNA.
Read the full piece by Abhishaike Mahajan at Owl Posting.
Digital Pathology is Rising
Katie Maloney, Partner at DeciBio, highlighted growing momentum in digital pathology, noting how Proscia recently raised $50 million in a funding round led by Insight Partners. This follows a series of major investments over the past six months in the digital pathology space:
Aignostics – $34M Series B (October)
Mindpeak GmbH – $15M Series A (September)
PathPresenter – $7.5M Series A (September)
According to Maloney, this wave of funding reflects increasing adoption of digital pathology solutions in clinical settings, though uptake has been gradual.
Andrii Buvailo, adds context around why digital pathology is attracting attention—traditional pathology relies on glass slides and manual examination, which is time-consuming and subject to human variability. Digital pathology replaces this with whole-slide imaging (WSI), producing high-resolution, digitized images of tissue samples that can be processed and analyzed at scale. AI models—particularly deep learning architectures like convolutional neural networks (CNNs)—automate pattern detection, biomarker measurement, and predictive diagnostics.
But the integration of AI doesn’t merely digitize an existing process—a point could be made that it fundamentally transforms the workflow itself, as the process evolves from a manual, time-consuming method to one that’s more efficient, consistent, and capable of uncovering deeper biology.
Predicting Kinetic Solubility
“Can AI filter out what won’t dissolve?” asks Pierre Llompart, from Sanofi R&D. Accurately predicting kinetic solubility—a key factor in drug development—remains a challenge despite improvements in AI models. Llompart summarized findings from a study led by Shamkhal Baybekov, published in Molecular Informatics, which analyzed a large dataset of 56,000 compounds to assess the feasibility of predictive QSPR (quantitative structure–property relationship) models for kinetic solubility.
One of the key takeaways is that kinetic and thermodynamic solubility are not interchangeable—meaning you can’t accurately predict one from the other, even though that’s been a common assumption in drug discovery. Surprisingly, the study found that kinetic solubility data are more reproducible across different labs than previously thought, which challenges the idea that this type of data is inherently unreliable. Interestingly, while models based on thermodynamic solubility data fail when applied to kinetic solubility, combining datasets from different sources significantly improves model performance.
Llompart’s core point is that AI-driven models for kinetic solubility can work—but only if the datasets are carefully curated and the models are properly trained.
This was an interesting read 👍
thanks for the interesting links! I wonder what your thoughts are on China's advancing biotech, which is a hot headline right now. I wrote about it here but would love to hear your pov https://chinahealthpulse.substack.com/p/biotech-in-china-four-important-truths