Weekly Tech+Bio Highlights #29
Also: This Company Suggests Six Criteria for Assessing AI’s Role in Drug Discovery, Modernizing Clinical Trials, AI-Designed Enzymes, New Ultrafast Protein Labeling...
Hi! This is
’s weekly newsletter, Where Tech Meets Bio, where we explore technologies, breakthroughs, and the great companies driving the biopharma and medtech industries forward.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!
Brief Insights
🔬 NVIDIA, in collaboration with Arc Institute and Stanford, has launched the largest publicly available AI model for biomolecular research. Evo 2 analyzes DNA, RNA, and proteins across species, aiding drug discovery, genomics, and biotechnology applications.
🔬 Isomorphic Labs expands its AI-driven drug discovery collaboration with Novartis, adding up to three more research programs on the same financial terms as the original 2024 agreement.
🔬 A STAT report by Brittany Trang examines AI-driven de novo antibody design, highlighting how some companies market their antibodies as AI-generated despite making only minor modifications to existing ones, raising questions about the true definition of de novo design and AI's role in antibody discovery.
🔬 AI-designed enzyme breaks down plastic by targeting ester bonds, demonstrating a multi-step catalytic process rarely found in nature.
🔬 A new Radiology paper from RSNA explores foundation models in radiology, detailing their potential for workflow automation, report generation, and diagnostics, while also addressing key challenges like bias, hallucinations, and regulatory gaps.
🔬 AI is reshaping drug discovery, but pharma’s biggest hurdles aren’t technical—WSJ reports industry leaders cite bureaucracy, risk aversion, and talent shortages as the real bottlenecks. While AstraZeneca, GSK, and Genentech report AI-driven breakthroughs, executives argue that rigid corporate structures and short-term thinking slow adoption.
🚀 Atomwise, an early AI-driven biotech, has named Steve Worland as CEO, succeeding founding CEO Abraham Heifets, per Endpoints News’ Andrew Dunn. The company, which applies deep learning for structure-based drug design, has downsized to 28 employees from over 100 and raised a $45M Series C last year, with a focus on "…candidates that move into the clinic and show clinical data."
🔬 Biopharma’s slow clinical trials may finally speed up—McKinsey reports that modernizing clinical IT systems could cut trial startup times by 15–20% and shorten study length by up to 30%.
🔬 Insilico Medicine sets industry benchmarks for AI-driven drug discovery, reporting a 13-month average to preclinical candidate nomination—cutting timelines in half.
🚀 The architects behind AlphaFold are back—Latent Labs, co-founded by DeepMind alumni, has emerged from stealth with $50M to push AI-driven protein design beyond prediction and into molecular engineering.
🔬 Machine learning isn’t a magic wand for biotech data—Eric Ma, Senior Principal Data Scientist at Moderna, breaks down the pitfalls of integrating public and lab-generated datasets, from inconsistent measurements to domain shifts. Instead of force-fitting everything into one model, he advocates keeping models separate and merging insights with human expertise.
🔬 The UK government has awarded £18.9M to PharosAI, a company using AI to analyze decades of NHS cancer data, accelerating AI-powered diagnostics and drug discovery. PharosAI aims to refine oncology datasets for AI model training to improve early cancer detection and treatment.
🔬 In a podcast interview, Zelda Mariet, co-founder and Principal Research Scientist at Bioptimus, discussed the company’s approach to foundation models for biology, emphasizing the importance of a holistic, multimodal model that goes beyond single-modality approaches to better capture biological complexity.
🔬 ProtGPS, a new deep learning model developed by researchers at MIT and the Whitehead Institute, predicts where proteins localize within cells, revealing hidden molecular sorting rules that influence biological organization and disease.
🔬 MIT researchers demonstrate a method for labeling proteins across millions of individual cells in intact 3D tissues, enabling single-cell protein profiling in whole organs within a day and revealing discrepancies between genetic and antibody-based labeling in brain tissue analysis.
🔬 ExpressionEdits partners with Boehringer Ingelheim to optimize gene therapy expression using AI-driven intronization technology, aiming to improve protein production and therapeutic efficacy without altering genetic sequences.
🔬 In an interview with BioTechniques, Ola Engkvist, Executive Director and Head of Molecular AI at AstraZeneca, says AI now assists 85% of the company’s small molecule and PROTAC drug discovery projects.
🔬 Oxford researchers have developed a new liquid biopsy method which preserves more DNA than traditional bisulfite sequencing, allowing simultaneous genomic, methylomic, and mutational analysis. This multimodal approach improves cancer detection accuracy, achieving 94.9% sensitivity and 88.8% specificity. Alex Dickinson, in his LinkedIn post, contrasted this with GRAIL’s Galleri test, which relies solely on methylation data, noting that TAPS' broader biomarker integration enhances signal detection and could make advanced cancer screening more accessible and cost-effective.
💰 Spatial biology startup Stellaromics raises $80M to support the commercialization of its 3D sequencing platform for spatial multiomics. The technology enables high-resolution mapping of gene expression and cell clusters in intact tissue, analyzing slices 100 µm thick or more and tracking the spatial distribution of thousands of genes, with early adoption in neuroscience and cancer research.
🔬 Researchers from University College London and the Broad Institute identify an antisense oligonucleotide (ASO) that slows the gene expansion process driving Huntington’s disease, showing in lab-grown human neurons that treatment halts expansion and reduces targeted protein levels in mouse brains, with dose-dependent effects observed.
A recent Cell study found that Huntington’s disease is driven not by toxic huntingtin protein buildup but by the progressive expansion of CAG DNA repeats in the huntingtin gene. The gene remains harmless until it reaches a toxic threshold, triggering neuron death, highlighting gene expansion as a potential therapeutic target.
🚀 Tasso launches a new at-home blood collection system that enables precise, 17.5-microliter capillary blood sampling for clinical trials, drug monitoring, and anti-doping tests, with dried samples that can be stored at room temperature.
📉 DNA sequencing firms Illumina, Pacific Biosciences, and 10x Genomics saw stock drops of 7% to 14% after the NIH moved to cap indirect research grant costs at 15%, limiting funding for lab equipment and facilities.
💰 Globus Medical acquires Nevro Corp. for $250M, expanding into neuromodulation and pain management with Nevro’s spinal cord stimulation and minimally invasive sacroiliac joint treatments.
🔬 The FDA approves AbbVie and Pfizer’s Emblaveo, combining aztreonam and avibactam to tackle multidrug-resistant Gram-negative infections in complicated intra-abdominal cases, U.S. launch expected in Q3 2025.
🔬 Researchers demonstrate how explainable artificial intelligence can improve predictions in targeted protein degradation by revealing key molecular structures influencing drug activity. In a new study, scientists applied AI models to Cereblon-based degraders targeting GSPT1 and found that explainability methods mirrored medicinal chemistry insights.
🔬 David Baker’s team at the Institute for Protein Design has computationally designed a multi-step serine hydrolase enzyme, achieving the first AI-designed enzymes capable of completing full catalytic cycles—though still far slower than their natural counterparts.
🔬 A ChemRxiv preprint by researchers at Schrödinger introduces an automated physics-based method for predicting absolute drug residence times, a key pharmacokinetic property.
🔬 Ginkgo Bioworks introduces a generative AI model for designing full-length mRNA sequences, leveraging discrete diffusion to optimize coding sequences and untranslated regions for stability, translation efficiency, and expression.
🔬 Researchers from the University of Tokyo and Waseda University develop the world’s largest biohybrid robot hand, integrating living human muscle tissue into a robotic skeletal structure. Published in Science Robotics, the study introduces a high-performance muscle actuator made of bundled cultured tissue, enabling independent finger movements and complex gestures.
🔬 In BiopharmaTrend Insight, Bea Mann, PhD, from ICON highlights the lag between oncology biomarker development and the rapid progress of precision cancer therapies, noting the potential of AI-driven biomarker discovery, alongside advanced trial strategies, to help close this gap.
📉 The FDA’s recent layoffs have disproportionately affected its AI and digital health division, raising concerns about the agency’s ability to regulate the growing use of AI in healthcare.
🔬 biotx.ai and 3Z Pharmaceuticals, building on their 2023 collaboration, applied causal AI modeling to mimic clinical trials in human genetic data, repositioning amlodipine—a widely used hypertension medication—for ADHD.
This newsletter reaches over 8K industry leaders interested in cutting edge technologies in drug discovery, biotech, and healthcare.
Interested in sponsoring?
Contact us at info@biopharmatrend.com
Measuring AI’s Role in Drug Discovery
A recent STAT News report by Brittany Trang examines how AI-driven biotech companies present their capabilities, highlighting the need for greater transparency in the field. The article looks at Absci and Generate Biomedicines, both focused on computational protein design, and discusses the distinction between de novo AI-generated molecules and modifications of existing antibodies. The broader issue raised is how AI’s role in drug discovery is communicated and the need for clear benchmarks to assess its actual contributions.
Model Medicines, a drug discovery company founded in 2019 and focused on AI-driven small molecule therapeutics, has proposed a structured framework called The AIDD Code to help evaluate AI’s impact in drug design. According to Model Medicines, these six technical criteria offer a measurable way to distinguish real AI-driven innovation from incremental drug development.
New Biology & Target Discovery – AI should identify new biological targets rather than just optimizing existing drugs.
New Chemical Entities (NCEs) – AI should generate entirely novel molecular structures, rather than small modifications of known drugs.
Zero-Shot & One-Shot Hit Rates – AI predictions should achieve a 30% success rate of active compounds at <20μM in initial testing.
Tanimoto Scores <0.5 – AI-generated molecules should be structurally distinct from known drugs, rather than minor variations.
A Diverse Pipeline – AI should be capable of identifying promising drug candidates across multiple therapeutic areas.
Preclinical Proof-of-Concept – AI-designed molecules should show real efficacy in disease models, beyond computational predictions.
These criteria are meant to provide a systematic way to assess whether AI is genuinely responsible for drug design, rather than being used as a tool to make small modifications to existing compounds. The company argues that without clear metrics, the industry risks conflating computational assistance with true AI-driven discovery, making it difficult to separate innovation from hype.
The broader discussion around AI in drug discovery has been shaped by a lack of transparency regarding key performance metrics. While some companies have disclosed specific data—such as Insilico Medicine, which recently published preclinical benchmarks detailing timelines and molecular synthesis rates—many others remain vague about how their AI systems actually contribute to drug development.
Modernizing Clinical Trials
In a recent publication, McKinsey authors outline how modernizing the R&D IT “application layer” can address issues such as fragmented data and labor-intensive trial processes. They cite analysis indicating that updated core systems may cut study start-up times by 15% to 20%, shorten trial length by 15% to 30%, and improve trial success rates by around 10%. Their research also notes an annual value opportunity of up to $50 billion from advanced analytical methods, including generative AI.
The article points out that although some top biopharma firms have started upgrading their systems, a large share have not yet achieved substantial returns on investment. McKinsey’s framework covers four connected layers—analytics, applications, data, and infrastructure—and emphasizes the importance of interoperability among them.
The authors propose a step-by-step approach:
Defining the scope of modernization, from end-to-end clinical processes to targeted data strategies.
Pinpointing areas to differentiate, such as AI-supported use cases or decentralized trials.
Choosing between platform, best-of-breed, or hybrid solutions.
Selecting vendor combinations based on interoperability and long-term compatibility.
Ensuring business and IT teams share joint ownership to align technical decisions with clinical needs.
McKinsey’s piece also discusses the need for clear metrics to track progress, such as accelerated database lock times and lower operating costs.
According to the authors, modernized tools can help companies meet rising demands for speed and quality in clinical development while supporting real-time analytics, patient enrollment strategies, and more flexible trial designs. They suggest that focusing on both technical upgrades and organizational changes—such as staff training and streamlined workflows—will be key to realizing the potential benefits.
AI-Designed Enzymes Complete Full Catalytic Cycle
Researchers at the Institute for Protein Design have used AI to create multi-step enzymes, marking a significant step in computational enzyme engineering. The study, published in Science, details how the team, led by David Baker, designed new serine hydrolases capable of completing full catalytic cycles—something not previously achieved through AI.
The challenge in enzyme design lies in accounting for dynamic structural shifts during catalysis. Most computational methods rely on static models, making it difficult to design enzymes that undergo multiple transformations. To tackle this, the team used RFDiffusion, a generative AI tool, to create 10,000 enzyme candidates. These were filtered through AlphaFold2 for structural predictions, followed by a more refined screening using PLACER, which assessed atomic interactions based on chemical and physical principles.
Through multiple iterations, the researchers improved the success rate of functional designs from 1.6 percent to 18 percent, with two enzymes achieving full catalytic turnover. However, their activity remains far slower than natural enzymes, a limitation the field aims to overcome in future work.
has a detailed write-up on this breakthrough by and —check it out!Mapping Millions of Cells in a Day
MIT researchers have developed a new method for ultrafast protein labeling, enabling the visualization of proteins across tens of millions of individual cells in intact 3D tissues within just 24 hours—compared to weeks required by traditional methods, Published in Nature Biotechnology.
The key challenge in large-tissue labeling has been the slow diffusion of antibodies, leading to uneven staining where outer layers are saturated while deeper regions remain unstained. To address this, the MIT team introduced CuRVE (Controlled Uniform Rapid Visualization of Elements), implemented in the eFLASH system, which:
Modulates antibody binding speed using deoxycholic acid, ensuring even distribution.
Uses stochastic electrotransport to accelerate antibody movement through tissue.
Achieves uniform labeling in whole organs, validated with over 60 different antibodies.
In testing, eFLASH revealed discrepancies between antibody labeling and genetic fluorescence-based labeling, commonly used to track protein expression. In some cases, genetic methods underreported or overreported protein presence, suggesting that antibody labeling provides a more immediate and accurate snapshot.
Beyond neuroscience, this technique could improve protein profiling in pathology, developmental biology, and regenerative medicine.
Read also:
AI Foundation Models in Biotech: New Paradigm
14 Foundation Models for Biology Research and Chemistry