'Google Maps' of Human Cells
ALSO: Word Models in Biotech — the Next Frontier; Antibody Design Without Binders, Transforming Clinical Trial Recruitment, and more...
Hi! This is our weekly newsletter, ‘Where Tech Meets Bio,’ where we talk about technologies, breakthroughs, and great companies moving the biopharma and medtech industries forward.
If you've received it, then you either subscribed or someone forwarded it to you. If that is the case, subscribe by clicking this button:
Let’s get to this week’s topics!
Brief Insights
🔬 Recursion finalizes its merger with Exscientia, creating a unified AI-driven drug discovery world model that combines Recursion's 60+ petabytes of data and machine learning with Exscientia's chemical design, advancing over 10 clinical/preclinical programs and targeting $20B in potential milestone payments.
🔬 Insilico Medicine receives FDA IND clearance for ISM5939, an AI-designed ENPP1 inhibitor for solid tumors, marking its 10th AI-driven molecule cleared for clinical trials, showcasing rapid development using Pharma.AI.
🔬 Formation Bio, Sanofi, and OpenAI launch Muse, an AI platform to streamline patient recruitment for clinical trials, cutting timelines by analyzing literature and real-world data, with Sanofi set to use it in Phase 3 multiple sclerosis trials.
🔬 Nabla Bio's AI platform JAM demonstrates success in designing functional de novo antibodies for challenging targets like GPCRs and viral proteins; with $550M+ partnerships with AstraZeneca, Bristol Myers Squibb, and Takeda, and funding from Radical Ventures to advance computational and experimental antibody discovery.
🔬 Vall d’Hebron Institute's next-gen CAR T-cell therapy achieves complete tumor regression in preclinical HER2-positive cancer models; a Phase 1 trial for 15 advanced cancer patients is set to begin, testing its safety and long-term effectiveness.
💰 Novartis acquires Kate Therapeutics for up to $1.1 billion, integrating its AAV-based gene therapy platforms and preclinical candidates for inherited neuromuscular diseases.
💰 Novartis inks a $745M licensing deal with Ratio Therapeutics for global rights to a radiopharmaceutical cancer therapy targeting SSTR2, leveraging Ratio’s radioligand technology for therapeutic and imaging applications, with Novartis leading development and commercialization.
🚀 Trogenix launches from stealth with $15M+ funding, unveiling its Odysseus platform powered by Synthetic Super-Enhancers to target aggressive cancers like glioblastoma; preclinical studies show curative responses and persistent immunity, with Phase 1/2 trials set for 2025 and plans for five INDs within 5 years.
🔬 Lantern Pharma doses its first patient in Japan as part of the Phase 2 Harmonic trial, expanding to Japan and Taiwan to target never-smokers with NSCLC, a subgroup 2–3 times more prevalent in Asia.
💰 Enveda Biosciences raises $130M in Series C, boosting its $360M total to advance AI-driven, nature-derived medicines, including a pipeline of 10 candidates and ENV-294, now in Phase 1 trials for inflammatory conditions.
💰 Converge Bio raises $5.5M to develop a unified platform for biotech LLMs, enhancing tools for antibody research and molecular design, while also planning to train its own foundational model to advance biotech-specific AI applications.
🔬 NVIDIA releases the open-source BioNeMo Framework, enabling faster AI-driven drug discovery with tools like AlphaFold2 and DiffDock 2.0, supporting scalable biomolecular model training with partners like Argonne and Genentech.
🔬 Owkin and Proscia partner to integrate Owkin’s MSIntuit CRC v2 AI diagnostic into Proscia’s Concentriq platform, streamlining MSI testing for colorectal cancer with 95% sensitivity.
🔬 Neuralink receives Health Canada approval for a clinical trial of its brain-computer interface, testing the N1 Implant and R1 Robot in individuals with quadriplegia or ALS to enable device control via neural signals, with an 18-month primary phase and long-term follow-up.
💰 Cytomos raises £5M to scale production of Celledonia, a benchtop analyzer for label-free, real-time single-cell analysis, aiming to cut biologics development timelines by six months and expand into the US market.
💰 LongeVC invests in Glyphic Biotechnologies, backing its single-molecule protein sequencing technology to analyze individual proteins in complex mixtures, with applications in drug discovery, diagnostics, and the growing $30B proteomics market.
🔬 Advarra’s 2024 survey highlights persistent barriers in clinical trials: communication gaps between sites and sponsors, severe staffing shortages with limited trained personnel, and technology tools often failing to deliver promised efficiencies.
The Human Cell Atlas Is Here
A milestone in systems biology!
The Human Cell Atlas (HCA), one of the most ambitious global collaborations in biotechnology since the Human Genome Project, has delivered its first draft atlas. This remarkable effort, involving 3,600 researchers from 102 countries, aims to create comprehensive reference maps of every human cell type, providing a foundation for understanding health and disease.
As HCA co-chair Aviv Regev put it during the 2023 HCA General Meeting, the project is about "creating comprehensive reference maps of the types and properties of all human cells as a way—a basis—for understanding, diagnosing, monitoring, and treating in health and disease."
So, this obviously isn’t just another dataset; it’s a blueprint of human biology, painstakingly assembled cell by cell, tissue by tissue, organ by organ:
"Thanks to multiomics, you can measure many features of human cells simultaneously, from chromatin organization and protein levels to metabolic profiles and dynamic events like transcription and translation rates. On top of that, to understand cells truly, we have to know where they are located—both in a common coordinate framework and in relation to each other within tissues. This is what we call the spatial branch of the atlas. Finally, for this to be a Human Cell Atlas, it must reflect human diversity—in ancestry, gender, age, geography, and more. As these additional views grow, we have an opportunity not just to integrate them, but to build a unified representation and understanding of what it actually means to be a cell."
The initiative catalogs over 62 million cells from 9,100 donors, representing every stage of human development, from embryonic to adult. The HCA focuses on 18 Biological Networks, with some detailed maps already available. These atlases harmonize cutting-edge techniques—single-cell RNA sequencing, spatial transcriptomics, and multi-omics—to offer an unprecedented view of cellular landscapes.
Notable highlights include:
Linking 100,000+ genetic variants to cellular functions and diseases.
Leveraging AI tools like GeneFormer and scGPT to accelerate discoveries.
Harmonizing data across systems to connect genotype to phenotype.
The first draft features integrated atlases for the lung, nervous system, and eye, offering diverse insights. For example, the lung atlas, published recently, uses transfer learning and harmonized ontologies (standardization of how biological data is labeled and organized across datasets) to reveal how cell profiles vary by factors like age, gender, and BMI.
The HCA opens transformative possibilities:
Target Discovery: Pinpoint disease-specific cell types and biomarkers for precision therapies.
Personalized Medicine: Tailor treatments to genetic and environmental variations in diverse populations.
Disease Models: Build realistic organoids for accurate drug testing.
Safer Drugs: Predict tissue-specific drug metabolism to reduce adverse effects.
Early Diagnosis: Detect subtle gene expression changes to catch diseases like cancer earlier.
AI Insights: Machine learning tools accelerate integration and refine predictions, with frameworks evolving to handle the project’s vast scale.
The atlas is built on collaboration and diversity, with dedicated Biological Networks ensuring representation across geographic, genetic, and ancestral lines.
In a recent article "The Human Cell Atlas: From a Cell Census to a Unified Foundation Model," (Rood & Regev et al.) discuss how cell atlases are evolving into powerful tools. They argue that these resources are transitioning from being simple cell censuses to becoming unified foundation models.
👉 Explore the atlas and its tools here.
Word Models in Biotech — the Next Frontier?
On November 20, 2024, Recursion and Exscientia finalized their merger, combining Recursion’s Recursion OS—an experimental-computational platform—with Exscientia’s AI-driven molecular design expertise, aiming to streamline drug discovery by integrating experimental and computational tools through an iterative system to accelerate the process.
The merger brings together a pipeline of over 10 clinical and preclinical programs, alongside discovery projects and partnerships, which have collectively generated $450 million in milestone payments, with potential for further growth.
What stood out in the announcement was a hint of a vision for developing virtual cells to simulate biological processes, framed within a two-part paradigm where an AI world model works in tandem with the real world.
What’s a world model, after all?
Recently, we discussed world models and how they are applied in biological context. This article gives an overview with examples, but to summarize: a world model is an AI system that builds a dynamic internal representation of its environment and predicts how actions might unfold, allowing forecasting and decision-making within complex systems.
Meanwhile, Large Language Models (LLMs) excel at predicting patterns in sequential data, such as language or other structured inputs, by identifying correlations from vast datasets. However, unlike world models, LLMs do not create dynamic internal simulations of their environment or model causal relationships that unfold over time.
Some of the foundational ideas behind world models come from researchers like Jürgen Schmidhuber & David Ha, and Yann LeCun, these initial WMs were primarily visual; there are parallels with control theory, which deals with designing systems that use feedback to guide actions toward a desired state, commonly applied in areas like robotics, aerospace, and industrial automation.
In one of the previous newsletters, we’ve already discussed a biology-specific world model, NOETIK’s OCTO, which simulates cancer biology by creating a unified, multimodal representation of tumor biology.
It integrates spatially aligned data such as gene expression, protein measurements, DNA sequencing, and structural tissue markers to capture the complexity of cancer at multiple biological scales. This representation allows OCTO to predict patient-specific responses to treatments, modeling counterfactual scenarios (what if we could virtually tweak a gene to see how it affects tumor growth?) about hypothetical biological interventions.
The model simulates these perturbations internally, enabling researchers to explore therapeutic outcomes without relying on exhaustive trial-and-error experiments.
Now, Recursion is also building a world model with the ambition to understand biology at a systems level.
The Recursion press release mentioned plans for Recursion OS to continue to “…drive iterative loops of hypotheses and active learning…” and eventually develop virtual cells to run in silico clinical trials.
Conceptually, this is a big step, as virtual cells could serve as a new domain-specific world model. While the exact details of this vision remain unclear, they may well leverage the mentioned 60 petabytes of proprietary data and integrate it with chemical design and synthesis capabilities from Exscientia, potentially allowing Recursion to drastically reduce the need for early-stage physical experiments.
In many ways, Recursion OS approaches the structure of a proper world model, but based on the brief announcement, it’s unclear how far along Recursion is in building these capabilities. So for now, we can only wait and see how this fares.
Antibody Design Without Binders
Fresh off a $26 million Series A financing round in May 2024, Nabla Bio has unveiled an interesting advance in computational antibody design with its AI platform, Joint Atomic Modeling (JAM).
Building on foundational work from George Church's lab at Harvard—where co-founder Surge Biswas helped pioneer protein language models like UniRep—JAM enables de novo antibody creation starting from target protein sequences or structures, without relying on known binders.
That means it doesn’t need to know what’s worked before. It can take the sequence or structure of a target protein, make up entirely new antibodies, and, more often than not, those designs bind well, function effectively, and look promising for further development.
This process allows Nabla to address some very challenging targets in drug discovery, such as multipass membrane proteins like GPCRs and Claudin-4, which are traditionally difficult to drug due to their complex structures and the challenges of recombinant protein production.
JAM operates by designing antibodies entirely computationally, leveraging libraries as small as 1,000 to 100,000 candidates to achieve high hit rates. The resulting antibodies demonstrate strong affinities (in the nanomolar range), high epitope specificity, and early-stage developability metrics, including low polyreactivity, good expression levels, and monomericity (for those who could read these instinctively, please refer to their technical paper)—properties critical for advancing potential leads toward therapeutic applications without extensive optimization.
Nabla reports JAM designed antibodies against Claudin-4 and GPCR CXCR7, showing high-affinity binding and functionality, including on-cell binding and SARS-CoV-2 neutralization.
For a deeper dive into Nabla's methods and findings, check out their blog post.
Transforming Clinical Trial Recruitment
Building on the May 2024 collaboration between Sanofi, Formation Bio, and OpenAI, the trio has introduced Muse—an AI-powered platform addressing a critical bottleneck in clinical trials: patient recruitment. With fewer than 10% of eligible patients enrolling in trials—often due to lack of awareness—delays in recruitment contribute to a slow progress in bringing therapies to patients.
Muse is designed to simplify and speed up this process by cutting timelines from months to minutes. Here’s how:
Data Analysis: It synthesizes disease-specific research, real-world evidence, and competitive insights to identify optimal patient profiles.
Diverse Strategies: Muse develops recruitment plans aimed at improving diversity and representation, addressing long-standing disparities in trial participation.
Content Automation: The platform auto-generates pre-screening questionnaires and outreach materials tailored for different subgroups, languages, and channels.
Compliance Built-In: An embedded large language model ensures that materials align with regulatory and Institutional Review Board (IRB) standards.
Sanofi plans to deploy Muse in its Phase 3 multiple sclerosis trials, leveraging its expertise and data to optimize the platform. Formation Bio will support broader applications across therapeutic areas, combining its tech-driven approach with OpenAI's cutting-edge AI models.
Read also:
11 Biopharma Trends to Watch in 2024