Weekly Tech+Bio Highlights #19
ALSO: Generative AI vs. Fatal Lung Disease; This Company Claims New “Gold Standard“ for Microscopy Data Analysis; Rethinking Gene Delivery with ML, 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
🔬 Insilico Medicine’s AI-designed drug, ISM001-055, shows promising Phase IIa results in treating idiopathic pulmonary fibrosis (IPF), demonstrating significant lung function improvement and potential to reverse lung damage in just 12 weeks.
💰 Tempus AI acquires Ambry Genetics from Konica Minolta for $600 million, enhancing its precision medicine capabilities with expanded genetic testing offerings in hereditary cancer, pediatrics, rare diseases, and more.
🔬 DeepMind releases AlphaFold3 as open source, advancing protein structure prediction by modeling complexes with biomolecules like DNA and drug targets, enabling high-accuracy predictions.
🔬 Recursion introduces MolE, a transformer-based model using molecular graphs for property prediction, achieving top performance in ADMET tasks and reducing reliance on experimental data through self-supervised pretraining on 842 million molecules.
💰 Schrödinger partners with Novartis in a $150M upfront deal, with potential for $2.3B in milestones, to scale computational drug discovery using Schrödinger's predictive modeling platform, accelerating Novartis's therapeutic pipeline.
🔬 Recursion launches OpenPhenom-S/16, a foundation model for microscopy data available on Google Cloud, enabling automated cellular analysis and advancing phenomics with superior performance over traditional tools like CellProfiler.
💰 Genesis Therapeutics partners with NVIDIA to enhance its AI platform, GEMS, targeting undruggable diseases using advanced neural networks and molecular modeling, supported by additional NVentures funding following a $200M Series B round.
🔬 Enveda advances ENV-294, an AI-discovered, nature-derived drug for atopic dermatitis, into Phase 1 trials, leveraging its PRISM foundation model and the world’s largest natural product library to accelerate therapeutic development.
💰 Immunai partners with Teva Pharmaceuticals to enhance oncology and immunology clinical trials, leveraging its AMICA immune cell atlas and AI-driven Immunodynamics Engine to optimize dose selection, biomarkers, and patient subgroup analysis.
🔬 Corti and BigHand partner to deploy an AI tool to 40,000 NHS healthcare workers, reducing administrative tasks by 80% and enabling staff to focus on patient care, as NHS aims to meet growing appointment demands.
🔬 Medable adds generative AI to its Studio platform, reducing eCOA generation time from days to seconds and accelerating digital and decentralized clinical trial workflows, enhancing efficiency and cost savings for sponsors and CROs.
🔬 Ansa Biotechnologies unveils a sequence-agnostic DNA synthesis platform capable of producing error-free clonal DNA constructs up to 5 kb, overcoming traditional synthesis challenges with advanced assembly and long-read sequencing quality control.
AlphaFold3 Goes Open Source
DeepMind and Isomorphic Labs have released AlphaFold3, expanding beyond AlphaFold2's single-protein modeling to include interactions with other biomolecules, such as nucleic acids, drug-like molecules, and DNA.
AlphaFold3, building upon it's predecessor’s architecture, introduces a diffusion-based generative framework that enhances predictive accuracy by over 20% for protein-ligand interactions and around 15% for nucleic acid interactions, opening new possibilities for exploring protein behavior under physiological and drug-binding conditions.
With the code now available on GitHub, for non-commercial use, researchers can run the model locally, which allows for more customized and expansive investigations. However, certain limitations remain, particularly regarding the non-commercial use policy and restrictions on training derived models.
Key improvements include:
Enhanced Accuracy: Consistently predicts protein structures with median root-mean-square deviation (RMSD) often below 1.6 Å, closely matching experimental results.
Increased Speed: Optimizations allow predictions within hours, reducing computational time by 50-75% compared to traditional methods.
Broader Scope: Trained on a dataset of over 170,000 protein structures and millions of sequence alignments, it generalizes across diverse molecular configurations, including interactions with RNA, DNA, and small molecules.
AlphaFold3’s release follows months of contention. Initially made available in May 2024 as a restricted-access tool, it prompted criticism for offering only pseudocode and a web server with significant constraints—such as limited daily quotas, non-commercial use restrictions, and prohibitions on training derived models. The current release, although more open, still upholds non-commercial use requirements, limiting its application within the commercial sector.
Generative AI vs. Fatal Lung Disease
Insilico Medicine has shared promising Phase IIa trial results for ISM001-055, an experimental drug developed using generative AI to treat idiopathic pulmonary fibrosis (IPF).
Leveraging their AI platform, Pharma.AI, Insilico Medicine utilized ML algorithms to analyze biological data related to IPF. They identified TNIK (Traf2- and NCK-interacting kinase) as a novel therapeutic target implicated in the disease's progression.
TNIK is a kinase involved in signaling pathways that regulate cell proliferation, differentiation, and survival—processes that, when dysregulated, contribute to pathological fibrosis in the lungs. ISM001-055 was designed to inhibit TNIK, with aim to interrupt aberrant signaling pathways and reduce the excessive scarring and tissue stiffening characteristic of IPF.
Now, results: the 12-week clinical trial demonstrated improvements in lung function and quality of life for patients, indicating the drug's potential to not only slow but possibly reverse this progressive disease that affects nearly 5 million people worldwide and currently lacks treatments capable of reversal.
The trial enrolled 71 patients across 21 sites in China, who received either a placebo or varying doses of ISM001-055.
Key findings:
Lung Function Improvement: Patients receiving the highest dose experienced a 98.4 mL increase in forced vital capacity (FVC), while the placebo group saw a decline of 62.3 mL.
Quality of Life Gains: Those on the high dose reported a 2-point improvement in the Leicester Cough Questionnaire score.
Safety and Tolerability: The drug was generally well-tolerated, with mild side effects such as gastrointestinal issues and minor liver function abnormalities.
Following these positive outcomes, Insilico is preparing to engage with regulatory agencies to plan a pivotal Phase III trial involving a larger patient population. They also intend to present the full data at medical conferences and publish their findings in a peer-reviewed journal. To oversee the progression of ISM001-055, the company has appointed Dr. Carol Satler as Vice President of Clinical Development.
Rethinking Gene Delivery Stealthily
In a previous issue, we covered the Evolved 2024 Hackathon, an event focused on computational biology and AI. Hosted by EvolutionaryScale and Enveda Bio, it featured support from partners like OpenAI and AWS, providing resources for participants to develop innovative solutions. The first-place prize went to EvoCapsid, a team rethinking gene delivery from the ground up.
Recent setbacks in gene delivery systems highlight ongoing challenges in the field, exemplified by Verve Therapeutics voluntarily halting a trial in April 2024 due to abnormal liver enzyme levels caused by lipid nanoparticle delivery, and Pfizer discontinuing an adeno-associated virus-based gene therapy in June 2024, after a patient fatality linked to the delivery agent.
The work of EvoCapsid’s team began with a question: what if machine learning could engineer a gene delivery vehicle that functions in the body without triggering immunity? They explored whether it was possible to recapitulate evolutionary history using foundational AI models such as ESM3 and AlphaFold to redesign viral proteins to look like human proteins for stealth gene delivery.
Their platform combines capsid-building tools with a flexible workflow to evolve gene delivery systems. This includes tailoring carriers, cargo, and targeting moieties with potential to address complex diseases such as cancer, muscle degenerative disorders, and Severe Combined Immunodeficiency (SCID) by targeting difficult cell populations like immune stem cells—a capability already being deployed in collaboration with Dr. Donald B. Kohn at UCLA.
To achieve this, EvoCapsid employs three specialized modalities:
EvoEnvelope: Engineers envelope proteins to deliver genetic material to human cells while evading immune detection.
EvoGag: Develops endogenous gag components compatible with human biology, enabling stable and efficient carriers.
EvoCargo: Optimizes CRISPR-Cas editors or non-CRISPR-Cas editors for RNA binding, enhanced specificity, and reduced immunogenicity.
As EvoCapsid transitions from a hackathon project to a startup, the team is committed to open-source development. They plan to release a beta GUI for their platform, followed by a full v1.0 version this winter, along with a preprint documenting their findings. They also aim to collaborate widely, inviting researchers and organizations to join in evolving gene delivery technologies.
For those interested, check out their teaser website and feel free to reach out!
This Company Claims New “Gold Standard“ for Microscopy Data Analysis
A new AI foundation model in bio is claimed to set the new "gold standard" for microscopy data analysis, outperforming CellProfiler.
Now, Utah-based Recursion launched a new model, OpenPhenom, which is a non-commercial, publicly available foundation model built on microscopy data that the company believes has set the new "gold standard" outperforming CellProfiler.
Large-scale cell microscopy assays are crucial for discovering new biological processes, drug candidates, and potential targets.
Traditional software for analyzing these assays, however, requires constant fine-tuning for each new experiment and can easily break down if assay conditions change.
As massive datasets like RxRx3 and JUMP-CP grow, these outdated tools struggle to keep up with the data demands.
OpenPhenom-S/16, designed by Recursion, offers a powerful alternative for non-commercial researchers. Company says it can replace old workflows with an adaptable, ready-to-use model that consistently beats traditional analysis methods, with no extra tuning or training needed.
For example, OpenPhenom-S/16 significantly enhances recall of known biological relationships in the JUMP-CP cpg0016 dataset, as cross-verified with the StringDB database.
Notably, Recursion also has models for commercial use for its own pipeline and partnership programs. Recursion has featured the performance of two generations of proprietary models trained on Recursion’s internal microscopy, Phenom-1 (introduced at CVPR 2024) and Phenom-2 (introduced at NeurIPS FM4S 2024) -- also see image below.
Recursion proprietary models are stronger than OpenPhenom (surprise!) but that does not diminish the value the public model has for the drug discovery community.
Interestingly, just yesterday we covered another foundation model by Recursion -- MolE, for molecular property prediction (link in comments). Undoubtedly, foundation models are on the rise in biotech!
We previously compiled a list of 19 companies pioneering AI foundation models in pharma and biotech, and two new arrivals by Recursion, and some other new foundation models on the market, prompted us to update the article.
Read also:
19 Life Science Marketplaces to Try in 2024