BioByte 101: de novo antibody design, measuring how fast your brain is aging, enhancing GWAS, commentary on trial results, and more
Welcome to Decoding Bio’s BioByte: each week our writing collective highlight notable news—from the latest scientific papers to the latest funding rounds—and everything in between. All in one place.
What we read
Blogs
De novo design of antibodies enabled by Joint Atomic Modeling [Nabla Bio, November 2024]
The team at Nabla has developed Joint Atomic Modeling (JAM), a new AI system that designs therapeutic antibodies computationally using just the target protein's structure or sequence. While AI has made strides in protein structure prediction, designing antibodies that could become drugs has remained challenging because of the multiple developability properties that need to be juggled. JAM demonstrates that from-scratch design is now possible, creating antibodies with good binding properties and drug-like characteristics in about 4-6 weeks.
The system has shown promising results across eight different proteins, including traditionally difficult membrane protein targets like the GPCR family. In testing, JAM successfully designed antibodies against both soluble proteins like SARS-CoV-2 and challenging membrane proteins like Claudin-4 and CXCR7. The antibodies demonstrated good specificity—for example, showing 100-fold selectivity for Claudin-4 over similar proteins.
An interesting technical finding was that allowing multiple rounds of computational design—allowing the system to “think”—improved results significantly. When combined with focused experimental testing, this approach led to substantial improvements in binding strength. The system works with both single-domain antibodies (VHHs) and full antibodies (mAbs), and can handle proteins not present in its training data.
While these antibodies still need optimization to reach final therapeutic standards, JAM represents a practical advance in computational drug design and the first of its kind for antibodies. The system's success with membrane proteins is particularly noteworthy, as these represent an important but challenging class of therapeutic targets largely gone undrugged. For those in the drug development world, JAM is worth keeping an eye on.
How fast is your brain aging? Proteins in blood offer clues [Miryam Naddaf, Nature News, December 2024]
A paper published in Nature Aging this week reported on a set of 13 proteins that seem to be correlated with a person’s brain age gap (BAG), which is the difference between how old their brain appears and their chronological age. Although preliminary, the identified proteins could be used as easily accessible biomarkers to discern potential neurological disorders and intervene as early as possible.
To calculate the BAG, the researchers first built an ML model that determined the age of almost five thousand brain scans from the UK Biobank. The model analyzed factors such as brain volume, surface area, and white matter in its calculation of brain age. Alongside those brain scans, the researchers also analyzed proteomics data, looking at 2,922 proteins to determine if the concentration of any of the proteins correlated with a person’s BAG. Of the 2,922 proteins, eight appeared to be correlated with faster brain aging, and five with slower brain aging. Brevican (BCAN)—a protein involved in learning and memory processes—is one such example where higher levels appeared to be related to slower brain aging.
In addition to this primary result, the team also found that there were distinct changes in the concentration of some proteins at ages 57, 70, and 78, where each of these chronological ages denoted a distinct phase of brain aging. The proteins that changed most dramatically were different based on age, from those linked to metabolism and mental health at 57, dementia and stroke at 70, and inflammation and immunity at 78.
The study is promising, but more follow-up experiments are necessary. One scientist pointed out that the samples were taken from the blood instead of the brain, which could falsely portray the true neurological impacts of these proteins. They also proposed the need for animal studies to better elucidate how fluctuations in these proteins affected the BAG. Additionally, most of the participants were of European descent, so greater diversity of the sample population is required for future studies.
Papers
Small-cohort GWAS discovery with AI over massive functional genomics knowledge graph (Huang et al., MedRxiv, Dec 2024]
A preprint out this week from a collaboration between GSK and labs at Stanford and CMU introduces KGWAS (Knowledge Graph GWAS), a novel geometric deep learning method that significantly improves the power of genome-wide association studies by integrating a functional genomics knowledge graph. GWAS is one of the most important methods for detecting genetic variants associated with disease, and is often used in drug discovery to identify and/or validate therapeutic targets (targets with genetic evidence tying them to disease are 2.6x more likely to succeed in clinical trials).
KGWAS leverages a network of 11 million molecular interactions between variants, genes, and gene programs to enhance the detection of disease-associated genetic variants, particularly valuable for rare and uncommon diseases where large patient cohorts are difficult to assemble. Through extensive simulations and replication experiments, the authors demonstrate that KGWAS can achieve the same statistical power with up to 2.67× fewer samples compared to conventional GWAS methods, while maintaining proper calibration and false discovery control. When tested on 554 uncommon diseases in the UK Biobank, this approach found nearly 50% more disease-associated genetic variants than standard methods, with even better performance (80% improvement) for rare diseases that have very few patients available for study. Beyond just finding genetic links to disease, KGWAS helps explain the biological mechanisms behind these associations by identifying relevant molecular networks and cell types that may be involved in the disease process. For example, when analyzing Alzheimer's disease, KGWAS highlighted specific biological pathways involved in fat metabolism and protein accumulation that have been independently validated by experimental studies. The authors have made their method freely available through a web interface.
The Virtual Lab: AI Agents Design New SARS-CoV-2 Nanobodies with Experimental Validation [Swanson et al., bioRxiv, November 2024]
Teams at Stanford and the Chan Zuckerberg Biohub designed a Virtual Lab, which consists of an LLM principal investigator agent which creates and guides a team of LLM agents with different scientific backgrounds (e.g. immunology, machine learning, computational biology). These agents interact through virtual meetings (see image above for the workflow), which tackle different phases of a research project.
The Virtual Lab was employed to tackle the open-ended scientific problem of designing new nanobodies that bind to the latest variant of SARS-CoV-2. The Lab also chose to incorporate ESM, AlphaFold-Multimer and Rossetta to mutate existing nanobodies that bind to the receptor binding domain of the spike protein of the original SARS-CoV-2 strain.
The team experimentally validates 92 nanobodies designed by the Virtual Lab. It found 90% of them were expressed and soluble, with two candidates showing unique binding profiles.
Notable deals
Chroma Medicine and Nvelop Therapeutics Unite to Form nChroma Bio, Securing $75 Million to Accelerate Genetic Medicines. Chroma is hastily pushing its lead asset CRMA-1001 (Hepatitis B) to the clinic in a race with Tune Therapeutics. The additional capital will help in this endeavour whilst securing it novel delivery vehicle technology from nvelop to backfill its pipeline.
GSK has partnered with UK startup Relation Therapeutics to explore multiple targets for fibrosis and osteoarthritis. The deal includes $45 million upfront ($15 million equity investment) and up to $63 million in milestones, plus average biobucks of $200 million per target and tiered royalties. Relation, backed by investors like Nvidia, Deerfield, and Khosla, leverages machine learning, genomics, and multi-omics to deepen understanding of human biology and disease. Their platform includes a bone atlas, built from sequencing nearly 300 hip samples, and will expand into a skin atlas for fibrosis, starting with scleroderma.
Tasca Therapeutics launches with $52M Series A to drug auto-palmitoylation sites on proteins. The Boston biotech hopes drugging such sites will be effective in various cancers and plans to enter the clinic with its first asset next summer. Tasca was the first investment from new VC firm Cure Ventures.
Dimension raises $500M in second fund to invest at the intersection of bio and tech. The SF/NYC-based firm focuses on investing in founders pushing boundaries across hardware, software, and traditional biotech, with a strong emphasis on harnessing machine learning to drive transformative advancements. Notable investments include Enveda Biosciences, Chai Discovery, Monte Rosa Therapeutics, and Kimia Therapeutics.
BioAge Labs is halting its Phase 2 trial of the anti-obesity drug azelaprag due to observations of liver transaminitis in patients, leading to a 72% drop in its stock post-market. While the setback dims the company’s prospects in the next-generation obesity drug field, BioAge plans to refocus on advancing its NLRP3 inhibitor program and other metabolic aging therapies.
What we liked & listened to on socials channels
Shelby’s thoughts on some trial results this week—Relay, Recursion/Exscientia, Bioage [@ShelbyNewsad]
Field Trip
Did we miss anything? Would you like to contribute to Decoding Bio by writing a guest post? Drop us a note here or chat with us on Twitter: @ameekapadia @ketanyerneni @morgancheatham @pablolubroth @patricksmalone