BioByte 112: improvements in artificial wombs, motivational pathways of cachexia in cancer, introduction of AgentRxiv, and efficiency increases for de novo enzyme design

Amee Kapadia

Pablo Lubroth

Mikaela Kimpton

, and 3 others

Apr 17, 2025

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.

🚨 [Last week to apply - applications close 24th April] We're looking to expand the Decoding Bio team with Product Managers to transition Your Missing Guide to Breakthrough Biotech into a media empire. If you're interested in envisioning and implementing the future of Decoding Bio, fill out the quick form linked to the button below! 🚨

Apply to Join

What we read

Blogs

Womb for improvement [Aria Babu, The Works in Progress Newsletter, March 2025]

“While for most women pregnancy is a miraculous experience that gives them a unique bond with their child, having children can be challenging. Some women face serious risks during pregnancy. For others it is not even an option.”

The risks of pregnancy are numerous: from acute illness and pain like morning sickness and sleeping issues, to the lack of clarity of whether medication that mothers were taking before pregnancy will be safe during pregnancy, to gestational diabetes, pre-eclampsia to vaginal tearing, the risk of blood clots, haemorrhage, sepsis from vaginal birth and infection risk from caesareans, to post-partum depression and psychosis.

For many of the women and couples that do not have the choice to undergo this process or risks are too great, artificial wombs could make it possible to grow a baby from conception to birth outside the mother’s body.

This could be life changing:

According to the CDC, 6% of married women (aged 15-44) in the US are infertile and 12% have impaired fecundity.
IVF success rates are still low at 45-53% chance of a successful pregnancy after undergoing three cycles.
In the US, one third of infant deaths are related to premature births
Babies at risk from maternal illnesses, such as diabetes, could be protected

The article highlights how women already spend parts of the gestation period not pregnant: zygotes created via IVF can be kept outside the body for 2-3 days before implantation, and babies born earlier than 24 weeks can survive in incubators. Artificial wombs would fill in this gap.

Whilst we’re still far off from a full artificial womb, there have been some recent advancements:

“Scientists have been able to grow embryos in the lab for thirteen days
US researchers have built a synthetic amniotic sac and umbilical cord that can keep lamb foetuses alive at eighteen weeks
In Israel, some researchers have managed to grow mice in an artificial womb for twelve days”

Artificial wombs are a polarizing idea, which is not a surprise given the resistance to new reproductive technology from the pill to IVF. In order to overcome resistance to new reproductive technology, the article mentions increased investment into blue-sky risky projects beyond egg freezing and IVF, results driven IVF model to fund new fertility startups, gathering more data from fertility clinics and monitoring hormone levels and extending the “fourteen-day rule” to study how embryos develop to 21-28 days.

A Benchmarking Crisis in Biomedical Machine Learning [Faisal Mahmood, Nature Medicine, April 2025]

As foundation models and multimodal systems begin to reshape drug discovery and clinical decision-making, the biomedical ML field is confronting a critical challenge: the lack of standardized benchmarks.

Unlike other domains such as computer vision or NLP—which benefit from widely accepted public datasets—biomedical research often relies on proprietary data, inconsistent preprocessing, and context-specific evaluation criteria. This fragmentation makes it difficult to compare model performance, assess generalizability, or tie outcomes to real-world clinical value.

In a recent Nature Medicine commentary, Faisal Mahmood calls for a coordinated response: consortium-led efforts to create shared reference datasets, enforce transparency in preprocessing, and align evaluation frameworks with clinically meaningful endpoints. He also underscores the need to address the ethical, regulatory, and logistical hurdles of data sharing, while preparing a workforce fluent in both computational methods and biomedical workflows.

Without common standards, biomedical AI risks becoming siloed—where progress is hard to measure, replicate, or trust. Mahmood’s piece is a timely call to action for infrastructure that can keep pace with the field’s ambitions.

Papers

Atom level enzyme active site scaffolding using RFdiffusion2 [Ahern et al., bioRxiv, April 2025]

De novo enzyme design has emerged as a major challenging frontier in protein engineering. Unlike general scaffold design, enzyme engineering demands precise placement of catalytic residues to stabilize transition states and mediate specific bond‐forming or bond‐breaking events. These idealized active‐site geometries—termed “theozymes”—describe the constellation of side‐chain functional groups and cofactors around a reaction’s transition state. However, traditional methods of motif scaffolding require enumerating backbone indices and rotamer libraries for each catalytic residue, limiting both efficiency and the diversity of designs, and have also been limited by existing limitations to co-model proteins with ligands.

To overcome these bottlenecks, the authors developed RFdiffusion2 by retraining the RosettaFold All‑Atom diffusion model (RFdiffusionAA) with creative motif scaffolding tasks. The new framework represents some residues at full atomic resolution, allowing the network to infer side‑chain conformations (“rotamers”) and sequence positions during generation. Additionally, RFdiffusion2 incorporates ligand‐burial controls by conditioning on Relative Accessible Surface Area labels for each ligand atom, and uses flow‐matching for stable, all‑atom diffusion training. These enhancements enable seamless scaffolding of side‑chain and ligand motifs in a single generative pass, given only atomic coordinates of scaffolding residues.

When benchmarked on a newly curated Atomic Motif Enzyme (AME) set of 41 diverse active sites, RFdiffusion2 succeeded in scaffolding all 41 cases, a dramatic improvement over the 16/41 successes of the original RFdiffusion. They experimentally tested designs of enzymes with theozymes sourced either from existing crystal structures of enzymes or theozymes derived from Density Functional Theory. For each of the 5 de novo enzyme design campaigns, they validated more than one functional enzyme in 96 designs.

By dramatically expanding the range and efficiency of de novo enzyme scaffolding, RFdiffusion2 paves the way for large‐scale, atomically precise enzyme design and broader applications in ligand‐binding protein engineering.

AgentRxiv: Towards Collaborative Autonomous Research [Schmidgall & Moor, AgentRxiv, 2025]

The landscape of artificial intelligence is shifting from isolated, task-specific agents to networks of interoperable systems working collaboratively. While traditional AI agents were designed to operate independently, the real potential lies in their ability to coordinate and leverage collective intelligence to solve complex problems more efficiently.

Recent initiatives reflect this evolution. Google’s Agent2Agent (A2A) protocol, for instance, introduces an open standard that enables diverse AI agents, regardless of vendor or framework, to communicate securely, share information, and coordinate actions. Backed by companies like Atlassian, Box, Cohere, Salesforce, and consulting firms including Accenture and McKinsey, A2A represents a significant step toward cross-platform agent interoperability. Beyond a technical upgrade, this standard offers strategic value by streamlining workflows, boosting productivity, and lowering operational costs.

Complementing this, the blueprint architecture proposed by Kandogan et al. at Megagon Labs outlines a modular framework for orchestrating compound AI systems. It connects various agents such as APIs, databases, and predictive models through components such as registries, data planners, and task coordinators. At its core is the concept of streams, which direct data and instructions across agents to ensure coordinated, transparent task execution.

Meanwhile, AgentRxiv, developed by Schmidgall and Moor from Johns Hopkins and ETH Zurich, extends multi-agent collaboration into scientific research. The platform enables AI-driven labs to share and build upon findings, accelerating discovery in a way that mirrors the collaborative nature of human science.

Together, these efforts mark a clear industry shift toward interoperable multi-agent ecosystems. Rather than relying on isolated tools, the future of AI lies in dynamic networks of agents working in concert. This model not only improves efficiency and scalability but also opens the door to deeper innovation and more effective problem-solving across sectors.

A neuroimmune circuit mediates cancer cachexia-associated apathy [Zhu et al., Science, April 2025]

Cachexia, a syndrome implicated in many chronic diseases including advanced stages of cancer, is characterized by severe wasting as well as certain neurocognitive symptoms such as depression and apathy. Despite being a well-established phenomenon, little is known about the biological mechanisms underlying its manifestation. Prior findings have elucidated a linkage between elevated levels of proinflammatory cytokines such as IL-6, but the neural circuits driving this correlation are unknown.

Researchers at the Washington University School of Medicine investigated this relationship using a highly validated mouse cancer model. After injection with C26 adenocarcinoma cells, the mice in the study were subsequently measured for physical hallmarks of cachexia, such as loss of muscle mass. They were additionally subjected to a series of behavioral tasks designed to measure neurocognitive states around motivation, anhedonia, and despair. When compared to tumor-free control mice, the cachexia mice exhibited increased effort sensitivity but did not exhibit other forms of motivation loss, like despair and anhedonia. They also confirmed these observations were not due to anorexia with another control group subjected to reduced feeding.

In the cachexia mice, brain screens found significant alteration in discrete networks, notably a hotspot in the medial nucleus accumbens (mNAc) and marked inhibition in the ventral tegmental area (VTA), indicating implication of the mesolimbic dopaminergic system. This observation was further validated in a foraging task, which revealed progressive reduction in encoding of reward size—an indicator of mesolimbic dopamine functioning in NAc—that was correlated with weight loss metrics in cachexia progression. Reward-evoked dopamine neural activity in VTA also diminished accordingly.

After demonstrating these behaviors were mediated by IL-6 by rescuing the effort sensitivity through anti-IL-6-antibody treatments, optogenetics, and pharmacological interventions, Zhu et al. sought to understand the rest of the neural circuit. Viral techniques revealed GABAergic inhibitory impacts in spontaneous inhibitory postsynaptic currents in NAc and additionally led the researchers upstream to the substantia nigra pars reticulata (SNpr), which in turn is driven by inhibitory neurons in the parabrachial nucleus (PBN). A whole-brain c-Fos screen also found the area postrema (ArP) to be a key region upstream of the PBN, one of particular interest due to its role in interpreting cytokines.

These findings enabled the authors to propose a new neural circuit for effort sensitivity due to cachexia, one that starts as an inflammation response in the ArP, travels through the PBN and SNpr to suppress dopamine neurons in the VTA, and ultimately results in reduced dopamine levels in the NAc. Discovery of this new circuit establishes cachexia as neuronally distinct from anorexia, and indicates that cachexia affects a specific isolated behavioral dimension—effort sensitivity—as separate from anhedonia and despair. This study also has significant therapeutic applications, as existing drugs that affect IL-6 could be repurposed to remediate cachexia.

Scaling unlocks broader generation and deeper functional understanding of proteins [Bhatnagar et al., bioRxiv, April 2025]

Generative protein language models (PLMs) promise to revolutionize enzyme and protein design by learning the complex distribution of naturally occurring sequences, but until now it has been unclear how model scale and data composition influence the quality and diversity of generated proteins.

To address this, the authors introduce ProGen3, a family of sparse mixture‐of‐experts autoregressive PLMs spanning 112 M to 46 B parameters. They train this on their compiled Profluent Protein Atlas v1, a cleaned database of 3.4B full-length proteins derived from genomic and metagenomic sources containing 1.5 trillion tokens. They derive compute‐optimal scaling laws to determine the most effective trade‐off between model size and dataset size under a fixed computational budget, ultimately selecting a 46 B‐parameter model that outperforms predictions on the compute‐optimal frontier.

The team systematically evaluated the impact of scale both in silico and in the wet lab: they generated millions of sequences across scales, applied rigorous quality filters, and used split‐GFP assays to measure soluble expression for clusters of natural and out‐of‐distribution proteins. Although models of all scales had comparable rates of designing proteins that express, the 46B model consistently produced higher expressing proteins. They further demonstrated that larger ProGen3 models also respond more strongly to alignment with experimental fitness and stability datasets. Interestingly, they corroborate previous understanding from Weinstein et al. 2022 that larger models with better understanding of natural protein distribution are not necessarily better zero-shot fitness predictors, motivating the use of model alignment.

This work offers a thoughtful evaluation of how data quality, model scale, and alignment strategies jointly shape a pLM’s ability to generate diverse, out‑of‑distribution sequences and enhance its predictive performance, advancing the field of protein language modeling forward.

Notable deals

Glycomine has raised a $115M series C led by CTI Life Sciences Fund, funds managed by abrdn and Advent Life Sciences. The funding will be used to advance Glycomine’s lead asset, GLM101 into a Phase 2b clinical trial treating phosphomannomutase-2 congenital disorder of glycosylation (PMM2-CDG). Other investors included Novo Holdings, Sanofi Ventures, Abingworth, RiverVest Venture Partners, Sanderling Ventures, Chiesi Ventures, Remiges Ventures and Asahi Kasei Ventures.
HepaRegeniX has raised €21.5 to complete its ongoing Phase 1b trial and to advance the Phase IIa clinical trial for HRX-215. HRX-215 is an orally available small molecule that selectively inhibits MKK4, a master regulator of liver regeneration. Wellington Partners joined the investor syndicate.
Attovia Therapeutics has raised a $90M series C led by Deep Track Capital. The proceeds from the financing will be used to advance Attovia’s two lead assets (ATTO-1310 and ATTO-3172) for the treatment of chronic pruritus and atopic dermatitis. Other investors included Vida Ventures, Sanofi Ventures, Mirae Asset Capital Life Science, Frazier Life Science, venBio, Goldman Sachs alternatives, Nextech Ventures, Cormorant Asset management, EcoR1 Capital, Marshall Wace and Illumina Ventures.
Imbria Pharmaceuticals has raised a $57.5M series B led by new investor Deep Track Capital. This funding will be used to advance ninerafaxstat through a Phase 2b clinical trial in non-obstructive hypertrophic cardiomyopathy (nHCM). Other investors include AN Ventures, Catalonia Capital Management, Cytokinetics, RA Capital and SV Health Investors.

In case you missed it

The Magic of Fast Feedback Loops [Stephen Malina, Asimov Press, April 2025]

FDA Announces Plan to Phase Out Animal Testing Requirement for Monoclonal Antibodies and Other Drugs [FDA, April 2025]
The FDA is making a push to remove animal testing for preclinical safety studies, suggesting instead the use of clinical data from other countries, in vitro models, or predictive AI models.