BioByte 134: ToolUniverse Democratizes AI Scientists, mBER Tackles Nanobody-Based Binder Design, Nephrobase Cell+ Comprehensively Models Kidney Biology, and the Release of PRISM for Neuron Tagging
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
E11 Bio Releases PRISM Technology for Self-Correcting Neuron Tracing [Payne et al., E11 Bio, September 2025]
Large-scale brain circuit mapping faces two persistent issues: reliable, long-range morphological tracing (light-microscopy pipelines routinely suffer from splits and merges of neurons, breaking continuity) and simultaneous readout of molecular identity at synapse-scale resolution. PRISM makes a substantial advance in resolving both by tagging neurons with combinatorial protein barcodes that can be read with expansion microscopy and iterative immunostaining. In a mouse hippocampal (CA2/3) pilot, the authors encoded neurons with stochastic combinations of 18 engineered protein ‘bits’ (implying 218 possible binary barcodes), expanded tissue ~5x to an effective voxel size of 35 x 35 x 80 nm, and read 23 targets (18 bits + 5 synaptic markers) over 7 staining rounds. The result: an estimated ~100K distinct barcodes in the imaged volume and an 8x increase in error-free automated tracing distance (contiguous neurite length reconstructed without splits and merges) compared to conventional grayscale imaging.
This gain stems from a workflow composed of three elements working in concert. Cell-filling GFP scaffolds carry short, antigenically-distinct peptide tags that produce dense, combinatorial labels when delivered stochastically and sparsely by AAV. Expansion microscopy physically separates processes, allowing antibody staining to resolve them on a standard fluorescence microscope. Finally, an image-processing stack involving denoising and embedding amplifies barcode separability, sharpens neuron segmentation boundaries, and reconnects fragments across damaged or missing sections. Barcode identities are also coupled with synaptic markers for molecular annotation of the wiring. Together, these steps significantly minimized common failure modes – illustrating how molecular augmentation can revamp connectomic pipelines.
PRISM offers an accessible method that demonstrably improves tracing in a focused volume, but scaling it to whole-brain connectomics will require addressing emergent engineering realities – including uniform viral delivery at scale, antibody stripping / reproducibility across many rounds, and improving imaging throughput. Nonetheless, PRISM opens the door to exciting questions. Can it be extended to denser, more heterogeneous circuits? Paired with lineage and transcriptomic readouts, can it reveal novel information about cell identity and function? Taken together, PRISM provides a promising lever for improving the fidelity and interpretability of large-scale neuronal reconstructions.
Papers
mBER: Controllable de novo antibody design with million-scale experimental screening [Swanson et al., bioRxiv, September 2025]
Why it matters: The first wave of protein design models demonstrated high effectiveness for designing small mini-binder proteins. Binders with antibody formats would be more desirable in therapeutic contexts because of their clinical precedent, pharmacokinetics, established manufacturing and regulatory pathways. However, antibody-based binders are still challenging to design de novo. To tackle this, researchers from Manifold Bio develop a model mBER to design nanobody-based binders, and test over 1 million of them against 145 extracellular targets, with significant on-design hit rates for 45% of targets and up to 38% hit-rate per-epitope. Along with unlocking a large library of potential tissue-specific biologic shuttles for therapeutic development, the team also released mBER as open source to enable the community to design antibody-based binders.
Manifold Bio aims to create nanobody “shuttles” that can deliver biologics to specific cell types and tissues. That is an unusually hard biology problem along several axes: designing nanobody binders to extracellular proteins is challenging, testing cell-type specificity in vivo at scale requires careful experimental design, and doing both together compounds the difficulty. In this work, they focus on the first - how to design VHH-format binders effectively in silico, and couple this with an ultra high-throughput phage-display based readout to ultimately develop a large library of potential cell-type specific nanobody binders. The first wave of protein design models worked well for designing small mini-binder proteins to targets of interest - but these models like RFdiffusion and BindCraft are structurally unconstrained. The open question that this project - along with the concurrent model Germinal from Stanford - tackles is how to imbue specific properties of interest in the design process to enable e.g. de novo nanobody design.
The approach mirrors Germinal from the Hie and Gao labs at Stanford but is tuned specifically for nanobodies. Both frameworks build on ColabDesign, the AlphaFold2-based toolkit that underpins BindCraft. Conceptually, ColabDesign “flips AlphaFold2 on its head”: after predicting a structure for a given sequence, it backpropagates gradients from AlphaFold2’s confidence metrics into sequence space, gradually optimizing a candidate binder. Germinal constrained this process by adding IgLM, a language model of immunoglobulin sequences, as a prior; Manifold’s method, mBER (Manifold Binder Engineering and Refinement), uses ESM2 to bias CDR loop sequences toward nanobody-like distributions. In practice, the AlphaFold-derived loss captures structural plausibility of the antigen–binder complex, while the ESM2 prior keeps the evolving sequences within the distribution of functional VHHs. mBER also incorporates structural templates from NanoBodyBuilder2 and truncated antigen epitopes to further stabilize design trajectories.
The team generated 1.15 million VHH binders across 436 human cell-surface proteins and experimentally screened 331,000 of them against 145 targets in multiplexed phage display assays. Of these, 65 proteins (45%) showed statistically significant enrichment of on-design hits by Fisher’s exact test. Per-epitope hit rates varied widely, with some epitopes yielding <1% and others reaching up to 38% after filtering for high ipTM scores. Across the screened subset, the assays recovered 2,207 on-design hits and 11,040 off-design hits. Crucially, the off-design interactions were not diffuse polyreactivity but typically specific binding to a single alternative target, underscoring that the pipeline generates real binders even if not always to the intended antigen. Compared with minibinder-oriented frameworks like BindCraft, mBER achieves similar or higher effective hit rates but in a clinically validated antibody scaffold. Like Germinal, it demonstrates that AlphaFold-backpropagation plus antibody priors is sufficient for experimentally validated antibody design without retraining folding models.
Democratizing AI scientists using ToolUniverse [Gao et al., arXiv, October 2025]
Why it matters: Fields such as omics have advanced through the integration of tools and standardized analyses (also discussed in last week’s Decoding Science newsletter). In this paper, the authors present ToolUniverse, an ecosystem for building AI scientists; infrastructure required to support AI scientists.
ToolUniverse combines LLMs and AI agents with an ecosystem of scientific tools, datasets and APIs in an ecosystem with more than 600 tools. To improve compatibility, it implements a ‘tool specification schema’ which standardizes tool definitions and an ‘AI-tool interaction protocol’ which governs how AI scientists issue tool requests and receive results, equating it to HTTP in internet communication. New tools can be registered and integrated without additional configuration.
The authors demonstrated the ecosystem’s potential by applying it to therapeutic discovery for hypercholesterolaemia. By connecting ToolUniverse with Gemini CLI, the AI scientist system begins with protein target identification (via literature mining, target profiling and tissue expression tools). It highlighted HMG-CoA reductase as the most promising candidate. It then accessed the DrugBank database and profiled existing drugs that target the protein; selecting lovastatin as the initial treatment to optimize due to its off-target effects. It then carried out in silico screening to evaluate existing drugs and small molecule libraries via structural analog retrieval from ChEMBL and used Boltz-2 and ADMET-AI for affinity and profiling respectively. It cross-checked the top candidates in patent-mining tools and selected prevastatin, which has a lower side effect profile than lovastatin, and a new, patented candidate molecule with higher affinity to the target and better PK/PD parameters.
Whilst far from suggesting entirely new chemistry or hypotheses, it is an important step towards a community-built ecosystem to expedite components of hypothesis generation.
Nephrobase Cell+: Multimodal Single-Cell Foundation Model for Decoding Kidney Biology [Li et al., arXiv, September 2025]
Why it matters: The rise of single-cell sequencing technologies has transformed our ability to probe the inner workings of cells in a variety of developmental and disease states. The tremendous amounts of data generated from these experiments have paved the way for machine learning approaches that aim to capture the biological context that underlies different cell states and make predictions about perturbations or unseen data. While prime examples of general single-cell foundation models like Geneformer and scFoundation have shown significant promise, it is unknown as to whether such approaches can extrapolate to organ-specific models that can faithfully model disease states across different cells and tissue types. In their latest work, Li et al. present Nephrobase Cell+, a multimodal organ-specific foundation model that “learn[s] robust representations of kidney cell identities, gene regulatory networks, and microenvironmental crosstalk.” This paper offers a new benchmark for organ-specific modeling approaches in comparison to generalist foundation models for use in computational disease modeling and further downstream tasks.
Effective organ-level disease modeling requires a deep understanding of two major obstacles - the broad range of molecular mechanisms driven by inherent cellular diversity, and the multifactorial nature of associated pathologies such as inflammation and fibrosis. Current methods rely on task-specific models that struggle to generalize across different datasets and predictive tasks (e.g. a method for dimensionality reduction and cell type clustering cannot easily be used for gene regulatory network modeling). Instead of finetuning existing foundation models, the authors of Nephrobase Cell+ train a new model from scratch on over 100 billion tokens and 39 million cells and nuclei, leveraging scRNA-seq, single-nucleus RNA-seq and spatial transcriptomics data. The combined dataset spanned humans, mice, brown rats, and wild boar samples, with a range of spatial and chromatin accessibility assay data to enable cross-assay representation learning. As for the model itself, Nephrobase Cell+ uses a transformer-based encoder-decoder architecture with a Mixture-of-Experts module to integrate the single-cell and spatial data. The model is trained to reconstruct an input cell-gene matrix with additional architecture elements to enforce minimum levels of dissimilarity between cell-embeddings and control for variation across assay types.
As a primary sanity check, Nephrobase Cell+’s representations and consequent dimensionality experiments were compared with benchmark methods like principal component analyses and autoencoders. The authors note that the model was able to produce more coarse clusters in accordance with known kidney cell types. For example, the Nephrobase Cell+ could separate between proximal tubule and think ascending limb cells whereas tools like Geneformer and scGPT group the two cell types into mixed clusters. Looking at the model’s ability to generalize across species, Nephrobase Cell+ showed the ability to segregate similar cell types across species boundaries in contrast to Geneformer and UCE embeddings that show significant species-level clustering. On zero-shot cell-type classification tasks, the model achieved promising performance with most misclassification errors being attributed to closely-related subtypes.
The final study of Nephobase Cell+’s capability revolved around an in silico perturbation test. After doubling the expression of target genes in kidney disease in a randomly sampled set of cells, the authors examined the impact on other genes (differential expression). Gene set enrichment analysis was used to “identify perturbed biological programs” of significance. Perturbations to CCL2 genes showed effects on immune recruitment and membrane transport, while perturbations to VCAM1 showed increased T-cell proliferation, vascular migration, and effects to cellular respiration pathways linked to endothelial remodeling and metabolism adaptation. The simulated activation of SOX4 resulted in significant changes to developmental and energy-related functions. Key findings indicated a major shift toward essential organizational processes, with pathways strongly enriched for organ creation (such as kidney and valve formation) and programmed cell death.
Overall, Nephrobase Cell+ demonstrates the potential of organ-specific modeling paradigms applied to disease modeling. The bespoke foundational model approach shows improvement over task-specific or generalist foundation models whose latent representation space is not as useful for learning tissue-level and finer cross cell-type differences. The authors point to the fixed gene feature space and expensive computational training footprint as the model’s main limitations, especially if one wishes to retrain to include more assay types and kidney conditions. Nonetheless, it will be interesting to see if other organ-specific modeling approaches follow and if they can be effectively used for downstream tasks like drug response prediction and patient-specific prognosis prediction from molecular signatures.
Notable deals
The Dublin, Ireland based biotech Aerska announced a $21M seed round, co-led by Age1, Backed VC, and Speedinvest. With the use of interference RNA (RNAi) drugs, the company seeks to address brain diseases by silencing the genes that cause them. The difficulty of transporting drugs—especially larger ones—across the blood-brain barrier spurred the company to also develop an antibody-oligo conjugate platform which utilizes “brain shuttles” to enable effective transport of their RNAi drugs to the brain. Other investors in the round included Ada Ventures, Blueyard, Kerna, Lingotto (Exor), Norrsken VC, PsyMed, and Saras.
DNA synthesis company Ansa Biotechnologies reigned in $45.2M with an additional $9.2M in secured commitments for their oversubscribed Series B. Ansa’s chemically mild approach to DNA synthesis enables them to access lengths and complexities of DNA sequences hitherto largely inaccessible. Facile synthesis of these sequences will greatly accelerate scientific research that relies on or would benefit from them. With the funds, the company is primed to further expand manufacturing capabilities and capacity. Cerberus Ventures led the round, with participation from both existing investors—including Altitude Life Science Ventures and Blue Water Life Science Advisors—and new investors such as AIM13, Black Opal Ventures, and Fall Line Capital.
Crystalys Therapeutics raised $205M for their Series A, seeking to use the funds for Phase 3 clinical trials of their lead asset, dotinurad, in the US and Europe. The drug is an oral URAT1 inhibitor that seeks to serve as a secondary treatment for gout, the most common form of inflammatory arthritis. Due to its successful efficacy and safety performance in prior clinical studies, it has already been approved for use in China, Japan, the Philippines, and Thailand. In Japan, dotinurad has already been shown to bring about positive responses in more than 1.2 million patients since its approval in 2020.
For their sixth biotech investment fund, Mission BioCapital secured $134M. Since their founding in 2009, Mission has backed 60+ life sciences startups, and boasts significant optionality in lab space offerings across the US and Europe via their partnerships with a host of established lab incubators. Though no official press release for Fund VI has as of yet been published, the firm historically targets early-stage biotechs for investment, such as March Biosciences, Tune Therapeutics, and Dren Bio. Mission previously raised $275M for Fund V in the height of the biotech frenzy in November 2021. Endpoints coverage cites the ongoing difficult atmosphere for biotech fundraising as postulation for the decline in amount raised, however the fund’s target has not been disclosed and reasoning behind the amount remains unclear.
What we listened to
What we liked on socials channels
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: @decodingbio.
 |
|