BioByte 152: The State of Biosecurity, a More Sensitive Alternative to CAR T-Cells, Immune Modulation for Drug Clearance, Latent-Y for end-to-end biologics discovery, and Programming RNA Condensates
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
Reasons to be pessimistic (and optimistic) on the future of biosecurity [Abhishaike Mahajan, Owl Posting, March 2026]
In another characteristically brilliant piece, Owl Posting neatly summarizes the current state of biosecurity—a particularly relevant topic given rising geopolitical tensions and conflict, and one that the market has evidently deigned more attention to of late. With VCs backing breakout biosecurity companies like Valthos, Red Queen Bio, and Aclid, it is clear there is a significant concern regarding the potential for bioweapon development, especially as AI continues to grow more capable.
Despite the capital flowing into these companies, Mahajan argues that the primary customer, governmental agencies, will only sufficiently fund the industry in light of another catalyzing event (i.e. a pandemic as was seen with SARS-CoV-2). DNA synthesis companies are proposed as another potential customer, given scaling bioweapons requires DNA synthesis at scale, but this argument relies on three faulty assumptions: 1) sequences will be long enough to be screened, 2) sequences for bioweapons will be related to known pathogens, and 3) DNA synthesis companies are a necessary step to making these bioweapons at scale. To the first point, it has already been shown by a Canadian lab in 2018 that pathogens can be assembled from small DNA fragments. Additionally, continued improvements in AI will allow for the creation of de novo pathogens, while existing pathogens can also be made invisible to screening tools. Furthermore, official channels can be sidestepped by purchasing personal DNA synthesis equipment, reducing the cost and barriers to entry.
Government crackdowns which focus on the resources needed for manufacturing are one way in which these threats can be mitigated, as was done with meth in the early 2000s, but the legislation to do so is severely lagging, partially due to the dual-use nature of most equipment where the benevolent use cases outweigh the malevolent ones. The conclusion Mahajan draws from this is that there exists a clear, difficult to regulate path by which rogue actors could relatively easily design and procure novel bioweapons in the very near future.
Under the assumption that pathogens are accessible, both human and agricultural bioweapons should then be considered. For targeting humans, history demonstrates that the creation and distribution of bioweapons is incredibly challenging, even with a pathogen design in hand. The failures of Aum Shinrikyo, a Japanese doomsday cult from the 90s who sought to spread anthrax, and state sponsored bioweapon programs from the US, Russia, and Iraq clearly demonstrate the issues with such an undertaking, largely mitigating concerns regarding human applications.
For Mahajan, the more concerning scenario is that of agroterrorism. Given the genetic similarity and population densities of farmed plants and animals, the lack of surveillance on farms, sparser disease detection in crops and livestock, and a disincentivization for farmers to report disease outbreaks for fear of their herd being culled, the case for agricultural bioterrorism is evident. It then becomes important to understand how significant of a threat these attacks pose. Some models project economic costs of agricultural terrorism in the US to be upwards of $200B+ if spanning multiple states. On the other hand, in the grander scheme of the US economy this is somewhat insignificant, resulting in a 16% decrease to agriculture revenues (which as a whole comprises $1.4T) which is only a 1% hit to the overall economy (US GDP being $29T). An attack could be potentially devastating, but also likely recoverable.
Mahajan goes on to ponder if this scenario is a strong threat wielding significant consequences, why has it not been taken advantage of by optimistic terrorist groups? He theorizes that terrorism (particularly something as comparatively painstaking to carry out as bioterrorism vs nuclear attacks for example) is meant to strike fear in the hearts of a nation and/or bring said nation to its knees, both aims he asserts aggroterrorism fails to accomplish.
Moving on to consider means of defense, many issues with the current modes of detection are surfaced. BAR air monitoring has never actually detected threats despite triggering a multitude of detection flags (it has been likened to an overly sensitive but incredibly expensive smoke alarm). The development of wastewater screening during the COVID-19 pandemic has been positive, but it still fails to detect most diseases (i.e. beyond Covid, influenza, RSV, and a few others) due to limitations in qPCR testing.
Underlying these detection systems is the shortcoming that detection is not defense. Theoretically it enables a defensive response, but there is little structure set up to provide said response, which also relies on the existence of previous research into potential vaccine and therapeutic solutions. While the decidedly rapid turnaround of the COVID-19 vaccine by Moderna and BioNTech was a remarkable accomplishment, it was ultimately enabled by the existing decades of research into coronaviruses and their characteristic spike proteins. As a result, existing systems and infrastructure are far less well-equipped to handle novel pathogenic threats, especially in the case of viruses given their inherent penchant for mutating quickly.
As one potential solution to the delayed detection and response conundrum, Mahajan suggests metagenomic sequencing and machine learning could be useful. However, given the opinions of several experts he spoke with, ML is a ways away from being competent enough to help. As another potential solution, mRNA holds a lot of hope as an efficient and versatile modality despite its shortcomings. The production of de novo mRNA-encoded antibodies guided by ML models like RFdiffusion could be potentially very powerful for finding effective binders. However, as was seen with monoclonal antibodies during the COVID-19 pandemic, if a pandemic goes on long enough, inevitable viral mutations will likely render the antibodies obsolete. One potential way this immunity could be mitigated is via drug cocktails which leverage negatively correlated escape routes and low escape potential sites, both of which are and will continue to be enabled by ML.
Mahajan concludes his piece by summarizing what he is and isn’t scared of for the future of biosecurity. Agricultural bioterrorism, the accessibility of novel pathogens, and the temporal reality of a potential attack top the list of scary scenarios, while individual and state rogue actors and the ability of AI to create extremely advanced pathogens are ones he finds less threatening. Although he agrees with the experts that the likelihood of a future attack is low, the threat it poses makes it worth decidedly preparing for.
Papers
Sensitive CAR T cells redefine targetable CD70 expression in solid tumors [Hanina et al., Science, February 2026]
Why it matters: CAR T cell engineering has shown impressive results in treating blood cancers but struggles with solid tumors due to heterogeneous antigen expression. This work demonstrates how more sensitive HIT T cells can target markers like CD70 that were thought to be absent in some tumor cells to more effectively clear solid tumors.
Chimeric antigen receptor (CAR) T cells have shown immense promise towards the treatment of certain B cell malignancies like leukemias and lymphomas but struggle to yield the same effects in solid tumors. This has been attributed to varying levels of target antigens across the cell population, ultimately causing “incomplete tumor elimination and overall disease progression.” Ideally, CAR T-cell platforms need targets that are expressed on all tumor cells and common to many tumor types while showing low expression on otherwise normal cells. Previous work has established the CD70 marker as lowly expressed in normal cells while also present in many tumor types, albeit with heterogeneous expression within a tumor population with some cells having seemingly no CD70 at all. In this paper from the Sadelain Lab at Columbia University, the authors hypothesized that these CD70-negative cells may actually be positive for the marker, just at levels undetectable by conventional methods, and could be effectively targeted with a more sensitive receptor.
The team began by replicating the clinical failure of CD70-targeted T cells, using patient-derived renal cell carcinomas xenografts (human tissue transplanted into mice) from treatment-resistant patients. Even with high doses of three different CD70 therapies, the tumors failed to eradicate, showing only delayed progression and complete failure in more aggressive cancer models. Only after the tumors were artificially uniformly overexpressing CD70 did the CAR T-cells completely eradicate them. This result led the team to investigate why certain native “negative” cells were escaping treatment. Using confocal and stimulated emission depletion (STED) microscopy to better image cells with low-level signal, the authors found a spectrum of CD70 expression, with up to 80-90% of supposedly negative cells showing CD70 levels above those in CD70-knockout baselines. This proved that supposedly negative cells actually did have very low levels of the antigen, rather than none as previously believed. With this information, the team turned to HLA-independent T (HIT) cell receptors, a more sensitive alternative to traditional CAR T-cells, that can co-opt the highly sensitive downstream signaling of CD3 complex for signaling activation upon antigen detection. In vivo tests showed that the CD70-HIT T cells quickly eliminated tumors and showed improved responses compared to standard CAR T cell platforms in more aggressive examples. To prove that HIT T cells were specific to CD70, the authors also mixed wild type tumors with CRISPR-ablated CD70-negative tumors, which the HIT T cells failed to eradicate.
To understand why certain cells had ultra-low expression, both CD70-low and high cells were cultured and tracked over time, with the low cells slowly re-expressing protein after nearly eight days. ATAC-seq experiments showed that early-stage CD70-low cells had closed chromatin at the CD70 promoter which gradually opened up over the course of culture. This mechanism was found to be driven by reversible epigenetic silencing driven by EZH2-mediated histone modifications. To ensure this epigenetic mechanism could be safely targeted, the team mapped CD70 chromatin accessibility across various normal human tissues and then tested the HIT therapy in humanized mice models. Tests confirmed that the CD70 HIT T cells maintained antitumor efficacy without inducing significant off-target specificity or depleting normal immune cell populations. Finally, the platform was extended to CD70-heterogenous ovarian and pancreatic cell models, where tumors were completely eradicated. In summary, this research demonstrates the potential of CD70 and similar markers as promising targets for the development of more effective immunotherapies for a variety of cancer types.
Commensal-driven serotonin production modulates in vivo delivery of synthetic and viral vectors [Wang et al., Science, March 2026]
Why it matters: Drug delivery strategies such as LNPs, AAVs, and nanoparticles still suffer from low efficiencies due to rapid immune clearance. While most efforts have focused on improving the delivery vehicles themselves, Wang et al. instead modulate the immune system responsible for this clearance, leading to broad improvements across multiple in vivo delivery strategies. This introduces a new axis for optimizing drug delivery.
In vivo delivery systems (IDSs) have rapidly become essential tools for enabling cell-type–specific delivery, reducing systemic toxicity, and transporting otherwise unstable payloads such as CRISPR-Cas9, mRNA, or chemotherapeutics. But despite their promise, these systems remain expensive and inefficient in practice, with a large fraction of the administered dose never reaching the target tissue. Wang et al. approach this problem from a different angle: rather than redesigning the delivery vehicle, they ask how the body clears these particles, focusing on the liver as the primary site of clearance. Given the tight gut–liver immune axis, they hypothesize that the gut microbiome may act as a tunable regulator of immune-driven hepatic clearance of therapeutic delivery vehicles.
The authors start with this broad hypothesis and progressively narrow their inquiry. First, they ask a simple question: does removing the gut microbiome affect delivery at all? It does. Depleting the microbiome using antibiotics or germ-free models substantially improves therapeutic efficacy across multiple systems, including liposomal doxorubicin (Doxil), liposomal mitoxantrone, and polymeric nanoparticle formulations. This effect generalizes across disease models, with markedly reduced tumor growth and improved survival, indicating that the microbiome is broadly impairing delivery efficiency.
Mechanistically, this improvement comes from prolonged circulation of delivery vehicles. In microbiome-depleted animals, particles remain in the bloodstream longer, leading to increased accumulation in tumors and other target tissues. Using liver intravital microscopy, the authors show that Kupffer cells, the resident macrophages of the liver, are the dominant sink for these particles under normal conditions. When the microbiome is present, Kupffer cells are large, highly phagocytic, and rapidly clear circulating delivery systems. When the microbiome is depleted, these cells shrink, lose pseudopodia, and show markedly reduced uptake, allowing delivery vehicles to evade clearance and persist in circulation.
The authors then map the signaling pathway linking the gut microbiome to Kupffer cell activity. Rather than acting directly, gut microbes signal through intestinal epithelial cells, which produce serotonin from dietary tryptophan. This serotonin acts systemically to activate Kupffer cells via HTR2c/HTR7–ERK signaling, promoting cytoskeletal remodeling and upregulation of phagocytic receptors. Multiple interventions can transiently suppress this pathway, including gut microbiome depletion (antibiotics or germ-free models), blocking microbial sensing (e.g., TLR4/MyD88 signaling in intestinal epithelial cells), dietary tryptophan restriction, and pharmacologic blockade of serotonin receptors. Across these perturbations, delivery efficiency improves broadly: chemotherapy and oncolytic virotherapy show >3× therapeutic enhancement, while gene delivery approaches such as LNP-mediated mRNA and AAV-based editing improve by ~5–15× in extrahepatic tissues.
Together, this work establishes a generalizable strategy to bolster in vivo delivery treatments via tuning down the immune system’s rapid clearance response.
Latent-Y: a lab-validated autonomous agent for de novo drug design [Schmon et al., Latent Labs, March 2026]
Why it matters: Drug discovery remains an expert-driven, iterative process that is difficult to scale and parallelize. While generative models can propose molecules, the broader workflow – literature review, target understanding, epitope selection, design iteration, and candidate triage – still requires significant human coordination. This creates a bottleneck between model capability and real-world therapeutic output. Latent-Y addresses this by packaging the entire design loop into an autonomous system that executes end-to-end biologics discovery from natural language input.
Latent-Y is an AI agent that runs full antibody design campaigns, starting from a text prompt and proceeding through target analysis, epitope identification, candidate generation, validation, and selection of lab-ready sequences. It operates within a tool-augmented environment with access to databases, structural biology tools, and literature, and uses Latent-X2, a generative model for antibody design, to produce candidates. The system performs iterative, closed-loop design. It explores large combinatorial design spaces, evaluates intermediate outputs, refines hypotheses, and allocates compute toward promising regions. This explore-then-exploit behavior enables efficient search over epitope, framework, and sequence configurations within a single campaign.
Empirically, Latent-Y demonstrates end-to-end capability across multiple campaign types, including epitope-directed binder design, cross-species targeting, and literature-driven discovery. Across nine targets, the system produces lab-validated nanobody binders for six (67% success rate), with binding affinities in the single-digit nanomolar range. In controlled evaluations, it correctly identifies functional epitopes from scientific papers (21/21 cases) and generates high-quality candidate sets within practical sampling budgets. The system also changes the timescale of design. Expert-in-the-loop workflows assisted by Latent-Y complete campaigns ~56x faster than baseline expert estimates, compressing weeks of work into hours and enabling parallel exploration across multiple targets.
Overall, Latent-Y reframes biologics discovery as an agentic, programmable process. By integrating reasoning, tool use, and generative modeling into a single loop, it shifts drug design from sequential expert workflows to scalable, autonomous campaigns capable of producing experimentally validated molecules directly from high-level specifications.
Programmable artificial RNA condensates in mammalian cells [Li et al., bioRxiv, January 2026]
Why it matters: Biomolecular condensates – membraneless compartments formed by phase separation – organize critical cellular processes. Engineering artificial versions offer a route to controlling intracellular biochemistry. RNA is a natural substrate for condensation: repeat-expansion RNAs, ribosomal RNA, and overexpressed circular RNAs all form intracellular puncta. However, these condensates lack compositional specificity – the interactions driving them are difficult to design around. Similarly, most existing synthetic condensates rely on proteins with intrinsically disordered regions, facilitating promiscuous interactions with limited programmability. The core advance of Li et al. is rational control over RNA condensates inside living mammalian cells, using short (~100-200 nt), single-stranded RNA ‘nanostars’ – star-shaped motifs whose arms interact through sequence-specific kissing-loop (loop-loop base-pairing) interactions. Their approach offers fine-grained control: tuning nanostar structural parameters modulates condensate localization, size, orthogonality, and molecular recruitment.
Nanostars are expressed via the TORNADO system (which circularizes transcripts to extend their half-life) and form condensates in HEK293T, HeLa, and U-2 OS cells. The team systematically varied three structural parameters – arm length, arm number, and kissing-loop strength – across over 20 nanostar designs and quantified condensate volume, number, and subcellular localization per cell at 48 hours. Shorter nanostars (10 nt arms, ~57 kDa) exit the nucleus rapidly through pores and condense exclusively in the cytoplasm; longer variants (15-25 nt arms) are retained and nucleate first in the nucleus, with cytoplasmic fraction increasing over time. A compartment model explains this cleanly: condensation occurs wherever the nanostar concentration first exceeds the critical threshold, which depends on the balance between transcription rate, nuclear export rate (regulated by nanostar size and interaction strength), and cytoplasmic degradation. Increasing the arm number or kissing-loop GC content enhances nuclear retention and yields fewer, larger nuclear condensates. Up to three orthogonal (non-interacting) condensate species – each with distinct kissing loops (visualized via fluorogenic aptamer labels) – coexist without mixing. Chimeric ‘linker’ nanostars carrying kissing loops complementary to two distinct species titrate sub-compartmentalization: at low linker ratios, condensates form Janus-like (two-faced) droplets; at high ratios, they fully mix.
The modular nanostar architecture supports molecular recruitment: small molecules via fluorogenic aptamers, proteins via an MS2 aptamer arm (which binds to a reporter fused to the MS2 coat protein), and target RNAs expressed from separate transcripts – either through matched kissing loops or, more broadly, through a complementary hybridization domain that eliminates the need to modify the target transcript. FRAP (fluorescence recovery after photobleaching) analysis characterized condensate dynamics (recovery time constants of ~54s for cytoplasmic and ~13s for nuclear protein cargo), revealing relatively viscous, partially solid-like material properties comparable to nucleolar condensates. No innate immune response was detected – IFN-β and interferon-stimulated genes (ISG15, IFIT1) were measured at 3-44 hours post-transfection and showed no significant difference from controls.
While the paper demonstrates condensate formation and molecular recruitment, crucial caveats remain. Interestingly, high expression levels induced stress granule formation and nuclear enlargement, suggesting potential metabolic burden at current expression levels. Moreover, additional work is necessary to contextualize downstream effects on biological processes (translation regulation, enzymatic activity) and endogenous transcript recruitment. Progress toward resolving these limitations would enable this approach to extend from a programmable scaffold to forming functional synthetic organelles.
Notable deals
Dyno Therapeutics launches Psi-Phi, an open-weight agentic AI suite for protein binder design, at NVIDIA GTC 2026. The platform pairs Dyno Psi-1 – a flow-matching generative model trained on DGX Cloud using Hopper GPUs, influenced by NVIDIA’s La-Proteina family – with Dyno Phi, a set of predictive filters calibrated on experimental data to estimate wet-lab success. Psi-1 prioritizes structural diversity and controllability over pure in silico scoring. The suite includes API access to external tools like OpenFold3 via BioNeMo services, plus a Claude Code Skills plug-in for chat-based protein design without local infrastructure.
ImmuneBridge raises $7.7M second seed to open its cell therapy manufacturing platform to external partners. The round is led by NFX with participation from One Way Ventures, M Ventures, Insight Partners, LongGame Ventures, T.Rx Capital, Healthspan Capital, and Sand Hill Angels. Total seed funding is now ~$20M. Dr. Nina Horowitz, who built the donor screening system, was elevated from CSO to CEO; Rui Tostoes appointed CTO. The company is currently collaborating with 12+ partners across cell types, targeting ten therapies to clinic over the next decade with initial human trials in 2028.
Cathy Tie launches Origin Genomics to advance germline gene correction in the United States following the conclusion of Manhattan Genomics. The New York-based company will operate exclusively in the US under independent IRB oversight, focusing on precision genome editing for severe inherited diseases caused by well-characterized nuclear DNA mutations. Origin also intends to offer Mitochondrial Replacement Therapy contingent on passage of emerging federal or state legislation. At present, the company’s research roadmap centers on improving editing fidelity, reducing mosaicism, and conducting rigorous off-target analysis in early embryonic contexts.
R3 Bio emerges from stealth with a radical approach to replacing animal testing: growing nonsentient “organ sacks”. The Bay Area startup has been quietly pitching investors on structures containing all typical organs except neural tissue, rendering them unable to think or feel pain. The long-term goal is to create human versions as a source of tissues and organs for transplantation. Backed by Immortal Dragons, a Singapore-based longevity fund, R3 is initially targeting monkey organ sacks to address a critical shortage (after China banned nonhuman primate exports in 2020, there aren’t enough research monkeys in the U.S. for pandemic preparedness). Cofounders Alice Gilman and John Schloendorn say they’re exploring stem cell technology combined with gene editing, disabling genes needed for brain development so that embryos grow into organized organ structures. UC Davis stem cell biologist Paul Knoepfler calls the approach plausible using induced pluripotent stem cells reprogrammed to an embryonic-like state. Timing aligns with the Trump administration phasing out federal animal experimentation.
In case you missed it
Selection-free whole genome transplantation revives dead microbes
What we liked on socials channels
Events
8 days left to apply for the Encode: AI for Science Fellowship! Backing 10 brilliant AI / ML engineers, and giving them 12 months of freedom to build what matters, which includes a £115k salary, >£100k compute, an opportunity to collaborate with a top UK science lab, with access to datasets, wet labs, hardware, and expertise, and mentorship from advisors like Simon Kohl (Latent Labs), Andrew Giessel (Moderna), Theofanis Karaletsos (CZI), and Amy Villeneuve (Amazon Robotics). Outside of the UK? We help with visas. No equity taken. No fees charged. If you are interested, or have a talented contact/colleague/friend you want to recommend, contact encode@pillar.vc and anjali@pillar.vc. Apply here for Cohort 2!
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.