BioByte 139: LipiGo Takes Steps Toward Precision mRNA Drug Delivery, Smallest 3D Bioprinter Built for Vocal Cord Surgery, an Evaluation of Kosmos, and BIOINT for Addressing Biosecurity Threats
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.
Salvador Dalí, Galacidalacidesoxyribonucleicacid (1963)
What we read
Blogs
World’s smallest 3D bioprinter could rebuild tissue during surgery [Jenna Ahart, Nature News, October 2025]
Why it matters: After surgeries which remove cysts or growths from the vocal cords, there is often significant scarring of the vocal tissue. This scarring can lead to difficulty in speaking, as scar tissue does not afford the same flexibility as skin due to the linear alignment of the collagen within it, which limits movement. One potential solution to prevent scarring in the first place is the local delivery of hydrogels after surgery, which has been shown to improve the healing process. However, delivery of these materials proves difficult because of the confined space in which they must be applied. A recent paper in Device addresses the challenge of hydrogel delivery with an elegantly designed, elephant-inspired bioprinter that is compatible with the scope surgeons use for operations.
Initially constructed as a prototype with a diameter of 8 mm, the team led by Swen Groen at McGill University took inspiration from the trunks of elephants to design a 2.7 mm, flexible bioprinter capable of maneuvering within the 1 cm confines of the scopes used during surgery without obstructing the surgeon’s view. They demonstrated the effectiveness of the bioprinter by drawing various shapes and then using it during a mock throat surgery. They found that the low stiffness of the device, which prevents tissue damage upon contact, results in significant vibration, impacting precision. Further optimization of the stiffness and minimization of the vibrations will be necessary.
Three cables control the printhead, allowing it to be maneuvered into the correct position. It is currently piloted by a PlayStation controller, but future aims for the project seek to largely automate the process as the team shifts to testing in vivo. If you’re interested in what the device looks like, check out the paper for some cool videos of it in action!
Biosecurity Really: A Strategy for Victory [Endy et al., Hoover Institution, October 2025]
Why it matters: Biology is no longer a slow, opaque domain confined to specialist labs – it’s an information-rich technology stack that scales quickly, with cheaper DNA synthesis, ubiquitous sequencing, and AI design tools transforming biological engineering into a fast-moving, dual-use frontier. The advent of generative biology unlocks the possibility of novel pathogen and toxin creation, and – paired with a global climate of growing political conflict and suspicion – risks a return to nation-state bioweapons programs. Biosecurity Really argues that this shift means traditional, reactive public-health playbooks aren’t sufficient: the US (and its partners) must treat biology as a strategic domain and build durable, doctrine-level capability to detect, attribute, and respond to biological threats in near-real time. The report’s narrative illustrates how modest, concrete technological and institutional investments – including expanded metagenomic and wastewater surveillance, a National Biosecurity Institute (NBSI), and a market-facing ‘bioaudit’ regime – could raise the bar on adversarial misuse while accelerating benign applications. Perhaps the most compelling aspect is BIOINT (biological intelligence): an operational layer that ingests distributed biosensing and contextual data, runs a suite of complementary models (statistical, ML, phylogenetic, causal, and provenance checks), fuses their outputs probabilistically, and then routes high-value signals to fast wet-lab validation and human analysts for response. Its value is in translating raw, noisy biological signals into high-confidence, actionable alerts with quantified uncertainty and audit trails. The report makes the persuasive case that without integration, more data risks producing more noise and more false alarms.
The team offers a strategic, actionable blueprint: standing up BIOINT prototypes in 1000 days, piloting bioaudit standards to make verification a commercial norm, creating the NBSI to sustain countermeasure development, and investing in deployable, low-regret defenses (sequencing capacity, diagnostics, improved ventilation and Far-UVC, stockpiled biologics). These are high-leverage because they simultaneously raise the cost and detectability of misuse while improving routine public-health resilience.
Several recommendations surface hard tradeoffs and escalation points, requiring delicate political choreography to enact. A ‘default-no’ policy for certain gain-of-function work, expanded environmental surveillance, and tighter controls on synthesis and AI design may raise questions regarding civil liberties, scientific freedoms, and diplomacy. Implementing these demands definite governance instruments – clear statutory authority, independent oversight, privacy safeguards, and negotiated international norms. Operationally, building trustworthy BIOINT faces classic intelligence challenges – adversarial data manipulation, classification drift (model performance degrading as biological baselines shift), and false positives (spurious alarms which waste resources and erode trust) – which is why the report’s emphasis on validated lab pipelines and human-in-the-loop adjudication is well-founded. Pragmatically, the highest-value near-term wins are the ones the team prioritizes – find and pilot BIOINT prototypes, launch interoperable bioaudit pilots for high-risk facilities, and accelerate low-regret defenses whose benefits are immediate and measurable. If policymakers treat this document as a staged engineering program, it offers a robust path from ad-hoc crisis response to a resilient, integrated national posture for biological risk.
Papers
Kosmos: An AI Scientist for Autonomous Discovery [Mitchener et al., arXiv, November 2025]
Why it matters: FutureHouse spinout Edison Scientific builds Kosmos, an AI scientist that uses a world model of hypotheses and experiments to run through the scientific process of data-driven discovery over long periods of time. Kosmos can run up to ~12 hours, coordinating 200 agent rollouts that collectively write ~42k lines of code and read ~1.5k papers per run, while keeping its reasoning traceable through a shared, structured memory. In seven showcased studies, three reproduced findings from preprinted or unpublished manuscripts that Kosmos did not access at runtime, and four contributed new results or methods across metabolomics, materials science, neuroscience, and statistical genetics; independent scientists judged 79% of statements accurate overall.
What is the crucial step in building out a more thorough AI scientist? According to this team, building a world model that allows coherent flow of data, hypotheses, and conclusions flowing from an initial scientific inquiry and dataset. In practice, Kosmos’ world model is a structured store that both data-analysis and literature-review agents read from and write to each cycle, which lets the system keep pursuing the same objective over far more steps than prior agents like Robin, PaperQA2, or Finch. Each run launches dozens of parallel tasks per cycle, rolls their outputs into the world model, and, when the objective is “done,” compiles 3–4 short reports where every claim cites either code notebooks or primary literature. When put to the test against experts: 85% of data-analysis statements and 82% of literature-based statements were verified.
They tested 3 capabilities of the system: can it replicate results, can it match results from unpublished papers, and can it discover new things in existing papers? They organize seven case studies across replication (including unpublished/preprinted) and novel discovery, showing both replication and novel findings, and that “value” scales with more cycles. As an example of matching unpublished research, Kosmos discovered SOD2 as a driver of myocardial fibrosis in humans. Using GWAS of myocardial T1 (a noninvasive fibrosis proxy) and plasma pQTLs, Kosmos ran MR, fine-mapping, and colocalization and identified SOD2 as causally associated with lower T1 (β ≈ −0.23) with near-certain colocalization, then floated a plausible 3′UTR regulatory mechanism. Independent human re-analysis recovered nearly identical effect sizes (β ≈ −0.26) and direction, validating the target while noting caveats on the exact miRNA site. As an example of discovering new things, it discovered a breakpoint in pseudotime and ECM score in an analysis of proteomes of neurons with gradually increasing tau, modeling Alzheimer’s. Kosmos proposed segmented regression over an ECM composite score along pseudotime, found a statistically supported breakpoint (~0.58), and cross-checked the same trend at the transcript level in an independent dataset, suggesting a timed decline in ECM programs along the disease continuum.
So what - haven’t we already had some “AI scientists,” that still aren’t being used today? From a scientific perspective, the main takeaway here is that using a world model allows an “AI Scientist” to maintain a legible, scientific inquiry necessary for real-world science. Digesting swaths of literature, maintaining a large context while skillfully diving deep into thorough analyses, generating hypotheses, proposing experiments to effectively validate those hypotheses, and stitching together all of this into cohesive scientific inquiry are skills that take years (and often a PhD) to master. While there are limitations with Kosmos - notably around 58% of synthesis statements were accurate - like anything AI, expect this number to increase as the models powering the AI scientist become more effective. This also points to the power of the system: grounded hypothesis generation. The scientists reviewing Kosmos outputs estimate that Kosmos reproduced what would take an average of 6 months of computational analysis work and hypothesis generation. In collaborator studies, a 20-cycle run was rated at ~6.1 expert-months of work, and both “expert-time” and count of valuable findings rose roughly linearly with more cycles; that scaling story is the practical reason this may stick in real labs. Additionally, for wet lab scientists with limited bandwidth to learn the skills needed for those 6 months of inquiry, Kosmos is showing to be nearly capable of handling the computational analysis and inquiry that complements modern data-driven biological research.
LipiGo: A Versatile DNA-Lipid Nanoparticle Hybrid for Precision Drug Delivery [Kadletz et al., bioRxiv, October 2025]
Why it matters: Systemic lipid nanoparticles (LNPs) are the backbone of modern mRNA therapeutics but they nearly always localize to the liver, constraining the tissues and cell populations that can be targeted with nucleic-acid drugs. LipiGo introduces a subtle, modular modification: replacing part of the cholesterol in LNPs with cholesterol-tagged single-stranded DNA (ssDNA) handles. That change both (a) passively retunes LNP localization toward lymphoid organs and immune populations, and (b) functionalizes particles with programmable DNA handles that hybridize to complementary, DNA-tagged ligands (aptamers, antibodies) – enabling the modular display of targeting molecules. This combination tackles two central delivery problems at once: reducing hepatic expression of the payload and facilitating receptor-mediated cell targeting after administration.
The team produced LipiGo by substituting a fraction of cholesterol with a 26-nucleotide cholesterol-linked ssDNA, producing particles that are slightly larger (from ~50 to ~68 nm in diameter) and modestly more negative in zeta potential (surface charge in solution) while retaining encapsulation efficiency and serum stability. LipiGo demonstrates slightly faster tissue uptake (circulation half-time = 120 mins approximately) and, strikingly, a redistribution of functional mRNA expression from liver toward lymphoid tissues (spleen, lymph nodes, thymus) as measured by whole-body clearing / 3D imaging and orthogonal luciferase readouts. Single-cell immune profiling revealed that the fraction of mRNA-receiving immune cells in the spleen rose from 26.51% to 65.16%, with notable increases in antigen-presenting populations (e.g. F4/80+ macrophages, CD35+ follicular dendritic cells) and across adaptive (T-cell) and innate (myeloid) compartments. Mechanistically, comparative proteomics of the adsorbed protein corona (serum protein layer coating nanoparticles after injection) showed enrichment of Angiopoietin-1 (Angpt1) on LipiGo surfaces. The authors propose that this enrichment increases interactions with Tie2-expressing endothelium in lymphoid tissues, partly explaining the altered vascular permeability that facilitates tissue entry.
Beyond passive retargeting, the DNA handles allow for versatile, modular ligand display without altering particle properties. Conjugating an Asc-1 targeting DNA aptamer produced robust, selective internalization into white (but not brown) adipocytes in vitro. Moreover, after its subcutaneous administration, strong luciferase expression localized to abdominal white adipose tissue with minimal off-target signal – a concrete demonstration that LipiGo can execute receptor-specific delivery in vivo. Complementary cellular assays indicate that this targeted delivery improves downstream delivery steps, exhibiting faster endosomal escape (release into cytosol) and stronger cytoplasmic mRNA signal relative to standard LNPs. Crucially, at therapeutic doses used in their efficacy experiments, cytokine panels and serum markers showed little inflammatory or hepatotoxic signal. At substantially higher, supratherapeutic doses the team observed transient, modest metabolic perturbations.
LipiGo offers an elegant design that meaningfully expands the toolkit of extrahepatic and cell type-specific nucleic-acid delivery – especially toward tissues that vaccines and immunotherapies want to engage. That said, the mechanistic explanation (the Angpt1–Tie2 link) is correlative – targeted perturbations (e.g. blocking Angpt1/Tie2 or depleting/reconstituting the corona) are needed to prove causality. Important caveats also remain: human serum coronas differ from mice, ssDNA handles may trigger innate nucleic-acid sensors, and adipocyte targeting was only shown after local subcutaneous injection – systemic delivery to deep parenchymal tissues is untested. For translational purposes, the field will want: dose- and time-course biodistribution across species, rigorous immunogenicity profiling (including chronic dosing), demonstration of therapeutic benefit in disease models, and manufacturability and biostability data. If those steps succeed, LipiGo’s dual-purpose architecture could become a broadly useful pattern for programmable nanocarriers, enabling more precise vaccines, in-vivo cell engineering, and tissue-targeted genetic medicines.
Notable deals
Roche has agreed to pay Manifold Bio an upfront sum of $55M for access to their AI-guided shuttle discovery platform. To better treat neurological and neurodegenerative disorders, Roche will leverage both Manifold’s mDesign—an AI-driven in vivo discovery engine—and Manifold’s existing shuttle portfolio to bring about tailored next-gen blood-brain barrier shuttles. Within the deal, Manifold will retain the ownership of the shuttles they develop, passing everything preclinical and beyond to the pharma. The deal has the potential to be worth up to $2B, with additional payments for research and development milestones as well as tiered royalties. Manifold was founded in 2019 as a spinout from the Church Lab and raised their $40M Series A over three years ago.
NEOK Bio, a Palo Alto based company building antibody drug conjugates (ADCs) for cancer, emerged from stealth with a $75M Series A. The funding round was led by the Korean biotech ABL Bio—a leader in the ADC space—and the funds will be used to carry two assets ABL initially developed into the clinic, with the aim of having both in Phase 1 by mid-2026 with readouts in 2027. Both assets NEOK will be developing are bispecific ADCs with potential applications in a range of cancers, including gastrointestinal, gynecological, and thoracic. NEOK001 (previously ABL206) targets ROR1 and B7-H3 while NEOK002 (previously ABL209) targets EGFR and MUC1.
Led by Artesian (Alternative Investments) and Topology Ventures, Coherence Neuro raised a $10M seed round to begin the company’s first human trials. Coherence’s product, dubbed SOMA-1, is an implantable BCI that is able to monitor the progression of a brain cancer and administer treatments via electrical stimulation in real time. The thesis of the company is that cancer is a system-state error, and that based on this understanding, a fourth pillar of electrical-based cancer treatment is necessary to supplement chemotherapy, radiotherapy, and surgery. Other investors in the round include Blackbird, Divergent Capital, Jumpspace Ventures, Possible Ventures, SmartGateVC, Spacewalk VC, and XEIA Venture Partners.
Azalea Therapeutics launched with an $82M combined seed and Series A led by Third Rock Ventures. The funds will be used to continue to develop several assets, including bringing an in vivo CAR-T therapy for B cell malignancies and autoimmune diseases to the clinic and the further development of another in vivo CAR-T for multiple myeloma. The fundraise will also assist in the further development of Azalea’s platform, which utilizes their proprietary Enveloped Delivery Vehicle (EDV) technology to selectively target T cells to deliver CRISPR-Cas9 cargo, precisely inserting large genes at desired locations. The company, which is currently hosted at UC Berkeley’s Bakar Bio Labs, was spun out of a collaboration between the labs of Dr. Jennifer Doudna of UC Berkeley and Dr. Justin Eyquem of UCSF. Other financiers who participated in the combined round included RA Capital Management, Yosemite, Sozo Ventures and select individual investors.
In case you missed it
Scientists complete first drafts of developing mammalian brain cell atlases
Events
Don’t miss our next AI x Bio Meetup in Boston on Nov. 18th! Sign up here.
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.