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
AI Versus Eroom’s Law [Elliot Hershberg, Century of Biology, April 2026]
Why it matters: Pharma R&D efficiency has been declining exponentially even since the 1950s, with many arguing that AI-powered drug discovery will do little to alleviate the crisis. Here, Hershberg argues for a synthesis of both the pro-AI and realist positions, and lays out how reducing R&D cost and making discovery more efficient would benefit emerging biotech most instead of pharma, resulting in a new wave of ‘digital biologics’ companies that choose to develop their own drugs instead of selling to pharma.
“Eroom’s Law” refers to the trend of pharma R&D efficiency halving every nine years, the inverse of “Moore’s Law” which led to the decline in compute costs that gave us the electrified modern era. Much hype has surrounded AI-powered drug discovery potentially changing this, but the evidence shows that discovery accounts for just 9% of out-of-pocket costs for bringing a drug to market while clinical trials account for 68%; on the surface, it appears pharma has more molecules than they can afford to test, having spent $1 trillion since 2010 on mergers and acquisitions (the primary method by which pharma companies expand their portfolios) yet only 50 drugs are approved every year. Hershberg takes us through several decades of pharma history to make the case for both points of view.
First, Hershberg argues that the cost of drug development is split: pre-clinical phases are borne by emerging biopharma that constitute the innovation pipeline big pharma has outsourced discovery to, while clinical development and scale-up is handled by pharma. This makes the numbers appear deceptive: for this ecosystem of emerging biotechs, efficient and differentiated discovery is their raison d’etre and necessary to support innovation, which pharma, given the astronomical sum spent on M&A, is still hunting for. In addition, the discovery phase holds more weight due to the time it exhausts, and because it determines whether the rest of the campaign will continue.
Secondly, Hershberg outlines that bringing down R&D costs is necessary and will likely support small biotech rather than pharma the most. There is a limited number of big pharma companies with limited infrastructure who will be rate-limited by the number of drug candidates they can take in. Reducing R&D costs could enable the creation of more ‘full-stack’ biotechs that go on to develop their own drug candidates, pursuing clinical directions that are important to pursue but that pharma will not: he points to the lack of M&A spend on obesity prior to Novo and Eli Lilly ‘de-risking’ the market by developing GLP-1s, potentially neglecting one of the biggest opportunities in biotech history because pharma was previously reluctant to invest in it. The increased focus on obesity by big pharma has now led to a flurry of startups pursuing obesity-related targets. He also outlines the 3-fold rise in the number of drug launches by ‘first-time launchers’ as evidence of this trend already growing.
In addition, Hershberg argues AI will enable the creation of both more and better drugs, a class of treatments he refers to as digital biologics: drugs like Moderna’s intismeran autogene, a personalized mRNA cancer vaccine, or Aspen Neuroscience’s sasineprocel, a dopaminergic neuron cell replacement therapy for Parkinson’s disease, that could not have been made without the machine learning algorithms involved in their development, such as the algorithms Moderna uses to rank cancer targets and those Aspen uses to assess cell quality. To conclude, he emphasizes that biotech in AI is not as black-and-white as a revolutionary discovery technology or a folly constrained by development costs: it is far more nuanced and could be transformative for the biotech industry at large.
Papers
B lymphocyte protein factories produced by hematopoietic stem cell gene editing [Hartweger et al., Science, April 2026]
Why it matters: Hartweger et al. crack an existing challenge of long-term installation of antibodies into the immune repertoire by gene editing hematopoietic stem cells rather than mature B cells. For diseases like HIV, for which widespread vaccination strategies have failed, this enables a potential new cell-therapy based immunization solution.
HIV is a hard disease to vaccinate against. For other widespread viruses like influenza, vaccines work because generally most individuals are able to develop antibodies to the vaccine. HIV has a highly glycosylated envelope which the human immune system has a hard time developing any antibodies against. This has stymied the development of HIV vaccines. On the other hand, we have discovered effective antibodies for HIV in rare cases of patients who were able to develop antibodies. As a result, the field has been attempting to deliver these pre-designed HIV antibodies in vivo to enable long term protection against HIV.
Several approaches in the last decade have installed these antibodies into B cells and plasma cells using gene editing methods. However, those responses were short-lived, as the antibody production by mature B cells gradually waned over time. To tackle this problem, Hartweger et al. climb up the cell differentiation ladder. By targeting hematopoietic stem cells, which can self-renew and give rise to B cells that respond to the proper stimulus.
The immunization strategy is as follows. Extract and isolate hematopoietic stem cells from mice. Use AAV6 and CRISPR-Cas9 to install the HIV antibody into the IgH locus. They carefully designed this gene cassette to downregulate expression of any endogenous heavy chain this stem cell would otherwise create by including an SV40 polyadenylation signal. This helps prevent expression of an endogenous heavy-chain product and biases the cell toward producing the antibody of interest. Then this population of stem cells is grafted back into the mouse, and the mouse is immunized with a cognate HIV antigen.
What is stunning about this approach is how few cells were needed to effectively generate the necessary amount of antibody. The AAV6-based gene editing of the stem cell population was actually rather inefficient - fewer than 10% of the grafted stem cells were successfully edited with the new antibody cassette. Yet even in the case where as few as 29 successfully gene-edited cells were delivered to the mouse, with proper priming and boosting procedures, the mouse was able to produce protective levels of the anti-HIV antibody! This is because B cells with useful antibodies rapidly expand, so those 29 initial cells can clonally expand and differentiate into plasma cells that begin secreting as much as 104 antibody molecules per second, enabling a large immune response from a small set of gene-edited cells. Most importantly, they show that the approach is durable, unlike previous methods that directly target mature B cells.
The authors validated this approach across different disease targets. They also validated that this approach could be performed combinatorially: when grafting populations that contained mixes of two unique anti-HIV antibodies, the mice developed strong expression of both anti-HIV antibodies - indicating that this approach can scale to multiple antibodies. For viruses like HIV which can quickly escape a single targeted epitope, this is huge - it allows the immune system to target multiple epitopes at a time, reducing the likelihood of epitope escape. While ex vivo gene editing and grafting of HSCs is expensive and not scalable, this approach can potentially be combined with modern in vivo gene therapy approaches for affordable installation of specific antibodies into a patient’s immune repertoire - and maybe a true HIV vaccine.
RosettaSearch: Multi-Objective Inference-Time Search for Protein Sequence Design [Kshirsagar et al., bioRxiv, April 2026]
Why it matters: Current state of the art protein sequence design models like ProteinMPNN frequently fail to produce high fidelity sequences that correctly fold into their target structure. In this preprint, RosettaSearch addresses this bottleneck by leveraging LLMs to help design models iteratively refine their designs using structural feedback metrics without requiring expensive task-specific retraining or finetuning.
Current approaches for backbone-conditioned sequence design problems like ProteinMPNN and LigandMPNN generate protein sequences likely to fold into an input target structure. However, generated sequences from such tools are known to exhibit a lack of fidelity, meaning that they don’t actually adopt a target structure when independently verified by structure prediction tools like AlphaFold. Specifically, the autoregressive decoding architecture of Protein/LigandMPNN-like tools means that they generate a sequence in a single pass and are unable to use knowledge from structural verification tools to revise their outputs. Sequence design then becomes a function of hoping you get lucky with the next candidates instead, meaning that such models are ultimately more optimized for sequence recovery rather than true structural fidelity. In this paper, the authors present RosettaSearch, an inference-time multi-objective optimization approach that leverages large language and vision-language models to act as optimizers for sequence generation using structural feedback. Crucially, RosettaSearch optimization occurs at inference time as it does not require task/context-specific retraining and fine-tuning seen with other approaches.
RosettaSearch builds on previous work attempting to use LLMs and textual feedback as the basis for optimization over internal model gradients. The pipeline begins by using RosettaFold3 to predict the structure of a candidate sequence, after which best prediction is scored using fidelity metrics like predicted Local Distance Difference Test (pLDDT), root mean square deviation (RMSD) of the alpha carbon backbone, and the template-matching (TM) score. These metrics are paired with textual descriptions (e.g. “0.3 is a very low reward. Significant improvements needed.”) to give an optimizer a sense of current global quality. Furthermore, the authors also incorporate local text feedback to help direct optimization towards known problematic regions, again based on the previously mentioned fidelity metrics. Interestingly, in an approach somewhat mirroring a scientist physically inspecting a protein structure, 3D renderings of the predicted protein structure were passed to vision-language models to highlight secondary structure, domain topology, and exposed/buried region information at the time of optimization.
The authors used two methods to actually perform the optimization - sequential revision and priority search. Sequential revision considers the protein sequence as a trainable text variable and creates a “structured problem instance from intermediate states, outputs, and scalar or textual feedback consisting of variables, 6 constraints, code, and observed outcomes” that is provided to the LLM optimizer at each iteration. In contrast, priority search maintains a running “buffer” of all design attempts and their evaluation results, and iteratively generates new candidates in parallel for the previous top-K designs. To enforce optimization across multiple objectives, the different fidelity metrics were combined with a weighted sum to produce a single scalar reward for a given sequence, with the majority of emphasis being split between TM score and RMSD and pLDDT accounting for less than 10%.
To evaluate RosettaSearch’s capabilities, the authors evaluated the pipeline against various datasets like the LigandMPNN dataset of PDB monomers (test low fidelity sequences that need fixing), the Dayhoff Dataset of de novo protein backbones from RFDiffusion (to test if the optimizer could improve already good designs), and binder sequences from BindCraft (to test improvements for already highly optimized functional sequences). On the LigandMPNN dataset, RosettaSearch significantly improved the design success rate from 7.9% to just over 20%, while improving from 72.4% to nearly 90% success rates on the Dayhoff dataset. On the BindCraft dataset, the authors found that RosettaSearch improved structural fidelity metrics across all binders without breaking the ligand-binding pocket interaction. As a test of robustness against potential biases from RosettaFold3 (the structural oracle), the team also verified their improved designs using Chai-1 which did lower absolute scores but still confirmed a strong improvement over baseline metrics. In summary, RosettaSearch demonstrates the potential of LLM-based reasoning to improve protein design and optimization pipelines without the need for expensive task-specific retraining or fine-tuning.
CHARIOT-AAV: Conjugation of diverse vectors to adeno-associated viruses for delivery of large genes [Nagao et al., bioRxiv, April 2026]
Why it matters: Adeno-associated viral vectors are currently the standard for cell-type-specific transgene delivery. Unfortunately, their delivery capacity is pitiably small, far below the size of many therapeutic payloads such as CRISPR gene editing systems or dystrophin, among others. Other vectors such as lentiviruses or LNPs have a much higher loading capacity but cannot deliver payloads in a tissue-specific manner. Nagao et al. attempt to achieve the best of both worlds by using click chemistry to physically conjugate AAVs and other vectors together to improve tissue-specific delivery of large payloads.
Adeno-associated viral vectors are widely utilized for their low immunogenicity, low pathogenicity, and tissue-specific tropism. However, the delivery capacity of most AAVs is far below the size of many therapeutic cargoes including the smallest CRISPR-Cas gene editing systems such as CasMINI. Larger vectors such as lentiviruses or lipid nanoparticles (LNPs) while having the ability to deliver up to 10kb, lack a specificity for specific tissue or cell types. The split-intein system enables splitting protein cargoes among two AAVs, but requires that both AAVs transduce cells simultaneously, reducing the number of cells that receive a full copy.
Nagao et al. instead used the already-characterized split-intein system but physically conjugated the AAVs containing either end of the split protein, hence bypassing the efficiency reduction having both AAVs transfect cells simultaneously causes. They tested this system, called CHARIOT-AAV, with two AAV serovars: AAV-DJ, known for its ability to transfect HEK293 cells, and AAV-9, which cannot, delivering GFP with AAV-9 and mRuby2, a red fluorescent protein, with AAV-DJ. They proved that AAV-DJ could escort AAV-9 into HEK293 cells it could not otherwise transfect, leading to a 13.3-fold increase in the number of cells expressing GFP.
The authors next tested if this approach could translate to other vectors, fusing lentiviruses to AAVs by means of the SnapTag system. Despite the lentiviruses being pseudotyped with a mutant version of the vesicular stomatitis virus G-protein (normally fused to them to allow them to transfect HEK cells) that impaired their transduction, this yielded an over 13-fold increase in the delivery of lentiviral cargoes. The same approach transferred to LNPs - which are scalable and easy to manufacture but cannot target extrahepatic delivery - resulting in a 72-fold increase in delivery efficiency to HEK293 cells.
There is a caveat: LNPs, when conjugated to the brain-specific AAV serovar AAV9 and used for brain-targeted delivery in vivo, released their mRNA cargo into endothelial cells instead of going through transcytosis along with AAV9. The authors conclude by emphasizing the need for LNPs with delayed endosomal escape kinetics and the potential conjugating three AAV capsids has to improve the delivery of even larger cargoes.
Notable deals
Eli Lilly to acquire Kelonia Therapeutics in a new deal worth up to $7B. Kelonia is pioneering personalized medicine via their proprietary in vivo gene placement system (iGPS0) which promotes generation of CAR-T therapies by and within the patient’s own body. This feat is accomplished by way of specially engineered lentiviral-based particles which selectively enter the recipient’s T-cells inducing the CAR-T cell generation, promoting more effective immune response to disease. The company’s lead program, KLN-1010, exemplifies this platform as a one-time, intravenous gene therapy which yields anti-B-cell maturation antigen (BCMA) CAR-T cells, targeting treatment of multiple myeloma (which expresses BCMA proteins on the diseased cell surfaces). Not only have the initial clinical data of KLN-1010 been promising, but the platform as a whole holds potential to revolutionize patient access to CAR-T therapies, a treatment class which has faced substantial challenges in manufacturing, safety, and overall accessibility, thus limiting availability of these transformative medicines to patients in need. Furthermore, the platform stands to be expanded to a wide variety of cancers and other diseases, skyrocketing its potential value proposition of bringing cell therapy rapidly into the mainstream through greater accessibility. As a part of the deal, Kelonia shareholders will receive $3.25B upfront, with additional payments predicated by the achievement of specific clinical, regulatory, and commercial milestones.
Odyssey Therapeutics reignites plans to IPO on Nasdaq. The biotech first voiced desires to go public in January 2025 but pressed pause on the initiative following a slow and uncertain IPO market experienced by mid-year. Nonetheless, this sentiment has since changed following the recent flurry of biopharma IPO announcements which seem to indicate signs of life in a thawing IPO market. Pricing details are not yet disclosed, however, indications have been made that funding from the offering will first and foremost go toward advancing Odyssey’s lead candidate, OD-001, an RIKP2 inhibitor in Phase 2 trials for ulcerative colitis as both a standalone and combination therapy with Takeda’s existing Entyvio. Funds will also reportedly support the entry of another asset, SLC15A4, a small molecule therapeutic targeting cutaneous lupus erythematosus, nephropathies, and B-cell mediated diseases, into a Phase 1/2a trial. Still, Odyssey’s founder previously emphasized funding allotment to preclinical pipeline buildout whilst still balancing this initiative with additional clinical progress according to further coverage by Fierce Biotech. The company previously raised several megarounds since its founding in 2021 and adds itself as another strong contender to the IPO fray.
Serif Biomedicines launches out of Flagship Pioneering with $50M commitment. The latest of Flagship’s spinouts, Serif is championing Modified DNA as the newest class of medicines. Described as a combination of the best features from mRNA and gene therapies, the company is capitalizing on burgeoning capabilities for foundational biological engineering to reenvision DNA as a broadly applicable therapeutic modality. By reshaping the structural and chemical form of DNA, Serif is seeking to build out a tech stack that enables direct-in-cell programmable DNA medicines that are not only cell-specific but also demonstrate reduced immunogenicity (historically a significant issue facing traditional gene therapies). The platform further encompasses everything from optimized mRNA-enabled nuclear entry, precise LNP delivery, AI-guided sequence design, and scalable manufacturing. Serif’s initial focus is rare disease and immune programming to address unmet clinical need, however, the technology’s generalizability signals likely applicability to a wide range of indications. The launch follows five years of platform development and the company enters the scene with strong preclinical data of tolerability and sustained expression in non-human primates. See the full debrief and exclusive coverage here from this week’s interview of Serif co-founder Jake Rubin by our very own Zahra Khwaja!
Tortugas Neuroscience emerges from stealth with $106M in seed/Series A funding co-led by Cure Ventures, The Column Group, and AN Venture Partners. Led by a pair of accomplished biotech veterans, Tortugas is in the process of developing several de-risked clinical assets for a handful of heavy-hitter neurological indications: schizophrenia, tinnitus, focal epilepsy, and reversible encephalopathies, among other critical CNS disease areas. The company’s pipeline was licensed from Japan-based Ensai Co., Ltd. and is comprised of small molecule new chemical entities with established mechanisms of action. With a clear path toward regulatory approval and a fervent devotion to address high unmet clinical need, Tortugas will use funds raised to progress their two unspecified lead candidates through Phase 2 trials as well as for further R&D.
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.
The HIV story is especially interesting because it blurs the line between vaccination and cell therapy. Instead of training the immune system in the usual way, it installs antibody producing capacity more directly. Do you think this kind of engineered immunity will become a practical vaccine strategy?