BioByte 155: Advancing Engineered Allosteric Control, Amplifying Therapeutic Effects of Cell Therapies, EVEE Enables Variant Effect Prediction, and Improving Organ Transplants with CAR Treg Cells
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
Papers
Artificial allosteric protein switches with machine-learning-designed receptors [Guo et al., Nature Biotechnology, April 2026]
Why it matters: Allostery underlies most biological sensing and regulation, but engineering new allosteric proteins has been constrained by the need for natural ligand-binding domains that undergo large conformational changes. This limits the space of inputs that can be coupled to functional outputs. This work shows that allosteric control can be engineered without global conformational change, using machine-learning-designed binding domains, expanding the design space for programmable protein sensing and control.
Guo et al. construct single-component protein switches by inserting ligand-binding domains into reporter enzymes such as β-lactamase, luciferase, and glucose dehydrogenase. These binding domains include both natural proteins and ML-designed minibinders, which are structurally stable and lack large conformational transitions. Despite this, they generate ligand-dependent changes in catalytic activity across a wide range of ligands, including small molecules, peptides, and proteins.
The system is generalizable across architectures. Artificial receptors can be combined with different reporters and composed into logic gates, including intramolecular YES and AND gates that increase dynamic range or require multiple inputs for activation. Fully synthetic switches are also demonstrated by pairing ML-designed receptors with ML-designed reporter enzymes, showing that both sensing and output modules can be engineered de novo.
Mechanistically, the switches operate through changes in conformational entropy rather than large structural rearrangements. Biophysical measurements (CD, NMR, HDX-MS) show minimal global structural change upon ligand binding, but detectable shifts in local dynamics that propagate to the reporter active site. Ligand binding reduces conformational entropy of the system, increasing catalytic activity, while circular permutation increases entropy and suppresses activity. Functionally, the approach enables practical applications. The authors engineer ligand-dependent antibiotic resistance in E. coli, where growth depends on the presence of specific small molecules, and build electrochemical biosensors that convert ligand binding into measurable electrical signals. These systems operate with high dynamic range and can be reset and reused.
Overall, this work reframes allosteric protein design as a modular, composable problem. By decoupling allostery from large conformational changes and leveraging ML-designed receptors, it enables construction of programmable protein switches with arbitrary inputs and outputs, forming a foundation for synthetic sensing, control circuits, and bio-integrated devices.
Combinatorial base editing couples disease correction with lineage amplification in hematopoietic stem and progenitor cells [Jia et al., bioRxiv, April 2026]
Why it matters: The first approved CRISPR therapy for sickle cell disease (SCD), Casgevy, works by reactivating fetal hemoglobin – a form normally silenced after birth – that can functionally compensate for defective adult hemoglobin. However, corrected stem cells do not gain any competitive edge in the bone marrow: hemoglobin genes only turn on late in red blood cell (RBC) maturation, so edited cells are no more ‘fit’ than unedited ones. Therapeutic success thus hinges on editing and engrafting enough stem cells to matter – which currently requires harsh chemotherapy conditioning. Jia et al. reframe the issue: rather than solely correcting the disease, they engineer edited cells to preferentially expand in the RBC lineage. They do so by combining fetal hemoglobin-reactivating edits with a naturally occurring erythropoietin receptor truncation (tEPOR) – first identified in a Finnish Olympic cross-country skier – that removes the receptor’s built-in off switch (the SHP-1 binding domain that normally terminates proliferative signaling), making edited erythroid progenitors hypersensitive to the body’s own growth signals. This results in a multiplex base editing strategy where corrected cells outgrow their unedited neighbors – amplifying therapeutic output (hemoglobin) from a given engraftment level.
The strategy rests on achieving efficient editing at up to three independent genomic loci (the 𝛾-globin promoter HBG, erythroid enhancer of the fetal hemoglobin repressor BCL11A, and tEPOR) in the same primary human stem cell (prior work has typically reached two loci with variable efficiency). Using an optimized, next-generation cytosine base editor (CBE6), the authors achieved ~75-86% editing at each target in healthy donor HSPCs, which they verified with clonal genotyping. The addition of TEPOR drove a >4-fold increase in erythroid cell output during in vitro differentiation relative to conditions without it – including Casgevy – across healthy donor, SCD, and β-thalassemia patient-derived cells. In SCD HSPCs, the triple combination (HBG + BCL11A + tEPOR) reached 84.6% fetal hemoglobin as a fraction of total hemoglobin, substantially exceeding both Casgevy (~34.5%) and the ~60% reported for the most advanced clinical base editing trial, BEAM-101. Spike-in experiments – where edited and unedited cells were mixed 1:1 to model post-transplant competition – showed even larger gains, consistent with progressive enrichment of edited cells during differentiation. When the authors attempted the same tEPOR combination using Cas9 rather than base editing, the proliferative advantage disappeared – likely because Cas9 generates heterogeneous indels at EPOR (only a subset encoding functional truncations) and dual double-stranded breaks (DSBs) cause compounding genotoxic stress. Notably, tEPOR appeared to rescue a known liability of BCL11A disruption: reduced erythroid output (ineffective erythropoiesis), which was evident in both Cas9 and base editor BCL11A-only conditions but compensated when tEPOR was co-edited.
The in vivo data is encouraging but carries important caveats. At 16 weeks post-transplant in immunodeficient mice, multiplex base-edited cells achieved 50-75% human cell engraftment with preserved multilineage (myeloid and lymphoid) output – outperforming Casgevy (~35-50%) and Cas9/AAV conditions, where engraftment was severely compromised (by the DSBs and viral vector delivery). Editing frequencies were maintained or increased at all base-edited loci in long-term engrafted cells, while Casgevy allele frequencies declined from ~85% to ~69% and Cas9/AAV editing effectively vanished. However, human RBC production is not detectable in this xenograft model, so the core claim – that tEPOR amplifies therapeutic RBC output – currently rests on in vitro differentiation and colony assays, not on demonstrated erythroid advantage in vivo. Whether tEPOR-driven expansion translates to meaningful clinical benefit at lower engraftment thresholds, and whether a proliferative erythroid allele raises any long-term safety risks beyond the benign phenotype observed in the original proband, remain open questions. If the in vivo erythroid advantage holds, this framework – layering fitness-enhancing edits onto disease-correcting ones – may meaningfully lower the engraftment bar for curative therapies and reduce dependence on toxic conditioning regimens.
EVEE: Interpretable variant effect prediction from genomic foundation model embeddings [Pearce et al., bioRxiv, April 2026]
Why it matters: Current foundation models are capable of learning powerful representations of the genome that can be used for variant effect prediction tasks; however, these models lack interpretable explanations for their predictions. Pearce et al. attempt to tackle this problem by training lightweight classifier and interpretability probes on Evo 2 embeddings and pairing them with natural language reasoning models to generate human-readable explanations of a variant’s predicted pathogenicity.
Understanding how variants in the genome relate to disease biology is still an open question, with most variants still being classified as variants of unknown significance (VUS). The use of foundation models offers a more tractable way to make predictions about variants; however, most models focus heavily on either missense mutations in proteins (like AlphaMissense) or noncoding regions of DNA (AlphaGenome). More importantly, most foundation models lack a robust internal framework to explain their predictions in a human-readable format which is critical for their use in clinical settings. In this joint work between Goodfire AI and Mayo Clinic, the authors detail the development of Evo Variant Effect Explorer (EVEE), a unified framework that utilizes representations from the Evo 2 genomic foundation model to deliver state-of-the-art variant effect prediction with accompanying human-readable interpretations.
Rather than developing a new model from scratch, the authors chose to use embeddings from the Evo 2 genome language model. Previous studies have shown that Evo 2 is capable of learning powerful representations of coding and non-coding regions of the genome that are useful for variant effect prediction, despite not having seen any human variants during training. The authors began with a genome-wide pathogenicity probe, training a small classifier on frozen Evo 2 embeddings. Crucially, the authors chose to use covariance-based sequence pooling rather than standard mean pooling to preserve feature co-occurrence patterns. When evaluated on over 800 thousand single nucleotide variants from ClinVar and zero-shot prediction tasks on short indels, the probe was highly performant with AUROCs of 0.997 and 0.991 respectively. Importantly, this method was able to outperform classical meta-predictors like CADD as well as protein-based models and current genomic foundation models.
Next, the team focused on making these predictions interpretable by training new probes on reference sequence embeddings. Different probes were geared towards predicting annotations like splice sites, regulatory elements, and post translational modifications while also predicting how much a certain variant would alter those properties. Apart from a DNA-level interpretation probe, the authors also trained probes focused on amino acid properties to capture variant effects at the protein level. To convert these probes into human-inteligible outputs, the “disruption profiles” were fed to large language reasoning models to generate natural-language explanations based on the information from the probes and Evo 2 embeddings. To judge the quality of interpretations, the team adopted a LLM-as-a-judge approach to compare the generated explanations against ground-truth expert panel evidence from 154 expert-reviewed ClinVar variants. Looking across mechanism coverage, biological accuracy, and specificity, the integration of Evo 2 probe predictions drove the largest improvement in explanation quality, demonstrating that the model’s representations can be extended to interpretable predictions for clinical use.
CAR Treg cells mediate linked suppression and infectious tolerance in islet transplantation in mice [Wardell et al., Science Translational Medicine, August 2025]
Why it matters: Wardell et al. develop a new way to suppress immune responses to HLA-mismatched transplants by engineering HLA-specific CAR-Tregs.
Organ transplants are often rejected because of differences in HLA proteins between donor and recipient. When an organ is transplanted, the cells in that organ present peptides on MHC class I molecules. Your immune system is trained to recognize your own peptides presented on your own HLA alleles as self. If you receive a transplant from a donor with a different HLA allele, such as HLA-A2 when you do not express HLA-A2, your T cells recognize those HLA-peptide complexes as foreign and mount an immune response, leading to rejection.
The authors develop a strategy to prevent the recipient immune system from attacking an HLA-A2-positive graft by engineering Tregs that are specific to HLA-A2. They do this by introducing a CAR targeting HLA-A2 into Tregs. After confirming that these Tregs maintain their identity, marked by FOXP3 and HELIOS, and remain functional in vitro, they test the approach in vivo. In a mouse model of islet transplantation, A2-CAR Tregs reduce expansion and cytokine production of diabetogenic T cells and protect most mice from hyperglycemia, performing substantially better than polyclonal Tregs.
Not only does this prevent immune-mediated graft damage, but it also leads to a phenomenon called infectious tolerance. The idea is that while the HLA-A2-specific Tregs suppress T cells attacking the graft, they also induce tolerance to other antigens in the tissue. This makes the effect more robust than HLA-A2 targeting alone: even after removing both the HLA-A2 graft and the CAR-Tregs, the mice remain euglycemic. In other words, the Tregs effectively teach the immune system to treat the transplanted organ, through HLA-A2 and other antigens, as self.
Notable deals
Kailera Therapeutics unveils pricing for IPO with plans to raise $528.5M to list on Nasdaq. Following Kailera’s announcement to go public last month, the company’s new filing with the SEC lends further details to their offering as well as planned uses of proceeds. The filing outlines a prospective offering of 33.33M shares priced between $14-$16 with estimated proceeds of $458.7M and up to the $528.5M should underwriters exercise their 30-day option to purchase 5M additional shares at the same price point, according to coverage by Fierce Biotech. Funds generated from the IPO will go toward advancing Kailera’s strong obesity portfolio licensed from China-based Jiangsu Hengrui Pharmaceuticals. Specifically, Kailera will purportedly allocate $625M to their injectable ribupatide, a potential competitor to Lilly’s Zepbound, currently in Phase 3 trials and aspiring for best-in-class designation; $150M toward their once-daily oral ribupatide, anticipated to enter Phase 3 trials in 2028; $50M toward their small-molecule, injectable GLP-1 candidate, KAI-7535, as it advances into Phase 2 trials (currently awaiting Phase 3 readouts in China from Hengrui); and an undisclosed amount to their fourth GLP injectable candidate, KAI-4729, also in clinical trials (albeit Phase 1) in China.
Seaport Therapeutics files for IPO with potential to raise up to $100M to advance neuropsychiatric drugs through clinical trials. Led by ex-Karuna CEO, Daphne Zohar, creator of the BMS schizophrenia blockbuster, Cobenfy, Seaport has set its sights on two more widespread neuropsychological conditions: major depressive disorder (MDD) and generalized anxiety disorder (GAD). Enabled by their proprietary platform, Glyph, the biotech is seeking to improve the treatment landscape for patients suffering from these disorders by improving oral bioavailability and reducing liver metabolism and hepatotoxicity as well as other unintended side effects. With lead candidate, GlyphAllo (SPT-300), in a Phase 2b trial for MDD and GlyphAgo (SPT-320) in Phase 1 for GAD as well as a steady pipeline of other preclinical assets such as SPT-348, a prodrug of a non-hallucinogenic analog of LSD, the company is well-positioned to be a strong contender in the neuropsychiatric space. Seaport would reportedly list on Nasdaq.
Hemab Therapeutics joins the IPO fray with intentions to also go public on Nasdaq for up to $100M. The company’s lead candidate, sutacimig (HMB-001), is a bispecific antibody aimed at treating Glanzmann thrombasthenia, a rare genetic disease which results in a failure of blood to clot due to a deficiency of the platelet integrin alpha IIb beta3. HMB-001 is presently set to enter Phase 3 trials following successful Phase 2 readouts, an aim to be supported by funds generated from the prospective IPO. The $100M would also be used to finance ongoing Phase 2 trials of the drug for Factor VII Deficiency, a similar bleeding disorder, as well as Phase 1/2 trials of second candidate, HMB-002, a monovalent antibody, for treatment of von Willebrand’s disease and bringing another undisclosed preclinical asset into the clinic later this year.
Avalyn Pharma declares plans to IPO with the same $100M placeholder value. The respiratory-focused biotech will reportedly use funds generated from the public offering to advance three main pipeline candidates, APO1, APO2, and APO3, all of which consist of modifications to existing drugs pirfenidone, a small molecule modulator of cytokines and growth factors, and nintedanib, a small molecule inhibitor of several tyrosine kinases. The former is currently marketed as Esbriet® by Genentech and the latter as Ofev® by Boehringer Ingelheim (with impending generics approval later this month). All three candidates target pulmonary fibrosis and other interstitial lung diseases, aimed at improving existing treatments. Pricing details are not yet disclosed but funds raised will specifically support the progression of APO1 from Phase 2b to Phase 3 clinical trials and likely APO2 into Phase 2 trials as well following Phase 1 results.
Alamar Biosciences files for $159M Nasdaq IPO amidst the ongoing frenzy. The company has developed novel proteomics-based biomarker platforms, Argo HT and NULISAseq, for disease detection—the former being an automated protein analyzer with multiplexing support for biomarker profiling and the latter consisting of an inflammation panel characterizing the body’s natural defenses to various disease states spanning autoimmune, neurodegenerative, cardiovascular, and oncological indications. According to Alamar’s SEC filing, the medtech will offer 9.4M shares at a price of $15-$17 per share.
Neomorph closes $100M Series B led by Deerfield Management. The clinical-stage, San Diego-based biotech is pursuing molecular glue degraders for undruggable protein targets across indications. This most recent fundraising supports Neomorph’s lead, NEO-811, a novel molecular glue degrader, in its ongoing Phase 1/2 trial evaluating its safety, tolerability, pharmacokinetics, and preliminary efficacy for patients with locally advanced or metastatic non-resectable clear cell renal carcinoma (ccRCC). The clinical candidate’s method of action consists of induced targeted degradation of ARNT (HIF-1β) whilst fully blocking a core signalling pathway in ccRCC, which would allow for a significantly improved standard of care. Confidence in the company and its tech has been further bolstered by partnerships with Novo Nordisk, Biogen, and Abbvie who have reportedly collaborated on areas spanning cardiometabolic, rare disease, immunology, oncology, and neurology, highlighting the versatility of Neomorph’s platform. Funding will serve to further advance other assets in Neomorph’s pipeline. Other participating investors in the round included new additions: Regeneron Ventures, Longwood Fund, Alexandria Venture Investments, and Binney Street Capital of the Dana-Farber Cancer Institute among other undisclosed contributors.
Helical raises $10M seed round led by Redalpine. The company has positioned itself as an application layer for biopharma, built on large foundation models and meant to translate them into computationally-powered experiments as a virtual lab. Helical’s goal is ultimately to “close the gap between AI engineers and physical biologists.” Doing so has the potential to vastly shorten drug timelines from discovery to approval, a tantalizing prospect that has already caught the eye of several major pharma players such as Pfizer which already entered into a partnership with Helical in pursuit of novel blood-based biomarkers predicting patient response to gene therapy. Other participants in the round include Gradient, Frst, and BoxGroup in addition to a syndicate of angel investors.
CrossBridge Bio announces agreement to be acquired by Eli Lilly for up to $300M. The six-person team spun out of UT Houston is developing next-generation dual-payload ADCs with a potential best-in-class designation pending clinical approval. Founded in 2023 and still preclinical at the time of this announcement, CrossBridge’s lead candidate, CBB-120, a TROP2-targeting TOP1i/ATRi dual-payload ADC for cancer treatment, is slated for IND application this year. The deal includes both upfront and milestone payments, totaling up to a $300M cash payout for CrossBridge shareholders.
In case you missed it
Novo Nordisk and OpenAI partner to transform how medicines are discovered and delivered
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.