A Cure Designed for One

KJ was six months old when he received an injection that had been designed for him alone. The *CRISPR* *base editors* targeted a single nucleotide error in his genome, a spelling mistake that would otherwise kill him. The therapy was developed in months rather than years, approved through a novel regulatory pathway, and by May, KJ was eating more protein, gaining weight, and requiring less medication. He was the first person in history to receive a gene-editing therapy designed for one patient. There had never been another like it, and there may never be one quite the same. That same year, a global consortium released the complete wiring diagram of an adult fruit fly brain: 139,255 neurons, 50 million chemical synapses, every connection mapped. The first complete *connectome* of an adult brain in a complex organism arrived just 18 months after the larval version. "There is a before and after connectome, B.C. and A.C.," observed Albert Cardona of the MRC Laboratory of Molecular Biology. He meant it seriously. For neuroscience, the coordinates had shifted. The year revealed biology's dual nature: intensely personal medicine for individual patients, and vast collaborative atlases that map life's fundamental architecture. One baby, one injection, one genome. And simultaneously, the entire wiring diagram of a mind. ## The Era of Personalized Gene Editing KJ was born with *carbamoyl phosphate synthetase 1* (CPS1) deficiency, a rare metabolic disorder that prevents the body from processing ammonia properly.[^1] Untreated, the condition is fatal. His mutation was specific to him, a typo in his genetic code that belonged to no one else. No existing therapy targeted it. Researchers at Children's Hospital of Philadelphia and Penn Medicine raced to develop a solution. They designed a base editor to correct the single nucleotide error, tested it in cells and animal models, and won FDA approval through the *Expanded Access* pathway. The timeline from diagnosis to treatment was exceptional for gene therapy, months rather than years. Long-term monitoring continues for off-target editing, immune responses, and durability of effect. Base editing differs from traditional CRISPR's DNA-cutting approach. Rather than creating double-strand breaks that the cell must repair, base editors chemically convert one nucleotide to another: $$\text{C} \rightarrow \text{T} \quad \text{or} \quad \text{A} \rightarrow \text{G}$$ Only these conversions are currently possible: cytosine to thymine, adenine to guanine. The precision reduces unintended insertions and deletions, making base editing safer for therapeutic applications. The success establishes proof of concept for *n-of-1* gene editing, therapies developed for individual patients with rare mutations. The regulatory pathway exists, as does the technical capability. What remains are questions of economics and will, of who pays for a therapy that benefits exactly one person. ## Genetic Medicine Milestones KJ's therapy was designed for one patient. But the same technologies enabling personalized treatment are advancing therapies for conditions affecting thousands or millions. Gene therapies targeting more common conditions advanced to late-stage trials in 2025, with results that hint at what medicine might become. The first genetic medicine for *Huntington's disease* showed striking results, slowing cognitive decline by 75% in trial participants.[^2] Huntington's is caused by a single dominant gene, making it a natural target for gene silencing. It uses *antisense oligonucleotides* (a distinct modality from gene editing) to reduce production of the toxic mutant huntingtin protein. For families who have watched the disease claim generation after generation, a 75% reduction is not a statistic. It is hope with a number attached. A first-in-human trial of CRISPR gene editing for lipid disorders safely reduced both LDL cholesterol and triglycerides in 15 patients with difficult-to-treat conditions. The one-time treatment inactivated PCSK9 and ANGPTL3[^3] via *in vivo* liver-directed editing, genes that raise blood lipid levels. This mimics naturally occurring mutations that protect against heart disease. Some people are born with inactive copies of these genes and enjoy extremely low cardiovascular risk throughout their lives. Now, for the first time, physicians can offer everyone else the same protection. ## Xenotransplantation's Longest Run Gene editing can repair a patient's own genome or redirect the immune system. It can also modify other species to make their organs compatible with human bodies. A pig kidney modified with 69 genetic changes functioned in a human patient for approximately nine months, nearly matching the 1964 record set with a chimpanzee kidney.[^4] The modifications prevent immune rejection and remove *porcine endogenous retroviruses* that could infect human cells. That patient ultimately died, but the kidney functioned until the end. That distinction matters. This demonstrates that *xenotransplantation* may become clinically viable for patients who cannot obtain human organs. With over 100,000 Americans on organ transplant waiting lists and only about 40,000 transplants performed annually, the arithmetic is stark: demand outstrips supply by more than two to one, every year. The 69 genetic modifications included knockouts of three pig genes that trigger acute immune rejection, insertion of seven human genes to modulate the immune response, and inactivation of all porcine endogenous retroviruses. The engineering required multiple rounds of editing: industrial-scale CRISPR, applied with a precision that would have seemed fantastical a decade ago. ## AlphaFold Becomes Infrastructure Xenotransplantation demonstrates what becomes possible when genetic engineering reaches industrial scale: 69 edits to make a pig kidney compatible with a human body. But designing those edits requires understanding how proteins fold, interact, and function. Something subtle but consequential happened in 2025: the tool that predicts protein structure stopped being news. AlphaFold became plumbing. AlphaFold 3, released in 2024, extended protein structure prediction to complexes involving DNA, RNA, and small molecules.[^5] Throughout 2025, the tool transitioned from scientific novelty to pharmaceutical infrastructure. Major pharmaceutical companies integrated AlphaFold into discovery pipelines. Regulatory agencies began accepting computational structure predictions. Hit rates in early-stage screening improved measurably. Drug discovery pipelines now routinely incorporate AlphaFold predictions. *Virtual screening*, which tests millions of candidate compounds against protein targets computationally, depends on accurate structures of binding sites. Where experimental structures are unavailable (and they often are), AlphaFold fills the gap. The days of waiting months for crystallography are ending. AlphaFold 3 predicts not just static structures but how proteins change shape in response to binding events. This *conformational flexibility* matters for drug design: a molecule that fits an inactive receptor may not fit the same receptor in its active state. AlphaFold 3's ability to predict complexes (protein with DNA, protein with small molecule, protein with protein) enables modeling of biological processes rather than isolated components. Limitations remain. *G-protein coupled receptors* (GPCRs),[^6] targets of approximately one-third of FDA-approved drugs, prove challenging. AlphaFold predicts inactive conformations reliably but struggles with the conformational changes that occur upon activation, precisely the states most relevant for drug design. The tool has transformed the field without solving it. ## CRISPR-GPT Accelerates Discovery AlphaFold predicts how proteins fold. But gene editing requires knowing where to cut, how to avoid off-target effects, and which delivery system will reach the right cells. Stanford Medicine researchers developed CRISPR-GPT, a large language model that accelerates gene-editing experiment design by addressing exactly these questions.[^7] CRISPR-GPT makes gene editing more accessible to researchers without deep expertise in *guide RNA* design, lowering a barrier that has kept many labs on the sidelines. It also speeds discovery by automating tasks that previously required manual literature review and computational analysis. It suggests designs. Laboratories still validate experimentally. Acceleration comes from reducing iteration cycles, not eliminating wet-lab work. Its architecture combines transformer language understanding with specialized training on published CRISPR experiments. When a researcher describes a target gene and desired edit, CRISPR-GPT generates ranked suggestions for guide sequences, predicts potential off-target sites, and recommends delivery vectors based on target cell type. Expertise encoded. Knowledge democratized. Whether AI-assisted gene editing raises new safety concerns remains debated. Proponents argue that computational optimization reduces off-target effects. Critics worry that accelerated timelines may introduce errors that slower, more deliberate processes would catch. Both arguments have merit. The tradeoff may be unavoidable. ## mRNA Engineering Matures CRISPR-GPT optimizes the design of gene-editing experiments. But gene editing is not the only technology the pandemic accelerated. mRNA vaccines, developed in months to fight COVID-19, proved that messenger RNA could be manufactured at scale and delivered safely. The platform built in crisis is becoming infrastructure for calm. Wood et al. introduced Helix-mRNA, a foundation model for full-sequence mRNA analysis that processes both coding regions and *untranslated regions* (UTRs).[^8] Previous models focused only on the coding sequence, but UTRs significantly affect translation and stability. The *5' UTR* controls ribosome recruitment. The *3' UTR* influences mRNA half-life and localization. Ignore them, and you're designing with one eye closed. Helix-mRNA uses only 10% of the parameters of existing foundation models while handling sequences 6x longer. The efficiency enables analysis of full therapeutic mRNA sequences, which can exceed 4,000 nucleotides including UTRs and *poly-A tails*. Helix-mRNA predicts expression levels, stability, and immunogenicity from sequence alone: the properties that determine whether an mRNA therapeutic will produce enough protein, persist long enough to be effective, and avoid triggering inflammatory responses. Sequence in, prediction out. The craft of mRNA design is becoming more science, less art. ## Precision Cyclic Peptides mRNA therapeutics instruct cells to produce proteins. But some therapeutic needs require something smaller. *Cyclic peptides* occupy a therapeutic space between small molecules and proteins: small enough to penetrate cells, large enough to bind protein surfaces that small molecules cannot reach. They offer access to *undruggable* targets, the proteins that conventional chemistry cannot touch. Au demonstrated precision design of cyclic peptides using AlphaFold, generating peptides that closely resemble known inhibitors of the HIV gp120[^9] protein. His method constrains AlphaFold to predict cyclic backbone conformations, then optimizes side chains for binding affinity. Cyclic peptides are challenging to design because their constrained geometry limits conformational space. The ring closure requirement means that not all sequences can fold: the backbone must return to its starting point, and not every path leads home. AlphaFold's structure prediction naturally accounts for these geometric constraints. The approach generated candidates for multiple targets beyond gp120, including protein-protein interactions that have resisted conventional drug discovery. Whether computationally designed cyclic peptides prove therapeutically viable awaits demonstration in biological assays. But the design problem, the hardest part, now has a solution. ## Mapping Cells in Space Cyclic peptides, mRNA therapeutics, and CRISPR all target specific molecular components of cells. But cells do not exist in isolation. They organize into tissues, organs, and bodies. Understanding that organization requires knowing not just what genes a cell expresses, but where that cell sits among its neighbors. *Spatial transcriptomics*, the ability to measure gene expression while preserving tissue architecture, reached whole-organ scale in 2025. Allen Brain Cell Atlas released a comprehensive map of the mouse brain containing over 4 million cells classified into 5,322 distinct types, organized hierarchically into 34 classes, 338 subclasses, and 1,201 supertypes. Classification captures not just what genes cells express, but where in the brain those cells reside. Location becomes identity, geography becomes function. Technology combines single-cell RNA sequencing with spatial imaging techniques like *MERFISH*[^10] (Multiplexed Error-Robust Fluorescence In Situ Hybridization). A panel of over 1,100 genes was imaged in approximately 10 million cells across entire adult mouse brains. The result is a coordinate system for the brain. Given any location, the atlas specifies which cell types should be present. Given any cell type, the atlas specifies where it can be found. A separate effort produced a spatial transcriptomic atlas of the macaque *claustrum*,[^11] identifying 48 transcriptome-defined cell types. The claustrum has long puzzled neuroscientists: a thin sheet of gray matter connecting to nearly every cortical area, its function unclear despite decades of study. This atlas provides a molecular foundation for understanding this enigmatic structure: a parts list, at last, for a machine no one fully comprehends. ## The Complete Fly Brain Spatial transcriptomics maps which cells are where and what genes they express. But brains are not just collections of cells. They are circuits. Knowing where neurons sit tells you nothing about how they connect. The FlyWire Consortium, a collaboration of over 200 researchers across 50 laboratories, released something different: the first complete connectome of an adult brain.[^12] Every neuron, every synapse. The wiring diagram contains 139,255 neurons and 50 million chemical synapses in an adult female *Drosophila melanogaster*. For the first time, we possess a complete map of how a mind is wired. This achievement required combining electron microscopy with AI and crowdsourced annotation. Researchers captured images of an entire fly brain at nanometer resolution, sufficient to identify individual synapses. AI algorithms traced neurons through the volume, generating an initial map refined by hundreds of scientists and citizen scientists worldwide. The effort was both heroic and mundane: breakthrough science powered by tireless annotation. Networks display *rich-club organization*, with approximately 30% of neurons highly connected, potentially serving as integrators or broadcasters of signals. Researchers identified subnetworks corresponding to 78 anatomically defined brain regions and characterized information flow between them. It revealed left-right symmetry not just in structure but in connectivity patterns. Information flows through matching circuits on both sides of the brain. Even visual processing, which might be expected to lateralize, showed symmetric organization. Architecture is beautiful in its precision, a blueprint that evolution refined over 500 million years. Why does a fly brain matter? Seventy-five percent of disease-related genes in humans have homologues in the *Drosophila* genome. Understanding how neural circuits compute in a tractable model system informs our understanding of computation in all brains. And the techniques developed for fly connectomics are now scaling toward mammals. Progression is striking. The larval fly connectome (3,016 neurons) was completed in 2023. The adult connectome (139,255 neurons) arrived in 2024. Mouse brain connectomics projects are underway, though the mouse brain contains roughly 70 million neurons, 500 times larger. One detail bears reflection: a fly brain operates on microwatts, fueled by sugar metabolism. AI systems attempting comparable computation require megawatts. We have mapped a mind. We have not yet understood how to build one. ## A Map and a Toolkit That map, combined with the molecular tools developed throughout 2025, provides what biology has long lacked: both the blueprint and the instruments to read it. KJ's personalized CRISPR therapy demonstrates that individualized gene editing is technically feasible and can navigate regulatory pathways in urgent cases. Scaling to rare disease populations requires addressing economic sustainability. Current costs reach millions of dollars per patient. The medicine exists, but the business model does not. Huntington's results suggest that gene therapy for neurodegenerative diseases is possible, though long-term efficacy awaits demonstration. A 75% reduction in cognitive decline, if it persists, would transform the prognosis for one of neurology's most devastating conditions. Families who have watched parents and grandparents disappear into the disease might, for the first time, have reason to hope. AlphaFold's maturation into drug discovery infrastructure represents a quiet revolution, from academic novelty to industrial application. Limitations remain, particularly for flexible proteins and conformational changes, but the baseline capabilities are now routinely useful. What was once a research breakthrough has become an everyday tool. The spatial transcriptomics and connectome atlases provide reference coordinates for biology. Just as genomics provided a parts list, these atlases provide an assembly diagram. The question shifts from "what components exist?" to "how are they arranged?", and the answer is finally coming into view. ## Distances Yet Unmeasured Personalized gene editing may not become economically sustainable for rare diseases. Current costs run in the millions. Thousands of rare diseases affect small populations. The math is brutal, and no business model has emerged to solve it. Gene therapy effects may not prove durable. The genome is edited permanently, but whether therapeutic effects persist requires years of follow-up. We are conducting an experiment whose results will arrive on a generational timescale. AlphaFold's limitations for active conformations may yield to future versions, or may require entirely different approaches. The gap between static structure prediction and dynamic conformational modeling remains significant. Connectomics may not scale to mammals in any reasonable timeframe. The fly brain required years and global collaboration. The mouse brain is 500 times larger. The human brain is another 1,000 times larger still. And even with a complete map, the FlyWire connectome provides wiring, not meaning. Understanding how computation emerges from connectivity remains the central challenge of neuroscience. We can see the wiring. We cannot yet read the code. But for the first time, both the map and the molecular toolkit exist to begin that reading. One baby, one injection, one genome. And simultaneously, the entire wiring diagram of a mind. Biology's dual nature revealed itself in 2025: intensely personal and vastly collaborative, molecular and architectural, therapeutic and fundamental. The year gave us both. The work of understanding continues. --- **Citations**: [1] "World's First Patient Treated with Personalized CRISPR Gene Editing Therapy at Children's Hospital of Philadelphia." CHOP News, 2025. [2] "Seven feel-good science stories to restore your faith in 2025." Nature, December 2025. [3] "First-in-human trial of CRISPR gene-editing therapy safely lowered cholesterol, triglycerides." American Heart Association Newsroom, 2025. [4] "Intriguing science discoveries of 2025." Rockefeller University News, December 2025. [5] "AlphaFold 3 predicts the structure and interactions of all of life's molecules." Google DeepMind, 2024. [6] Chib, G., et al. "Characterizing the Conformational States of G Protein Coupled Receptors Generated with AlphaFold." arXiv:2502.17628, February 2025. [7] "AI-powered CRISPR could lead to faster gene therapies, Stanford Medicine study finds." Stanford Medicine News, September 2025. [8] Wood, M., et al. "Helix-mRNA: A Hybrid Foundation Model For Full Sequence mRNA Therapeutics." arXiv:2502.13785, February 2025. [9] Au, C.S. "Precision Design of Cyclic Peptides using AlphaFold." arXiv:2510.13127, October 2025. [10] "A high-resolution transcriptomic and spatial atlas of cell types in the whole mouse brain." Nature, 2023; Allen Brain Cell Atlas updates, 2024-2025. [11] "Single-cell spatial transcriptome atlas and whole-brain connectivity of the macaque claustrum." Cell, April 2025. [12] Dorkenwald, S., et al. "Neuronal wiring diagram of an adult brain." Nature, October 2024. **Footnotes**: [^1]: CPS1 deficiency affects approximately 1 in 800,000 newborns. Without treatment, ammonia accumulates to toxic levels, causing brain damage and death. The urea cycle, which CPS1 initiates, is the only pathway for excreting nitrogen from protein metabolism. [^2]: Huntington's disease is caused by an expanded CAG repeat in the huntingtin gene. Normal copies contain 10-35 repeats; disease-causing copies contain 36 or more. Gene silencing approaches aim to reduce production of the toxic mutant protein. [^3]: PCSK9 and ANGPTL3 are genes that regulate cholesterol and triglyceride metabolism. Inactivating these genes mimics naturally occurring mutations found in individuals with extremely low cardiovascular disease risk. [^4]: The 1964 chimpanzee kidney transplant was performed by James Hardy. The recipient survived nine months before dying of causes unrelated to the transplant. Ethical concerns have since ended primate organ transplantation research. [^5]: AlphaFold 2 predicted only single protein structures. AlphaFold 3 can predict complexes of proteins with DNA, RNA, ions, and small molecules, enabling modeling of biological interactions rather than isolated components. [^6]: GPCRs are seven-transmembrane receptors that change conformation upon ligand binding. The active (ligand-bound) conformation is often the therapeutically relevant one, but also the most difficult to predict computationally. [^7]: CRISPR-GPT was trained on published CRISPR experiments and uses the transformer architecture underlying modern language models. The system generates guide RNA suggestions and predicts off-target effects. [^8]: UTRs (untranslated regions) flank the coding sequence of mRNA and regulate translation initiation, mRNA stability, and subcellular localization. The 5' UTR controls ribosome binding; the 3' UTR influences degradation rate. [^9]: gp120 is the HIV envelope glycoprotein that binds to CD4 receptors on human immune cells, initiating infection. Blocking this interaction could prevent HIV entry into cells. [^10]: MERFISH (Multiplexed Error-Robust Fluorescence In Situ Hybridization) allows simultaneous detection of hundreds to thousands of RNA species in intact tissue sections, preserving spatial information that is lost in conventional sequencing. [^11]: The claustrum is a thin, irregular sheet of gray matter beneath the insular cortex. Francis Crick (of DNA fame) proposed it might integrate information across sensory modalities, but its function remains debated. [^12]: The FlyWire Consortium used petabyte-scale electron microscopy data processed by AI algorithms and refined by crowdsourced human annotation. The full dataset is freely available for research use.