Precision Medicine

Established Precision medicine is the proposal that treatment should be matched to the individual rather than to the diagnosis, and after three decades the measured record splits cleanly along one seam. Where the task is to identify in advance who a drug will harm, the field works: a twelve-gene pharmacogenomic panel cut clinically relevant adverse drug reactions from 28.6 per cent to 21.5 per cent across 6,944 patients in seven countries, an odds ratio of 0.70. Established Where the task is to identify who a drug will help, it mostly does not: the best-argued sceptical accounting puts end-to-end benefit from matched cancer therapy at about 1.5 per cent of patients with relapsed and refractory solid tumours. Frontier That asymmetry, rather than the enthusiasm or the backlash, is the actual subject.

Established The oncology funnel is the cleanest case anywhere of one dataset being quoted to opposite effect. NCI-MATCH's own framing is that nearly two in five patients — 38 per cent — may have a candidate treatment revealed by next-generation sequencing. Established MD Anderson matched 6.4 per cent of sequenced patients to a targeted drug, NCI-MATCH itself had matched 2 per cent as of May 2016, and roughly 30 per cent of matched patients respond at all; multiply the chain and you get about 1.5 per cent. Established All four numbers are correct and all four describe different points on the same funnel, which is why a brief printing only 38 or only 1.5 has misrepresented one dataset in opposite directions.

Established Two structural facts govern everything downstream. Polygenic prediction is several times more accurate in people of European ancestry than in anyone else, and the applied gap is measured rather than rhetorical: 0.53 against 0.23 years of education for the same selection procedure. Established And six approved genetic medicines list at $2.125 million to $4.25 million, against which the flagship CRISPR therapy had infused 39 people worldwide roughly 22 months after approval. Speculative Whether individualised medicine can exist at population scale, or is structurally a luxury good, is the question those numbers actually pose.

Established The archetype carries its own failure rate, and the failure rate is the larger number. Trastuzumab was approved by the FDA in September 1998 and in the EU in August 2000, conditional on HER2 testing by immunohistochemistry or fluorescence in situ hybridisation, and HER2 is amplified in 20 to 30 per cent of early-stage breast cancers. Established In metastatic disease it moved median overall survival from 20.3 to 25.1 months; in early-stage disease the hazard ratio is 0.66 for overall survival and 0.60 for disease-free survival, with absolute risk reductions of 3 per cent for three-year death and 9.5 per cent for three-year recurrence. Established The part usually dropped is that about 70 per cent of HER2-positive patients do not respond to the drug their biomarker selected them for, and resistance develops in virtually all of those who do. Frontier A companion diagnostic that is right about the target and wrong about the patient seven times in ten is this field's best case, not its worst, and an honest account starts there.

Established Companion diagnostics are a regulated device class rather than a research concept, which is why this half of the field is durable. The FDA classifies them as medical devices and the EU In Vitro Diagnostic Regulation defines them as devices essential for the safe and effective use of a corresponding medicinal product; EGFR testing gating gefitinib and erlotinib follows the HER2 template exactly. Established The total number of FDA-approved companion diagnostics could not be verified here, and no figure for it is printed.

Established Pharmacogenomics produced the strongest randomised evidence the field has, and it is evidence about avoidance. PREPARE was open-label, multicentre and cluster-randomised with crossover across seven European countries, enrolling 6,944 patients — 3,342 to genotype-guided prescribing and 3,602 to standard care. Established Clinically relevant adverse drug reactions occurred in 628 of 2,923 evaluable patients under genotype guidance, 21.5 per cent, against 934 of 3,270, or 28.6 per cent, under standard care: an odds ratio of 0.70, confidence interval 0.61 to 0.79, p < 0.0001, which is roughly a 30 per cent relative reduction in harm from a twelve-gene panel. Frontier The design is open-label and the primary outcome partly patient-reported, and the journal that published it published a simultaneous commentary arguing that the benefits of pharmacogenetic testing are unclear; that commentary was not obtained for this assessment and no claim is drawn from it, but its existence belongs in the record.

Established The gene–drug pairs that change practice are few, specific, and mostly about poisoning people less. CYP2D6 ultra-rapid metabolisers converting codeine to morphine faster than expected, implicated in fatal reactions in children after tonsillectomy and adenoidectomy; VKORC1 and warfarin dosing; CYP2C19 and clopidogrel activation; G6PD deficiency and oxidative drugs; EGFR genotype gating the first-generation tyrosine kinase inhibitors. Established And the negative that the pharmacogenomics literature states about itself: sufficient evidence does not at this time exist to validate the cost-effectiveness of testing. Frontier The field's best-evidenced intervention is proven effective and explicitly not proven worth paying for, a distinction almost never preserved downstream.

Established Precision oncology at platform scale produced four numbers, all correct, describing one funnel. NCI-MATCH screened nearly 6,000 patients, assigned 1,593 to treatment across 38 substudies of which 27 reported, and returned 7 positive arms — a 25.9 per cent success rate by arm alongside a 25.9 per cent average response rate; it ran eight years and closed in 2023, and its own headline is that nearly two of five patients may have a candidate treatment revealed by sequencing. Established Vinayak Prasad's counter-accounting in Nature gives the other end: 6.4 per cent of sequenced patients matched to a targeted drug at MD Anderson, 2 per cent matched within NCI-MATCH as of May 2016, roughly 30 per cent of matched patients responding, median progression-free survival of 5.7 months, and a product of about 1.5 per cent end-to-end benefit for relapsed and refractory solid tumours. Established SHIVA, the first randomised test of molecularly guided selection, returned 2.3 months median progression-free survival against 2.0 for physician's choice. Established The reconciliation is not a compromise: 38 per cent is the actionable-finding rate, 6.4 and 2 per cent are the treated-on-a-matched-arm rates, 25.9 to 30 per cent is the response rate among those treated, and 1.5 per cent is the chain multiplied out.

Established The largest randomised test of population-scale precision screening missed its primary endpoint. NHS-Galleri randomised 142,000 asymptomatic participants aged 50 to 77 to a cell-free-DNA multi-cancer early detection test; the primary endpoint, a reduction in combined stage III–IV diagnoses over three years, was not met, and the stage IV reduction by the third screening round was 26 per cent with a confidence interval crossing 1.0. Established The assay itself performed extraordinarily: 30.7 per cent of all cancers and 54.7 per cent of the twelve pre-specified cancers detected within twelve months, 99.55 per cent specificity, 52.0 per cent positive predictive value against roughly 6 per cent for an NHS urgent cancer referral, and 92.5 per cent accuracy in calling the cancer signal of origin. Established The sponsor's release, headlined on a four-fold higher detection rate and a substantial reduction in stage IV diagnoses, states the missed primary endpoint in its body; Professor Richard Houlston's independent assessment is that the findings were presented far more positively than the overall results justify, and that there is no evidence base upon which to justify implementation at a population scale. Frontier Both readings survive the data, which is the point: this is an excellent test that did not, over three years, move the thing screening exists to move.

Established Polygenic risk scores discriminate at the population level and barely at the individual one. Area under the curve from genotype alone runs about 0.64 for coronary disease, 0.63 for breast cancer and 0.71 for both hypothyroidism and schizophrenia; adding age and sex lifts coronary prediction to roughly 0.80, which is to say most of the clinically useful discrimination is not genomic. Established For schizophrenia the score explains 7.3 per cent of liability, with an odds ratio of 39 comparing top and bottom centiles and 5.6 comparing the top centile against everyone else — and those two ratios are the whole story, because the dramatic contrast is the one no clinician uses. Established Applied to embryo selection, Turley and colleagues put the absolute risk reduction from choosing the top-scoring of ten embryos at 5.5 percentage points for type 2 diabetes and 1.9 for breast cancer, while the largest relative reduction in the set, 77 per cent for idiopathic short stature, buys about 2.5 cm of height.

Established The ancestry gap is not a caveat attached to polygenic scores; it is a property of them. Martin and colleagues stated it in 2019 from inside the field that built the scores: polygenic risk scores are several times more accurate in individuals of European ancestry than in other ancestries, because discovery cohorts, linkage-disequilibrium structure and allele frequencies are all European-weighted. Established The applied consequence is measured — the same selection procedure on ten embryos yields an expected 0.53 years of education under European weights and 0.23 under African weights, with non-European groups 25 to 65 per cent worse across every condition studied. Established The same failure mode arrives by a second, non-genomic route: the Framingham Heart Study risk equations produce biased outcomes applied to non-white populations, which suggests the problem is reference-cohort composition rather than anything peculiar to genomics. Frontier A predictor 25 to 65 per cent worse for most of the world is not a tool with an equity problem attached; it is a tool built for part of the world.

Established Where precision medicine has unambiguously worked, the pattern is identical every time: one gene, one companion genotype, no score. Casgevy, approved 8 December 2023, edits the erythroid-specific enhancer of BCL11A to de-repress fetal haemoglobin and does not correct the sickle mutation at all. Frontier DB-OTO, replacing otoferlin in congenital deafness, improved hearing in 11 of 12 children. Established The approved antisense oligonucleotides — nusinersen, tofersen, eteplirsen and successors — follow the same shape, and milasen, built for a single patient, is the limiting case of it. Speculative All are monogenic and none involves a polygenic score, so the honest summary of thirty years is that precision medicine works where Mendelian genetics already worked, and that the extension to common polygenic disease is the part not yet demonstrated.

Established And where it works, it is priced beyond the systems meant to deliver it. US list prices as of 30 November 2025 run Lenmeldy $4.25 million, Hemgenix $3.5 million, Elevidys $3.2 million, Skysona $3 million, Zynteglo $2.8 million and Zolgensma $2.125 million; Casgevy's US list price is widely quoted, was not verified for this assessment, and is not stated here. Established Against those prices, sponsor figures at 30 September 2025 record roughly 165 patients having completed a first cell collection, 39 actually infused, about 300 referred and 25 authorised treatment centres having initiated more than five patients each — against a stated eligible US population of about 16,000 people with severe sickle cell disease. Frontier Roughly 22 months after the first CRISPR approval anywhere, 39 people had been dosed, and the binding constraint was not the editing but autologous cell collection, busulfan conditioning, transplant-capable centres and payment.

Established The national substrates are real, large, and built for exactly the deficit above. The UK's 100,000 Genomes Project ran from December 2012 through 13 Genomic Medicine Centres recruiting across 85 NHS Trusts on over £300 million of investment, reaching its milestone in December 2018; the rare-disease diagnostic yield routinely quoted for it could not be verified here and is not printed. Established All of Us reported, at 30 June 2026, more than 883,000 enrolled participants, over 535,000 whole genomes linked to roughly 482,000 electronic health records, 1.3 billion variants, more than 14,500 long-read genomes, and 645,000 participants — 86 per cent — from historically underrepresented groups, with 733,000 personalised DNA results returned to 277,000 participants; its 2024 Nature paper reported 245,388 clinical-grade genomes, 77 per cent underrepresented, and more than 275 million previously unreported variants. Frontier The input problem behind ancestry portability is therefore being fixed, on a visible schedule, by one programme. The response-phenotype problem is being fixed by nobody.

Established The asymmetry this brief finds in genomics reappears in clinical AI, and the headline number quoted for it measures the wrong thing. The FDA’s public list of authorised AI-enabled medical devices has passed a thousand entries and keeps growing, the large majority of them in radiology. That list counts authorisations, not effects. Almost all of those devices cleared through the 510(k) pathway, which requires substantial equivalence to a marketed predicate rather than a demonstration of clinical benefit, so the count measures regulatory throughput and predicate-chain depth. Established An analysis of 130 cleared AI devices found 126 evaluated retrospectively only, none of the 54 classed as high risk evaluated prospectively, and most reporting neither multi-site assessment nor a sample size. Frontier Authorisation count and evidence base are two quantities that have grown at different rates, and citing the first as though it were the second is precisely the error this brief documents in the oncology funnel.

Established Where randomised evidence exists it measures workflow rather than diagnosis, and that turns out to be the useful measurement. The Swedish MASAI trial randomised mammography screen-reading to AI-supported single reading against standard double reading across roughly 80,000 women and reported about 20 per cent more cancers detected without a rise in recall rate, at screen-reading workload down about 44 per cent. Frontier The workload figure has the clearer mechanism: the intervention replaced one of two human readers, so the effect is a labour substitution with a detection side-effect rather than a new diagnostic capability. Established Screening mammography is also the most favourable case available — one image type, a defined population, an existing double-reading baseline to substitute for, and outcomes ascertained by a registry that already exists. Handwave None of it transfers to an unstructured clinical setting without being retested there.

Established The measured failures are post-market, and they are large. External validation of a widely deployed proprietary sepsis prediction model, running at hundreds of hospitals, found an area under the curve of about 0.63 against the vendor’s reported 0.76 to 0.83, identified only a small minority of sepsis cases not already recognised clinically, and generated alerts on roughly one in five hospitalised patients. Established A commercial risk-stratification algorithm applied to tens of millions of people was shown to use health cost as a proxy for health need, so that patients of one racial group had to be substantially sicker to reach the same score; correcting the target variable would have more than doubled the share referred to additional care. Neither failure was found by a regulator, a manufacturer or a surveillance system. Both were found by academics who went looking.

Established The most consequential result of the decade in this field is a negative at n = 142,000, and it points the wrong way. Multi-cancer early detection is a settled negative on its primary endpoint, with real secondary signals attached: stage IV diagnoses down by around a quarter in later rounds, a four-fold detection rate, a 52 per cent positive predictive value. Frontier Whether those secondaries mature into mortality benefit over a longer horizon is genuinely open, and the trial as designed cannot answer it. Speculative The structural reading is sharper than either camp's: a test at 99.55 per cent specificity needed an age-enriched population to reach a defensible predictive value, and at lower prevalence the same test produces mostly false positives — which makes multi-cancer early detection a risk-stratified technology rather than a population one.

Frontier In vivo cell therapy collapses the personalisation timeline from months to days, on four patients. Kelonia's KLN-1010 went from consent to infusion in 13 to 18 days with no apheresis, no lymphodepletion and no ex vivo manufacturing, reporting 4 of 4 partial responses at month one and 4 of 4 negative for measurable residual disease, first dosed 19 August 2025. Frontier That is n = 4 and should be read as a demonstration of a manufacturing route rather than of efficacy. Speculative If it holds, the entire industrial argument in this brief — 25 centres, busulfan conditioning, 39 infusions — is a transitional problem rather than a structural one.

Frontier N-of-1 therapy has two documented instances and both are regulatory precedents more than scientific ones. Milasen, in 2019, was a customised antisense oligonucleotide built for one patient's CLN7 Batten mutation under an FDA expanded-access protocol, and its authors stated plainly that it is not suited for the treatment of other patients with Batten's disease. Established In 2025 a patient-specific adenine base editor was designed, manufactured and dosed for KJ Muldoon's CPS1 deficiency in roughly six months against an eighteen-month estimate, delivered as three doses over seven weeks; the load-bearing fact is that the FDA permitted a patient-specific three-dose protocol at all. Speculative If the unit of precision medicine becomes one patient rather than one biomarker-defined subgroup, the prediction problem dissolves — you are no longer predicting response, you are correcting a known cause.

Frontier Base editing has produced the first in-human dose–response curve for a permanent genomic edit. VERVE-102 reported mean LDL-cholesterol reductions of 21, 41 and 53 per cent across three dose cohorts, with a maximum individual reduction of 69 per cent. Speculative A titratable, individually dosed, permanent edit is a new kind of object — not a drug you can stop, and not a one-shot binary either — and nobody has a regulatory or ethical framework for an irreversible intervention whose magnitude is chosen at the moment of dosing.

Established The safety denominator is now large enough to see rare events, which is itself the frontier. Intellia's nex-z had dosed more than 450 of roughly 650 planned patients in MAGNITUDE when a grade 4 hepatotoxicity occurred in a patient dosed on 30 September 2025; the FDA placed both phase 3 trials on hold on 29 October and the patient died on 5 November 2025, against a grade 4 transaminase rate below 1 per cent. Frontier A sub-1-per-cent event is invisible below a few hundred patients and definitive above them, so in vivo editing safety has just crossed from assumed to measurable.

Frontier Long-read sequencing has reached biobank scale, and structural variation is where the missing monogenic yield is expected to sit. More than 14,500 long-read whole genomes in All of Us is enough to begin systematic structural-variant discovery in a genuinely diverse cohort. Speculative That a material fraction of undiagnosed rare disease is structural rather than single-nucleotide is plausible, widely assumed, and untested at this scale.

Established The largest population-scale experiment in returning genomic information to people is running with no outcome arm. All of Us has returned 733,000 personalised DNA results to 277,000 participants. Speculative No evidence was obtained for this assessment about whether receiving a result changes behaviour, screening uptake or outcomes, and a cluster-randomised comparison of return against non-return at biobank scale would be cheap relative to everything else described here. Frontier Nobody appears to be running one.

Frontier Meanwhile polygenic scores are sold to consumers for embryo selection ahead of any clinical-utility trial. Orchid advertises screening for more than 400 genetic conditions including autism, Alzheimer's, ALS and schizophrenia, and claims to read 99 per cent of an embryo's genome, without publishing sensitivity or specificity. Speculative The commercial frontier of this field has moved past the evidentiary frontier, in the direction the evidence does not support.

Frontier The open question is whether a model authorised once can be known to still work later, and the regulatory answer is now a change-control document rather than a measurement. The predetermined change-control plan lets a manufacturer specify in advance the modifications it may make without a new submission, converting model updating from a licensing event into a documented process. Established That is a real improvement in the update problem and does nothing about the drift problem, because drift is a change in the population rather than in the model: frozen weights degrade when case mix, coding practice, instrument or referral pattern move underneath them. Frontier No jurisdiction requires a deployed clinical model to report its own performance against outcomes on a schedule, so the field has a change-control regime and no performance-monitoring regime, and the two are routinely discussed as one thing.

Frontier Automation bias is the second open question and the one most likely to be underestimated. Reader studies have found that an incorrect machine suggestion degrades human performance rather than being filtered out, with the effect largest in less experienced readers — which inverts the usual deployment argument that assistance is most valuable where expertise is thinnest. Speculative If that inversion holds at scale, the highest-value deployments and the highest-risk deployments are the same deployments. Frontier The WHO guidance on large multimodal models names automation bias and deskilling explicitly and recommends evaluation before deployment rather than after. Handwave Guidance is not a mechanism, and no measurement of deskilling across a clinician’s career exists in any published series.

Established The workback chain to a precision medicine that actually predicts individual response has seven links, and the binding one is not genomic. Link 1 is a predictor with individual-level discrimination: area under the curve above 0.85 for a common disease from genotype plus routine clinical data, validated out of sample and across ancestries, against a current best of roughly 0.80 for coronary disease with age and sex and 0.64 from genotype alone. Established Link 2 is ancestry portability, requiring association studies at 100,000 or more cases in each of at least four ancestry groups; All of Us at 535,000 genomes and 86 per cent underrepresented participants is the first plausible substrate, so this link is a supply problem currently being solved.

Frontier Link 3 is response phenotyping, and it is the binding constraint. What a response predictor needs is health-record-linked, structured treatment-response data on the order of 100,000 patients per drug class — who was given what, and what happened. Established That dataset does not exist for any drug. Frontier Everything upstream of it — sequencing cost, cohort size, statistical method, ancestry coverage — is either solved or on a funded trajectory, so the input the whole enterprise is missing is not genetic. It is a large, structured, comparable record of who responded to what, and assembling it is an institutional and data-collection problem that no genomics budget line pays for.

Speculative Links 4 and 5 separate individual prediction from subgroup assignment. Link 4 is within-person longitudinal data — repeated multi-omic sampling on the same individuals over time, the only design that can beat a subgroup mean for a named person — and neither the cost base nor the industrial capability for it exists at cohort scale. Established Link 5 is clinical informatics that can act on a result at the decision point, and this is historically where the field has failed: the largest attempt cost MD Anderson $62 million and was stopped.

Established Links 6 and 7 are market constraints, and they bind on delivery even where the science is finished. No national programme has demonstrated net cost reduction from precision medicine, and the pharmacogenomics literature says its own cost-effectiveness is unvalidated. Established Reimbursement for a one-time product priced at $2 to $4 million requires outcome-based contracting at national scale, and the state of the art is 39 infusions in 22 months. Frontier A therapy that is approved, effective and unpurchasable is a bottleneck of a different kind from an unsolved experiment, and the two are routinely conflated in both directions.

Established The binding constraint on clinical AI is the same bookkeeping absence that binds the rest of this brief. The field cannot say which patients a deployed model helped, because nobody records the counterfactual, and the adverse-event machinery that does exist was built for devices that break rather than for models that degrade. A model quietly losing calibration produces no event to report, so it generates no entry in a malfunction database and no signal to any regulator. Frontier Post-market surveillance for clinical software reduces in practice to complaints, academic validations and the manufacturer’s own reporting — in that order of volume and roughly the inverse order of reliability.

Frontier Liability is the next bottleneck, and it is unallocated rather than contested. A clinician who follows a model and is wrong, and a clinician who overrides it and is wrong, face a standard of care that no court has settled and no professional body has written. Speculative The predictable equilibrium is defensive use: the model is consulted, recorded, and then agreed with, which converts a decision-support tool into a decision-making one without anyone deciding that it should. Handwave That is a prediction from incentives rather than an observation, and no published series measures override rates against outcomes at scale.

Established This subject consumes results other fields produce for their own reasons. From statistical genetics it needs ancestry-diverse discovery cohorts at case counts no single national programme has yet reached in more than one population. Established From health informatics it needs structured phenotyping of treatment response, which is a record-keeping problem rather than a scientific one and is therefore funded by nobody. Frontier From delivery science it needs editing chemistry that reaches tissue other than liver, since every in vivo success named in this brief is hepatic, and the diagnostic half of precision medicine can only cash out through the therapeutic half. Speculative None of those three is a result this subject could produce for itself, and none of the three fields producing them has this subject's needs as its objective, which is the ordinary condition of an applied discipline and worth stating rather than assuming.

Established Two dependencies are shared with neighbouring subjects and visible in their records. Non-genotoxic conditioning would simultaneously lift the throughput ceiling on autologous products and address the insertional-oncogenesis signal seen in one approved therapy; that is a stem-cell-transplant problem rather than a genomics one, and it is the same constraint Genetic Engineering identifies as binding on ex vivo delivery. Frontier Structural-variant calling from long reads at biobank scale is the dependency most likely to convert undiagnosed rare disease into diagnosed rare disease, and it is arriving as a byproduct of sequencing economics rather than as a clinical programme.

Speculative The dependency nobody lists is a theory of which diseases are predictable at all. The field proceeds as though prediction accuracy is uniformly improvable with more data, but a liability-threshold trait with substantial environmental and stochastic variance may have a ceiling well below clinical usefulness that no cohort size can raise. Handwave There is no accepted method for estimating that ceiling in advance for a named condition, which means the field currently cannot tell a hard problem from an underfunded one, and spends accordingly.

Established Six measurements would settle most of what is contested here, and four of them are cheap. Frontier First: an out-of-sample, cross-ancestry validation of any polygenic predictor reaching an area under the curve above 0.85 for a common disease from genotype plus routine clinical variables. Nothing in this brief's sources approaches it, and the distance from 0.64 to 0.85 is the distance between a research instrument and a clinical one.

Speculative Second: a cluster-randomised trial of returning polygenic risk information to patients. All of Us has already returned 733,000 results to 277,000 participants without a control arm, and a return-against-non-return design at biobank scale, with screening uptake, behaviour change and downstream diagnosis as endpoints, would cost a rounding error against the sequencing budget. Frontier Its absence, given the number of results already returned, is the most striking evidentiary gap in the field.

Frontier Third: a blinded or registry-controlled replication of PREPARE. The trial is open-label with a partly patient-reported primary outcome and remains the strongest randomised result precision medicine has. Frontier A replication with adjudicated, blinded outcome ascertainment would either consolidate the one clearly positive intervention in the field or reveal it as reporting bias, and both answers are worth having.

Frontier Fourth: multi-cancer early detection in an enriched population, powered for mortality rather than stage shift. The 142,000-person trial was powered on stage distribution and missed. Speculative The arithmetic of positive predictive value says such a test will only ever be defensible where prevalence is elevated, so the next trial should stratify by risk before screening — and the risk stratifier available is a polygenic score at an area under the curve of 0.64, which is the circularity the field has not confronted.

Speculative Fifth and sixth are harder. A within-person longitudinal multi-omic cohort with repeated sampling and recorded treatment response is the only design that can demonstrate individual rather than subgroup prediction, and nobody in this brief's sources has run one at useful scale. Speculative And a national programme reporting audited net cost, rather than cost per quality-adjusted life year in a model, would settle the pharmacoeconomic question the field's own literature says is open.

Established The engineering constraint on delivered precision medicine is cell manufacturing, and it is unglamorous. An autologous ex vivo product requires, per patient, apheresis, cell collection, shipment, editing, expansion, release testing, return shipment, myeloablative conditioning and transplant-capable inpatient care. Established Twenty-five authorised centres had initiated more than five patients each roughly 22 months after the first CRISPR approval, and 39 patients had been infused. Frontier Scaling that to a stated eligible population of about 16,000 in the United States alone is a factor of several hundred in throughput, and no part of it is a genomics problem.

Frontier Two engineering routes bypass the bottleneck rather than widening it. In vivo delivery removes manufacturing entirely, at 13 to 18 days from consent to infusion with no apheresis and no lymphodepletion; non-genotoxic conditioning would remove busulfan, which is simultaneously the throughput limit and the plausible driver of the malignancy signal in one approved product. Speculative Either would change the deliverability of this field more than any conceivable improvement in prediction accuracy, which is an uncomfortable thing for a field organised around prediction.

Established The other engineering system is informatics, and it has one very expensive negative result. IBM Watson for Oncology launched with Manipal Hospitals on 29 July 2016 and Manipal discontinued in December 2018; the MD Anderson pilot begun in 2013 was stopped after $62 million without meeting its goals; IBM sold Watson Health to Francisco Partners in January 2022 for roughly $1 billion, completing that June, and a 2019 analysis called the episode a cautionary tale of hubris and hype. Established The precision-medicine literature separately notes data-processing errors on the order of 30,000 per processed human genome, insufficient physician training, and the need for an end-to-end change in implementation. Speculative Together those support a specific hypothesis: that the binding constraint is curation, integration and workflow rather than biology or machine learning, which would mean the field has spent a decade optimising the wrong layer.

Established The nearest neighbour is delivery, and it sets this subject's ceiling. Every therapeutic success named here is a genome-editing or oligonucleotide product, so the manufacturing, conditioning and delivery constraints assessed in Genetic Engineering apply in full, and the n-of-1 route in particular is a genetic-engineering capability wearing a precision-medicine label. Frontier The traffic runs both ways: a companion genotype is what makes an editing product approvable, so the diagnostic half of this subject is the gate the therapeutic half passes through, and the two halves fail together. Established The autologous manufacturing chain that limits an approved editing therapy to 39 patients is the same chain that would limit any diagnostic-led therapy to the same throughput, whatever the prediction accuracy upstream of it.

Established The polygenic half shares its statistical machinery, and its failure modes, with behaviour genetics. Within-family analysis shrinks population effect estimates by 47 per cent for educational attainment and 22 per cent for cognitive ability against 10 per cent for height, and that shrinkage applies to any clinical score built from population associations; the machinery and the correction are assessed in Neurogenetics. Frontier A clinical polygenic score has the same architecture as an educational one and has generally not had the same audit performed on it, which is the most portable criticism in this brief.

Frontier Three further adjacencies are load-bearing. Targeted delivery systems assessed in Nanomedicine are the route to in vivo editing outside the liver, which is the single technical fork deciding whether the manufacturing constraints above are permanent. Frontier The biomarker-and-intervention architecture examined in Longevity Therapies faces an identical response-phenotyping deficit and would be unblocked by the same dataset. Speculative And the selection technologies discussed in Biological Enhancement use the same polygenic scores, at the same accuracy, with the same ancestry gradient, for a different purpose entirely — which means an advance in clinical scoring is automatically an advance in selection, whether or not anyone intended it.

Established Three institutional facts, not three scientific ones, decide whether any of this reaches patients. The first is payment for a one-time multi-million-dollar product: six approved genetic medicines list between $2.125 and $4.25 million, and health systems are built to pay recurring costs from recurring budgets, which a single-administration cure inverts. Frontier Outcome-based contracting exists in pilot form and not at national scale, which is a substantial part of why 39 people had received the flagship CRISPR therapy 22 months after approval against an eligible US population of roughly 16,000.

Established The second is the composition of reference cohorts, which is a procurement decision that determines who the medicine works for. Polygenic scores are several times more accurate in European-ancestry individuals, applied gains fall 25 to 65 per cent outside that group, and the same pattern appears in non-genomic risk equations built on non-representative cohorts. Established All of Us is the counter-example and it was built deliberately: 86 per cent of its 645,000 participants come from historically underrepresented groups, and its 2024 genomic paper reports 77 per cent underrepresented across 245,388 clinical-grade genomes. Frontier That is what fixing this looks like — a decade, a national programme and an explicit recruitment mandate — and it has been done once.

Established The third is regulatory architecture for products with a denominator of one. Milasen was authorised under expanded access and the CPS1 base editor was permitted as a three-dose patient-specific protocol, but neither is a licensing pathway. Speculative What an n-of-1 field would need is platform approval — regulating the manufacturing process and the design rules rather than the product — and no agency has established one. Frontier Meanwhile the least regulated corner of the field is the consumer one: polygenic embryo screening is sold directly to prospective parents with no published sensitivity, specificity or clinical-utility trial, and no companion-diagnostic-style device review.

Established One institutional gap is worth naming precisely because it is boring. The pharmacogenomics literature states that evidence sufficient to validate the cost-effectiveness of testing does not exist. Frontier A health system cannot commission at scale against an unvalidated economic case however good the odds ratio, and the twelve-gene panel that produced this field's best randomised result is therefore standard practice nowhere.

Established The equity problem here is unusual in that it is quantified. Most arguments about fairness in medicine concern access; this one has a number for how much worse the instrument itself performs, and that number is 25 to 65 per cent across every condition studied. Frontier Deploying a predictor with that gradient into a mixed population does not merely fail to help some people, it redistributes screening attention and clinical suspicion toward the ancestry group the score was trained on.

Established Screening ethics turn on the other half of the predictive value. A 52 per cent positive predictive value in an age-enriched trial population means roughly half of everyone told they may have cancer does not, which is far better than the roughly 6 per cent of an NHS urgent cancer referral and is still a coin flip delivered as a health alarm, followed by diagnostic work-up. Frontier In a lower-prevalence population the ratio worsens, and the harms — anxiety, invasive follow-up, overdiagnosis of indolent disease — scale with the number screened, while the benefit, on the trial evidence, did not reach its endpoint.

Frontier Consumer polygenic screening raises a distinct problem: irreversibility without evidence. Embryo selection on scores for autism, schizophrenia or Alzheimer's is sold before any clinical-utility trial, and the decision it informs cannot be revised later. Established The peer-reviewed estimate of what such selection buys is an absolute risk reduction in the low single percentage points for most conditions, and about 2.5 cm of height. Speculative A market selling irreversible choices on that evidence base is a regulatory failure rather than an ethical dilemma, and treating it as the latter has been convenient for everyone involved.

Established And there is the distributional question the price list poses directly. A $4.25 million therapy delivered through twenty-five specialist centres is available, in practice, to people living near those centres with insurers who will pay. Speculative Precision medicine's characteristic harm may not be discrimination by algorithm at all, but the quieter one of a two-tier system in which the individualised arm is real, effective, and reachable by very few.

Speculative The strongest structural claim against this field is that its economics are permanently adverse rather than temporarily immature. Tests are cheap and the drugs they select for are not, so identifying responders raises cost per treated patient while lowering population cost only if the non-responders would otherwise have been treated — which holds for expensive oncology drugs and fails for most of medicine. Established No national programme has demonstrated net cost reduction, and the field's own literature says the cost-effectiveness case is unvalidated. Speculative If that is structural rather than transitional, individualised medicine is not a phase of medicine but a segment of it.

Speculative The counter-case is that the unit cost of individualisation is falling fast enough to invert the argument. A bespoke base editor was designed and dosed in six months; an in vivo cell therapy went from consent to infusion in 13 to 18 days. Speculative If per-patient manufacturing cost falls by two orders of magnitude and regulators approve processes rather than products, the marginal cost of the thousandth bespoke therapy approaches the cost of the first, and the luxury-good argument collapses. Handwave Nothing in the fetched record demonstrates either condition; both are coherent, and both are assertions.

Speculative Declare the tie, because the evidence genuinely supports both terminal positions. Against: 1.5 per cent end-to-end benefit in relapsed solid tumours, a failed 142,000-person screening trial, a best-in-field 30 per cent reduction in adverse reactions that cannot be shown cost-effective, and six approved genetic medicines at $2.125 to $4.25 million of which the flagship has treated 39 people. Established For: trastuzumab moved metastatic survival from 20.3 to 25.1 months and has been standard of care for 27 years, PREPARE cut clinically relevant adverse reactions by 30 per cent across seven countries, and 11 of 12 congenitally deaf children now hear. Frontier Both lists are correct, and the honest verdict is that this is neither a revolution nor a bubble but a narrow, expensive, real capability that has not generalised.

These horizons track results that would change what can honestly be said about matching medicine to individuals, not predictions of clinical availability.

• 10 yr: Frontier Ancestry portability is the link most likely to move, because the substrate now exists: All of Us at more than 535,000 genomes and 86 per cent underrepresented participants can produce non-European discovery cohorts at case counts that have never existed. Frontier Expect polygenic score performance outside European ancestry to close substantially without individual-level discrimination improving much for anyone, which will be reported as progress and is a different thing. Speculative A second multi-cancer early detection trial, powered for mortality in an enriched population, is the most decision-relevant experiment available and can be started now; whether anyone runs a randomised trial of returning polygenic results is a test of the field's seriousness rather than of its capability.
• 25 yr: Speculative Either response-phenotype datasets on the order of 100,000 patients per drug class exist, in which case individual response prediction becomes testable for the first time, or they do not, in which case the field remains subgroup medicine indefinitely and should say so. Speculative N-of-1 manufacturing either reaches a platform approval pathway and a hundred-fold cost reduction, making bespoke therapy a category rather than a case series, or the two documented instances remain two documented instances. Frontier In vivo delivery beyond liver is the technical fork: with it, autologous manufacturing stops being the bottleneck; without it, the price and throughput numbers in this brief are the steady state.
• 50 yr: Speculative If within-person longitudinal multi-omic monitoring becomes routine, the object of prediction changes from a population to a trajectory, which is the only version of this field that would deserve the name individualised. Speculative The alternative endpoint is equally coherent: common complex disease turns out to have a prediction ceiling set by environmental and stochastic variance, and precision medicine consolidates permanently as pharmacogenomic safety plus Mendelian therapeutics plus oncology genotyping. Handwave No method currently exists for telling which of those two futures is being observed from inside it.
• 100 / 250+ yr: Handwave At this horizon the interesting question is not whether medicine becomes individualised but whether the individual remains the right unit: continuous internal sensing, editing on demand and self-updating models would make the distinction between diagnosis and treatment stop being meaningful. Handwave Every component of that is asserted rather than demonstrated, and it is worth ending on the observation that this field's hardest constraint — knowing who responded to what — is a bookkeeping problem that has been solvable at every point in the last thirty years and has not been solved.

Frontier One clinical-AI clock runs on a different mechanism from the biology above. Change-control plans and the European high-risk obligations for medical-device software land inside the ten-year window and will certainly produce documentation; whether they produce a performance series is a separate question with a different answer, because none of them requires a deployed model to report measured outcomes on a schedule. Handwave This brief does not forecast which regulator writes that rule first.

• Depends on This brief waits on results that neighbouring subjects are producing for their own reasons, and it names none of them as a typed dependency because none is a result this subject is currently blocked on. It consumes editing and delivery chemistry from genome engineering, statistical machinery and within-family correction from behaviour genetics, and targeted-delivery physics from nanomedicine. What it is actually blocked on is a dataset that no research programme produces as an output: a health-record-linked, structured account of which patients responded to which treatments, at the scale of a hundred thousand patients per drug class. That is a record-keeping capability rather than a scientific result, which is why it appears nowhere on a technology map and why it is nonetheless the binding link.
• Requires (not on this map) Three non-scientific constraints bind this subject harder than any unsolved experiment. Health systems have no routine mechanism to pay once for a cure costing between $2.125 and $4.25 million, which is a large part of why 39 people had received the first approved CRISPR therapy some 22 months after its approval, against an eligible US population of about 16,000. Reference cohorts remain European-weighted, so the same predictor performs 25 to 65 per cent worse outside that ancestry group — a procurement fact rather than a biological one, and one national programme has now shown it is fixable by building an 86-per-cent-underrepresented cohort of 645,000 people. And autologous cell therapy still runs through apheresis, busulfan conditioning and a few dozen transplant-capable centres, so throughput is set by hospital capacity rather than by editing efficiency. A fourth constraint has arrived with clinical AI, and it is the same kind of thing. No jurisdiction requires a deployed clinical model to report measured performance against patient outcomes on a schedule, so a model that degrades produces no reportable event and no regulatory signal: the authorisation list grows and the performance record does not exist. That is a reporting rule a regulator could write rather than a research result, which places it with the other three.
• Enables What this subject supplies to others is a validated selection layer. Any therapeutic programme that treats a subgroup rather than a diagnosis depends on identifying that subgroup in advance, and the companion-diagnostic architecture assessed here is the mechanism by which that is done under regulation. It also supplies a benchmark that is easy to state and hard to beat: a proposal claiming to individualise an intervention now has to clear an area under the curve of 0.64 from genotype alone, a 52 per cent positive predictive value in an enriched screening population, and a 1.5 per cent end-to-end benefit rate in matched oncology. No typed enabling edge is recorded, because that contribution is a standard rather than a technology.
• Adjacent The adjacent subjects share failure modes rather than components. Longevity therapeutics faces the same missing response-phenotype dataset and would be unblocked by the same collection effort. Neurogenetics shares the polygenic scoring machinery and, more importantly, the within-family shrinkage correction that clinical scores have generally not had applied to them. Nanomedicine supplies the delivery route that would move in vivo editing off the liver. Genetic engineering owns the manufacturing and conditioning constraints that decide whether an approved precision therapy reaches anyone at all.

Established “Sequencing a tumour finds a treatment.” The actionable-finding rate and the actually-treated rate differ by a factor of five to twenty: 38 per cent with a candidate treatment on NCI-MATCH's own framing, against 6.4 per cent matched at MD Anderson and 2 per cent matched within NCI-MATCH itself as of May 2016, roughly 30 per cent of those responding, and about 1.5 per cent end-to-end benefit; a decade of searching found 32 exceptional-responder case reports against more than 18,000 commercially sequenced patients. Frontier The correct statement is that sequencing a tumour finds a mutation, that finding a mutation usually does not find a drug, and that finding a drug usually does not find a response.

Established “A blood test can screen the population for cancer.” The largest randomised trial of that proposition, at 142,000 participants, missed its primary endpoint of reducing combined stage III–IV diagnoses over three years, while the sponsor's release led on a substantial reduction in stage IV diagnoses, increased stage I and II detection and a four-fold higher detection rate, stating the miss in the body. Established An independent expert's assessment was that the findings were presented far more positively than the overall results justify, and that there is no evidence base to justify population-scale implementation. Frontier Read the release and the result side by side rather than choosing one: this is the most instructive interested-party contrast in the subject.

Established “A polygenic risk score tells you your personal risk.” Areas under the curve from genotype alone run 0.63 to 0.71, the schizophrenia score explains 7.3 per cent of liability, and what a score supports is a relative risk against a reference group rather than an individual probability. Established The odds ratio of 39 that impresses in headlines compares the top centile with the bottom centile, while the clinically usable contrast, top centile against everyone else, is 5.6. Frontier That distinction between an outlier statement and a population statement is the single most abused feature of this literature.

Established “HER2 testing means trastuzumab works” and “pharmacogenomic testing is proven cost-effective.” About 70 per cent of HER2-positive patients do not respond, and the same drug moved metastatic median survival from 20.3 to 25.1 months and has been standard of care for 27 years; both facts describe one drug, and the field's rhetoric has kept only the second. Established Pharmacogenomic testing is likewise proven effective — PREPARE, odds ratio 0.70 — while the literature states explicitly that evidence sufficient to validate its cost-effectiveness does not exist. Frontier Conflating effective with worth paying for is how a positive trial fails to change practice for years while everyone assumes someone else has done the economics.

Established “AI will deliver precision medicine.” The most heavily funded attempt cost MD Anderson $62 million before being stopped without meeting its goals, was discontinued by its flagship international partner, and was sold for roughly $1 billion after being called a cautionary tale of hubris and hype. Speculative The instructive part is that the failure was in curation, integration and workflow rather than in the oncology or the machine learning, which supports the hypothesis that clinical informatics is the binding constraint on the whole field — supported independently by the literature's own report of some 30,000 data-processing errors per processed human genome.

Established “Gene therapy is a one-time cure.” Three approved products say otherwise, each failing differently. Skysona: 6 of 67 patients developed myelodysplastic syndrome and a seventh acute myeloid leukaemia, roughly 10 per cent, five clones carrying a vector insertion in MECOM and one in PRDM16, with onset 14 to 92 months after treatment. Established Elevidys: approved in June 2023 on gene expression rather than clinical benefit, missed its primary motor endpoint in the phase III EMBARK trial at n = 125, and drew an FDA clinical hold after three deaths from acute liver failure in July 2025. Established Roctavian: factor VIII expression declines after roughly 12 to 18 months. Frontier “One-time” describes the administration, not the outcome, and the reimbursement architecture being built around one-time payment assumes the second meaning.

Established “Genomic medicine benefits everyone equally” and “approval means patients get it.” Prediction accuracy falls several-fold outside European ancestry and applied gains fall 25 to 65 per cent; the non-genomic version of that error is older, in Framingham-derived risk equations that produce biased outcomes applied to non-white populations. Established And the flagship CRISPR therapy recorded 39 infusions in its first 22 months against about 16,000 eligible US patients. Frontier The gap between approval and delivery is currently wider than the gap between research and approval, and it is measured in hospital capacity and payment rather than in science.

Speculative The deepest objection, and the one worth stating carefully: precision medicine may mistake variance for tractability. The founding intuition is that because patients differ, medicine tailored to those differences must be better — and that does not follow, because variation between patients can be large, real, and mostly unpredictable from any measurement available before treatment. Established The evidence pattern fits: every unambiguous success in this brief is a binary genotype defining a subgroup — HER2-positive, EGFR-mutant, CYP2D6 ultra-rapid, otoferlin, BCL11A — and not one is an individual prediction. Frontier On the fetched record, precision medicine is subgroup medicine with better marketing, and the honest test of whether it is more is whether any model has beaten the subgroup mean for a named person; nothing in this brief's sources demonstrates that, and doing so would need within-person longitudinal data rather than cross-sectional genotype.

Handwave The far end: the digital twin. The proposal is a continuously updated computational model of an individual patient — physiology, genome, exposures, prior responses — against which candidate treatments are simulated before being given. It is coherent, and it is funded in engineering domains where the governing equations are known. Speculative In medicine it presupposes exactly what this subject does not have: a validated model of individual response. A digital twin built on population parameters is a subgroup model with a person's name on it, and the input that would make it more — longitudinal within-person response data — is the same missing dataset that binds every other link in the chain. Frontier The proposal is not wrong so much as downstream of the bottleneck it is usually offered as a solution to.

Established Finally, four numbers this brief deliberately does not print, because they could not be verified. The US list price of the first approved CRISPR therapy; the rare-disease diagnostic yield of the 100,000 Genomes Project, routinely given as a round percentage; the number of gene–drug pairs covered by clinical pharmacogenetics guidelines; and the total count of FDA-approved companion diagnostics. Frontier Each is a figure a reader would expect a brief on this subject to state, and stating an unverified number in a piece about the misuse of numbers would be the one unrecoverable error.