Amyotrophic Lateral Sclerosis

Section 01Clinical definition

ALS is the simultaneous degeneration of two separate motor pathways, and that dual involvement is what distinguishes it from every mimic. The upper motor neurons are the Betz cells and other layer-V pyramidal neurons of primary motor cortex, projecting through the corticospinal and corticobulbar tracts. The lower motor neurons are the anterior horn cells of the spinal cord and the motor nuclei of cranial nerves V, VII, IX, X and XII.

Losing the upper set produces spasticity, hyperreflexia, clonus, Babinski signs and pseudobulbar affect. Losing the lower set produces weakness, wasting, fasciculations and cramps. Finding both in the same muscle is the clinical signature.

What the disease spares

The sparing is as characteristic as the damage. Oculomotor, trochlear and abducens neurons persist through to end stage. In one post-mortem series the oculomotor nucleus was altered in only 4 of 50 patients. Onuf's nucleus, in sacral segments S2 to S4, is likewise preserved, which is why bladder and bowel continence is generally retained. Sensation, cognition in roughly half of patients, and autonomic function are relatively spared.

Eye-tracking communication devices work precisely because of this anatomy.

Diagnostic criteria

Diagnosis is clinical and remains one of exclusion. Four body regions are assessed: bulbar, cervical, thoracic and lumbosacral.

The move from graded certainty to a single category reflects a hard-won lesson. Categories built for research eligibility were being used as clinical thresholds, and patients were waiting for a label they did not need in order to begin care.

The ALS to FTD spectrum

ALS and frontotemporal dementia are two ends of one continuum. They share TDP-43 pathology and, in C9orf72, a single causal mutation. The distribution among patients who are not overtly demented, from a series of 808 using Strong's revised criteria: Replicated

• Normal cognition and behaviour: 46.7%
• Cognitive impairment only: 22.9%
• Behavioural impairment only: 20.0%
• Both: 10.4%

A separate figure of roughly 15% meet full criteria for frontotemporal dementia. Executive dysfunction is the most common deficit and apathy affects around 28%. Cognitive status carries real prognostic weight: frontotemporal dementia raises the hazard of death or ventilation by 34%.

Section 02The disease spectrum

ALS sits inside a family of disorders that share its machinery, and several of them are not separate diseases so much as different amounts of the same one. Knowing which neighbour a patient actually has changes the prognosis by a factor of five or more, which makes this one of the highest-stakes distinctions in the field.

Why PLS progresses more slowly

The mechanical answer is that PLS spares the lower motor neuron, and the lower motor neuron is what kills. Upper motor neuron loss produces spasticity and weakness, which are disabling. Lower motor neuron loss produces denervation, muscle wasting, and, when it reaches the diaphragm and bulbar muscles, respiratory failure and aspiration. Since 61% of ALS deaths are respiratory muscle failure, a disease that does not denervate the diaphragm has removed its own principal cause of death.

That is also why PLS patients frequently plateau and why survival stretches into a second decade, while classic ALS runs 2.6 years.

Whether PLS converts to ALS

Conversion is not assured, and the risk is strongly time-dependent. The 2020 consensus criteria encode this directly: probable PLS requires the absence of significant lower motor neuron degeneration for two to four years from onset, and definite PLS requires four years or more. The four-year threshold exists because conversion risk falls sharply past it. A patient who has gone four years with pure upper motor neuron disease is very unlikely to become an ALS patient.

Two qualifications matter clinically. Minor EMG abnormalities, including slightly increased insertional activity and scattered fibrillations, are permitted within the criteria and do not by themselves signal conversion. And the direction of travel is one-way: there is no mechanism by which established lower motor neuron loss reverts, so reclassification from ALS back to PLS does not happen.

What the spectrum implies

These disorders are not merely similar. They share specific genes (C9orf72 across ALS and FTD, VCP across ALS and multisystem proteinopathy) and specific lesions (TDP-43 mislocalisation across ALS, FTD, and both chronic traumatic encephalopathy and limbic-predominant age-related TDP-43 encephalopathy).

The practical consequence runs in one direction that matters. Spinal muscular atrophy, which is the one member of this family caused by simple loss of a single gene product rather than by proteinopathy, is also the one member with effective approved therapies. That is not a coincidence, and it reinforces the pattern in Section 15: the therapies that work are the ones aimed at a single defined upstream cause.

Section 03Epidemiology

ALS is rare at any moment and not rare over a lifetime. Global pooled incidence is 1.65 per 100,000 person-years (95% CI 1.43 to 1.91) and prevalence 5.05 per 100,000, from a 142-study meta-analysis published in JNNP in 2026. Because survival is short, prevalence badly understates how many people the disease touches. Lifetime risk is roughly 1 in 350 for men of European ancestry and 1 in 400 for women. Replicated

Geographic and ancestry variation

The gradient is large and has been stable across four decades of study.

Within the United States, incidence in white Americans runs roughly 5.4 per 100,000 against 2.3 in African Americans. The gradient tracks ancestry more than geography, and gene frequencies explain part of it. The C9orf72 expansion derives from a single Northern European founder haplotype and is roughly tenfold rarer in Asian populations.

Age and sex

Incidence peaks at ages 70 to 79. Typical onset is 55 to 75. About 10% of cases begin before 45 and roughly 1% before 25. Familial cases begin about five years earlier than sporadic ones, and specific genotypes shift this sharply: mean onset is 48.5 years in SOD1 carriers and as young as the teens or twenties in FUS p.P525L juvenile ALS.

Men are affected roughly 1.5 times as often as women, with an incidence rate ratio of 0.74 female to male. The gap narrows with age and approximately closes after 70. Bulbar onset is proportionally more common in older women.

Survival

Survival varies sharply by where the disease starts. Respiratory onset gives a median of 1.4 years with 0% alive at 10 years. Bulbar onset gives 2.0 years, 3% at 10 years. Limb onset gives 2.6 years, 13% at 10 years. SOD1 p.A5V, the most common SOD1 variant in North America, gives a mean disease duration of 12.9 months.

Where the numbers are going

An estimated 32,893 people in the US had ALS in 2022, projected to 36,308 by 2030 in National ALS Registry modelling. Globally, Arthur et al. projected a rise from 222,801 cases in 2015 to 376,674 by 2040, a 69% increase driven almost entirely by population ageing, with the steepest growth in Africa at +116% and Asia at +81%.

Section 04Genetics

The familial and sporadic distinction is an artefact of family size and record-keeping, not biology. About 10% of patients report a family history, but known genes explain only 70% of those and 15% of the rest, while twin studies put the heritability of sporadic ALS at 0.61 (95% CI 0.38 to 0.78) and an Irish population register put it at 52.3%.

A Brazilian cohort of 1,911 patients showed the gradient directly. Diagnostic yield was 89.7% in definite familial cases, 24.0% in probable, 15.4% in possible, and 4.9% in apparently sporadic disease. The line is continuous, not binary.

Three things blur it further. Penetrance is age-dependent, so a true carrier family can look sporadic. De novo mutations, with FUS variants accounting for a reported 43% of patients whose onset is before 35, produce genetic disease with no family history at all. And about 6% of patients carry rare variants in two or more ALS genes, regardless of which label they were given. Replicated

The major genes

Frequencies differ enough by ancestry that a single global table is misleading. These are the case-series figures from Zou et al., pooling 111 studies.

Note the inversion. C9orf72 dominates in Europeans and is nearly absent in East Asians, where SOD1 and FUS carry the load.

C9orf72

A GGGGCC hexanucleotide repeat expansion in intron 1. Healthy alleles carry up to 8 repeats. Pathogenic alleles usually carry hundreds to thousands, with a conventional threshold above 30 and a growing case for lowering it to 24. Carrier frequency in the general population is about 1 in 1,600. Median onset is 58.0 years. Cumulative penetrance estimates range from 80 to 90% by age 80 up to 99.5% by age 83, and the higher figure comes from a symptomatic-ascertained cohort, so it is probably inflated. The expansion is unstable, and repeat length is substantially larger in CNS tissue than in blood, which confounds blood-based sizing.

SOD1

201 variants catalogued across 30 years, roughly 12% of familial and 2% of sporadic ALS. It has a distinctive clinical signature: 95% spinal onset, only 4.4% bulbar, and mean onset at 48.5 years. Variant identity matters enormously. p.A5V gives a mean survival of 12.9 months. Homozygous p.D91A is recessive and slow, and in a Scandinavian series 139 heterozygous D91A carriers on the recessive haplotype developed no ALS at all.

TARDBP and FUS

The two genes coding for the aggregating proteins themselves. FUS matters out of proportion to its frequency because it causes juvenile ALS. The p.P525L variant, which disrupts the nuclear localisation signal, is usually de novo, presents in the teens or twenties, and progresses very fast. Severity tracks how badly the variant disrupts nuclear import.

Second tier

NEK1 carries risk variants in nearly 3% of cases. ATXN2 intermediate CAG repeats account for 4.7 to 6.5%. TBK1, KIF5A, ANXA11, OPTN, VCP, UBQLN2, CHCHD10, SQSTM1, CFAP410 and DNAJC7 each contribute at the 1 to 2% level or below. Van Rheenen et al., pooling 29,612 cases, identified 15 genome-wide significant loci converging on vesicle transport and autophagy.

The most recent yield estimate is the largest. A March 2026 harmonised exome analysis across 17,919 cases reported risk variants in over 20% of all patients, validated ARPP21, DNAJC7 and CFAP410, and added five new single-variant associations. Single study

Models of how it starts

The dominant framework has been multistep pathogenesis, adapted from the Armitage-Doll model of carcinogenesis. Al-Chalabi et al. found a linear relationship between log incidence and log age across five European registers, with a slope implying six steps. Chiò et al. then showed that carrying a high-penetrance mutation pre-pays some of them.

This is the cleanest available explanation of why mutation carriers are not born with ALS and why penetrance rises with age. It has been replicated in South Korean, Japanese and Australian cohorts, and an Australian study found the step count varies by phenotype, from 5.1 for bulbar onset down to 2.7 for cervical.

A further puzzle sits underneath all of this. Common variants explain very little. SNP-based heritability is 0.028, under 3% of liability, against 50 to 60% from family designs, and polygenic risk scores show no effect on survival or age at onset. That gap is unexplained.

Genetic testing

The 2023 consensus guidelines recommend offering testing to every person with ALS regardless of family history, at minimum a C9orf72 repeat assay plus sequencing of SOD1, FUS and TARDBP. The oligogenic data add a second recommendation that is easy to miss: keep testing even after one pathogenic variant is found.

Practice lags badly. In the 1,911-patient Brazilian cohort, 17.3% of patients underwent any genetic testing, and those tested were younger and better educated. That is a documented equity gap. The clinical stakes changed in 2023, when tofersen made SOD1 identification directly actionable.

Section 05Environment and exposure

No environmental exposure has been established as a cause of ALS, and one methodological problem contaminates nearly every study in this literature. The disease has a long prodrome, so exposures measured shortly before diagnosis are frequently consequences rather than causes. What follows separates what survives that test from what does not.

Military service

Veterans have roughly 1.5 to 2 times the ALS risk of non-veterans. The two strongest studies are Horner et al. 2003, comparing Gulf War deployed against non-deployed personnel at RR 1.92 (95% CI 1.29 to 2.84), and Weisskopf et al. 2005, a prospective cohort of 408,288 men at RR 1.53 (95% CI 1.12 to 2.09). The Institute of Medicine graded this in 2006 as "limited and suggestive evidence of an association", below their threshold for sufficient evidence.

The VA nonetheless established presumptive service connection in 2008 under 38 CFR § 3.318, covering any veteran with 90 days of continuous active service, at a minimum 100% rating. This is a documented case of a benefits presumption deliberately outrunning causal proof. It is reasonable policy. It is not evidence.

On Camp Lejeune, a point that is widely misreported. The 2024 mortality cohort study of 159,128 Camp Lejeune personnel did not evaluate ALS as an outcome at all. The words "amyotrophic" and "motor neuron" do not appear in it. ALS is not among the eight VA Camp Lejeune presumptive conditions. The Parkinson's disease link there is reasonably supported. The ALS link is untested, not disproven.

Athletes and physical activity

This is the sharpest paradox in ALS epidemiology, and the resolution is that two different questions are being conflated.

Elite professional athletes genuinely have elevated risk, with internal dose-response. Chiò's Italian football cohort covered 7,325 professional players and found 5 observed cases against 0.77 expected, an SMR of 6.5 (95% CI 2.1 to 15.1), rising to 15.2 for careers over five years. Mean age at onset was 43.4 years, two decades early, and 3 of the 5 were bulbar onset at p = 0.003. Daneshvar et al. followed 19,423 NFL players and found 38 diagnoses, an SIR of 3.59 (2.58 to 4.93), with OR 1.2 per additional season played and mean diagnosis at 51.0 years. Established

Ordinary physical activity is a different matter. A 34-study meta-analysis gives vigorous activity OR 1.26 (1.06 to 1.49) and leisure-time activity OR 1.08 (1.04 to 1.12), small effects either way. A prospective Norwegian cohort of 373,696 people followed for 27 years found the opposite: high midlife activity associated with roughly 41% reduced ALS risk in men, with no effect in women.

Mendelian randomisation splits the difference in an interesting way. Julian et al. found causal support for strenuous leisure exercise at p = 0.01, found that exercise-responsive genes are enriched for ALS risk genes including C9orf72 at p = 0.013, and found that among C9orf72 carriers, higher lifetime physical activity predicted earlier onset at p = 0.007. Their reading is that exercise injures motor neurons mainly in people carrying a risk genotype.

Premorbid leanness does not survive the same test. Observational data show each unit of BMI associated with about 3% lower ALS risk, but Mendelian randomisation is flatly null at OR 1.04 (95% CI 0.97 to 1.11), and the same analysis correctly detected causal BMI effects for multiple sclerosis and stroke. The leanness is almost certainly prodromal hypermetabolism, not a cause. Not supported

Head trauma

Pooled meta-analyses give OR 1.45 to 1.47. Two features demolish the causal reading.

First, the association is concentrated immediately before diagnosis. A 2025 UK study of 85,690 TBI patients found HR 6.18 (3.47 to 11.00) within two years of injury and no significant association beyond two years, and a time-lag meta-analysis found odds ratios falling to 1.16 at three years or more and 1.05 at ten years or more. Second, there is no dose-response. Multiple head injuries carry no more risk than one.

The parsimonious explanation is reverse causation. People with undiagnosed early ALS fall and hit their heads. Neuropathological co-occurrence of CTE and ALS is real but uncommon, at 9 of 155 cases (5.8%) in a referral-biased brain bank, and does not generalise to ordinary head injury.

Guam and the BMAA hypothesis

The Western Pacific ALS-parkinsonism-dementia complex was the largest neurodegenerative cluster ever recorded. Among Chamorros on Guam in the 1950s, prevalence ran 50 to 100 times the world average, and lytico-bodig was the leading cause of death at roughly 420 per 100,000. No familial genetic pattern was found, and migration studies showed risk shifted with environment in both directions, which is strong evidence of an environmental cause. Incidence peaked between 1950 and 1954 and has since collapsed to near zero.

The leading hypothesis runs: cyanobacteria in cycad roots produce BMAA, cycad seeds are processed into flour, and the toxin biomagnifies in flying foxes, a local delicacy. Cox's measured gradient is striking, at roughly 0.3 µg/g in cyanobacteria, 37 µg/g in cycad seeds, and 3,556 µg/g in flying foxes. Vervet monkeys fed BMAA developed both amyloid plaques and neurofibrillary tangles, reduced about 70% by co-administered L-serine.

Other exposures

The exposome framing

The field's current organising concept is the gene-time-environment hypothesis: ALS emerges from interaction between genetic liability and accumulated environmental exposure during ageing. The practical advance is pooled environmental risk scores computed from a single blood draw, which index cumulative exposure to dozens of toxicants at once and predict both risk and survival.

Gene-environment interaction is well motivated but thinly evidenced. The C9orf72 exercise finding supports it. Euro-MOTOR found no C9orf72 modification of the physical-activity association, and the Michigan metal findings were independent of polygenic risk. One supportive result, several null tests.

Section 06The TDP-43 lesion

About 97% of all ALS, familial and sporadic, shows the same pathology: TDP-43 cleared from the nucleus and accumulated in the cytoplasm as ubiquitinated, hyperphosphorylated inclusions. The exceptions are SOD1-mutant and FUS-mutant ALS, which carry their own proteinopathies and are TDP-43-negative. This convergence is the single most important organising fact in the field. It means a therapy aimed at TDP-43 dysfunction would, in principle, address nearly the whole disease. Established

What TDP-43 normally does

TDP-43 is a 414-amino-acid nuclear RNA-binding protein with two RNA recognition motifs and a C-terminal low-complexity domain that hosts most ALS-causing TARDBP mutations. Two of its jobs matter here.

The first is splicing repression. TDP-43 binds long UG-rich intronic stretches and suppresses the inclusion of cryptic exons, non-conserved sequences that would otherwise be spliced into mature mRNA. The second is autoregulation. TDP-43 binds its own 3'UTR and throttles its own production, keeping nuclear levels in a narrow window. Age-related DNA demethylation of that autoregulatory region degrades the feedback loop, which has been proposed as one link to sporadic onset.

A third function was established only in late 2025. Two independent Nature Neuroscience papers showed TDP-43 also controls alternative polyadenylation, with Zeng et al. finding 3,206 genes affected and Bryce-Smith et al. finding 227 cryptic APA events. Notably, some of these increase protein output rather than depleting it. TDP-43 loss is not purely subtractive. Single study

The cascade

Gasset-Rosa et al. (Neuron, 2019) showed cytoplasmic TDP-43 assemblies sequester the nuclear import machinery, making nuclear clearance a consequence of cytoplasmic aggregation. A 2025 seeding study using recombinant TDP-43 fibrils confirmed this order by live imaging: cytoplasmic aggregation first, nuclear depletion after, with cryptic splicing appearing within two days.

The counter-position is that the visible inclusion is not the toxic species at all. Inclusion burden does not track degeneration. TDP-43 C-terminal fragments accumulate widely in both ALS and FTD, yet only FTD shows marked atrophy. On this reading the soluble nuclear deficit is what kills, and the aggregate is a tombstone. Contested

One older model has been substantially revised. Stress granules are no longer considered the obligatory crucible of TDP-43 aggregation, since cytoplasmic TDP-43 droplets form independently of them. Treat stress granule involvement as modulatory.

Cryptic exons, the conceptual breakthrough

Ling et al. (Science, 2015) established the principle that TDP-43's dominant nuclear job is repressing non-conserved cryptic exons. The words non-conserved carry the weight. Because these sequences differ between species, mouse models had been systematically blind to the human targets for two decades. This reframed the disease from "aggregate equals toxic" to "nuclear loss equals a definable, measurable, druggable splicing lesion."

Two genes converted the principle into specifics.

STMN2, 2019

Klim et al. found that TDP-43 loss triggers inclusion of a cryptic exon in STMN2, producing a truncated, non-functional transcript. Baughn et al. (Science, 2023) resolved the mechanism: TDP-43 binding sterically blocks a cryptic splice site, and its loss unmasks premature polyadenylation. López-Erauskin et al. then showed that depleting stathmin-2 in adults by three independent routes produces axoplasmic collapse, loss of large-caliber axons, slowed conduction and NMJ denervation, with the motor neuron cell bodies still alive.

UNC13A, 2022

Two back-to-back Nature papers showed TDP-43 loss permits a cryptic exon between exons 20 and 21 of UNC13A, introducing premature stop codons and depleting the protein that primes synaptic vesicles for fusion. What made this a landmark is the genetics. UNC13A had been a top GWAS hit since 2009 with no known mechanism, and the risk variants turn out to be conditional. One sits 16 bp inside the cryptic exon, another 534 bp downstream, and neither does anything until TDP-43 is lost, at which point they amplify cryptic splicing. This is the clearest example anywhere of a common risk variant acting through the core proteinopathy. Established

More than 3,000 cryptic events have now been detected across post-mortem datasets. Only about 210 are reliably predictive of TDP-43 proteinopathy. Confirmed targets include HDGFL2, AGRN, ACTL6B, POLDIP3, RANBP1, CAMK2B, MYO18A and ATG4B, and the repertoire varies by cell type.

Why this matters practically

Cryptic exons gave the field two things it had never had.

First, a biomarker of TDP-43 dysfunction itself. Most cryptic transcripts are degraded by nonsense-mediated decay and are invisible in biofluids. HDGFL2 is the exception, because its cryptic exon inserts in-frame, escapes decay, and yields a stable peptide. Irwin et al. (Nature Medicine, 2024) detected it in CSF of symptomatic C9orf72 carriers, sporadic ALS patients, and presymptomatic carriers. Its trajectory runs opposite to neurofilament: cryptic HDGFL2 peaks before symptoms and declines, while NfL rises after. The ratio of the two has been proposed as a staging metric.

Second, a druggable target that bypasses TDP-43 entirely. Antisense oligonucleotides can mask a cryptic splice site without touching TDP-43. QurAlis's QRL-201, an STMN2-targeting ASO, reported in February 2026 that it confirmed target engagement, corrected STMN2 mis-splicing, and showed a post hoc slowing of ALSFRS-R decline at six months in 69 patients. That is a company-reported post hoc finding in a Phase 1/2 trial, not a registrational result. It is also the first clinical evidence that the cryptic exon route is actionable in sporadic disease. Single study

Section 07Converging pathways

ALS pathogenesis is a network, not a chain, which is precisely why single-target drugs keep failing. The nodes below are mutually reinforcing, and several genes sit in more than one.

The two non-TDP-43 genotypes

SOD1 is a toxic gain of function, not a loss of enzyme activity. Sod1-null mice do not develop motor neuron disease, many pathogenic variants retain full catalytic activity, and severity does not track enzyme function. What the toxic species actually is remains unresolved. Candidates are metal-deficient monomers, soluble oligomers, and large inclusions, and the evidence weighs against the large inclusions, since eliminating them fails to improve mice.

Whether misfolded wild-type SOD1 contributes to sporadic ALS is genuinely disputed. Conformation-specific antibodies detect it in some sporadic samples, but Da Cruz et al. tested seven such antibodies and found no signal in sporadic post-mortem tissue, and one widely used antibody stains all non-neurological controls. Antigen-retrieval protocols can themselves induce false-positive misfolding. The defensible statement is that it may occur in a subset, and its causal role is unproven. Contested

C9orf72 works through three non-exclusive mechanisms.

1. Haploinsufficiency. The expansion causes promoter hypermethylation and abortive transcription, roughly halving C9orf72 protein. The protein regulates autophagy initiation, autophagosome-lysosome fusion and lysosomal homeostasis. But C9orf72-null mice develop autoimmunity, not motor neuron disease, so loss of function alone is insufficient.
2. RNA foci. Sense and antisense repeat transcripts form G-quadruplexes and nuclear foci that sequester nucleolin, hnRNPs, Pur-α and TDP-43 itself, driving nucleolar stress. Foci abundance correlates poorly with degeneration.
3. Dipeptide repeat proteins. Repeat-associated non-AUG translation produces five species from both strands. Poly-GA is most abundant and sequesters proteasome components and phosphorylated ATM. Poly-GR and poly-PR are by far the most toxic, binding ribosomes and nucleoli, disrupting membraneless organelle dynamics, with poly-GR binding mitochondrial ribosomal proteins specifically and linking the repeat directly to bioenergetic failure. Poly-GP is the least toxic and the most useful CSF pharmacodynamic biomarker.

The weak consensus is synergy. Gain of function initiates, haploinsufficiency sensitises. Reduced C9orf72 expression worsens poly-GR toxicity in patient-derived motor neurons.

Nucleocytoplasmic transport

The motor neuron loses the ability to move things across the nuclear envelope, and the ordering is contested in an important way.

Downstream evidence: aggregated TDP-43 sequesters 37 nucleocytoplasmic transport proteins, with FG-repeat nucleoporins co-aggregating via their own prion-like domains. RanGAP1 mislocalises, the Ran-GTPase gradient collapses, and nuclear import and mRNA export both fail.

Upstream evidence: Coyne et al. (Science Translational Medicine, 2021) found that nuclear accumulation of the ESCRT-III protein CHMP7 initiates nuclear pore injury and that TDP-43 dysfunction follows. That places pore damage before TDP-43 mislocalisation, a direct challenge to the ordering used everywhere else in this report. Contested

Proteostasis, where a striking number of ALS genes land

Motor neurons carry an additional handicap here. They have an unusually high threshold for inducing the heat-shock response, so chaperone upregulation is sluggish in exactly the cell type that needs it most.

Mitochondria and calcium

The ER-mitochondria tether formed by VAPB and PTPIP51 governs calcium transfer that drives the TCA cycle and ATP output. Blair et al. (2025) showed both TDP-43 and FUS disrupt this tether, FUS by activating GSK-3β, and that tether loss precedes symptom onset in transgenic models. Restoring it rescues calcium and synaptic defects caused by mutant TDP-43. The ALS8 VAPB P56S mutation sits directly in the binding domain.

In CHCHD10 disease, mutant protein aggregates in the mitochondrial intermembrane space, producing two separable lesions: defective cytochrome c oxidation with reduced mitochondrial copper, and activation of the integrated stress response through the OMA1-DELE1-HRI axis. Suppressing the stress signalling delayed cardiomyopathy but did not rescue respiration. The two arms come apart.

Axonal transport

KIF5A mutations cluster in the cargo-binding tail, usually causing exon 27 skipping. One proposed mechanism is loss of autoinhibition producing a constitutively hyperactive motor. The other, from a 2025 study, is remarkable in its specificity: the mutant splice product appends 39 amino acids that shift the domain's isoelectric point from 4.12 to 12, driving aggregation. Engineering the pI back down to 4.35 abolished both the aggregation and the cytoplasmic TDP-43 mislocalisation it caused. That places KIF5A upstream of TDP-43 pathology.

Neurofilaments connect this node to the biomarker story. TDP-43 loss reduces NEFL protein through polyadenylation site selection, and stathmin-2 loss collapses inter-neurofilament spacing. The neurofilament light chain measured in patient plasma is the debris.

Excitotoxicity

The astrocytic glutamate transporter EAAT2 is markedly reduced in ALS motor cortex and spinal cord, leaving glutamate in the synapse. Riluzole, the only broadly approved ALS drug, works partly by raising EAAT2 activity.

The more precise link runs through RNA editing. GluA2 mRNA is normally 100% edited at the Q/R site by ADAR2, which renders AMPA receptors calcium-impermeable. Aizawa et al. found that more than half of motor neurons in ALS cases were ADAR2-negative, against none in controls, and that every ADAR2-negative neuron showed phosphorylated cytoplasmic TDP-43 with nuclear clearing. Unedited GluA2 means calcium-permeable AMPA receptors and calcium overload, in a cell type poorly equipped to handle it.

Why motor neurons specifically

Four intrinsic features make them vulnerable.

• Low calcium buffering. Vulnerable spinal and bulbar motor neurons express little parvalbumin or calbindin-D28k, while resistant oculomotor and Onuf's nucleus neurons express both highly.
• Extreme geometry. Axons up to a metre long, with axonal volume vastly exceeding the cell body, imposing enormous transport and local translation load.
• High tonic firing and metabolic demand, hence deep mitochondrial dependence.
• High Cav1.3 expression in spinal relative to oculomotor neurons.

The best-characterised molecular determinant is MMP-9. Kaplan et al. (Neuron, 2014) found it expressed only in fast motor neurons, the selectively vulnerable ones, and absent from slow and oculomotor neurons. Ablating it delayed denervation by about 80 days and extended lifespan about 25% in SOD1 mice, and the protection generalises to TDP-43 models. Conversely, overexpressing oculomotor-enriched genes in spinal motor neurons extends survival in ALS mice, which is the strongest evidence that the vulnerability programme is causal rather than merely correlated. Replicated

Additional nodes, in brief

DNA damage. C9orf72 disrupts repair through at least four routes: poly-GA sequestering phospho-ATM, repeat-length-dependent R-loop accumulation, impaired non-homologous end joining, and oxidative lesions from poly-GR-driven mitochondrial dysfunction. TDP-43 is itself required for double-strand break repair.

ER stress. Motor neurons show early UPR activation, and in SOD1 mice stress markers rise in fast-fatigable neurons before denervation. The therapeutic direction is genuinely unsettled. Guanabenz, which prolongs the stress response, improved disease in one mouse study and accelerated it in another, while ISRIB, which acts oppositely, also improved survival.

Oxidative stress sits downstream of nearly every other node and is now framed as an amplifier rather than an initiator. The original loss-of-dismutase hypothesis is dead. Edaravone is the clinical embodiment of this pathway, and its efficacy remains contested.

Ferroptosis, iron-dependent cell death driven by lipid peroxidation, is a more specific version of the oxidative argument and has been proposed as the route through which edaravone acts. It is mechanistically attractive because it connects the oxidative node to the lipid abnormalities in Section 09, particularly the very-long-chain fatty acid elevation and sphingolipid dysregulation. The evidence is preclinical, and the clinical arm of the argument rests on a drug whose effect size is disputed. Contested

Section 08Proteostasis beyond ALS

ALS is one instance of a general failure mode, and putting it beside the others clarifies what is specific to motor neurons and what is not. Every cell runs a continuous protein folding operation: ribosomes produce polypeptides, chaperones fold them and refold the ones that slip, and the proteasome and autophagy clear what cannot be salvaged. That system has finite capacity. When production of a misfolding-prone protein exceeds it, aggregates accumulate, and the tissue that depended on that protein fails.

The same arithmetic produces very different diseases depending on which protein, which tissue, and how fast.

What the comparison shows about ALS specifically

Three features of the motor neuron make it a uniquely bad place for this failure to occur.

It cannot dilute the problem. Dividing cells halve their aggregate burden every division. A beta cell can be replaced. A hepatocyte regenerates. Motor neurons are post-mitotic and are never replaced, so aggregates accumulate monotonically across a lifetime with no reset.

The logistics are extreme. A motor neuron may maintain an axon a metre long, with axonal volume vastly exceeding the cell body. Chaperones, proteasomes and autophagosomes must be trafficked out along that axon and the debris trafficked back, which is why the axonal transport genes and the autophagy genes both appear on the ALS gene list. No other cell in this table has that problem.

The stress response is sluggish exactly where it is needed. Motor neurons require a higher threshold than other cell types to induce the heat-shock response. The cataract comparison is instructive here: in the lens, the chaperone itself is the thing that wears out. In motor neurons, the chaperone system works but is slow to switch on.

The therapeutic read-across, and why it has not worked yet

This frame makes an obvious prediction: boost chaperone capacity or clearance, and ALS should improve. That prediction has been tested and has failed twice, which is worth stating plainly rather than treating the frame as a promissory note.

• Arimoclomol, a heat-shock response co-inducer aimed directly at this mechanism, showed no clinical benefit in a 239-patient trial.
• Trehalose, which induces autophagy through TFEB activation and had a clean preclinical rationale in misfolded-protein toxicity, missed its primary endpoint as HEALEY regimen E.

Two readings are available and the field has not settled between them. Either the proteostasis frame is correct but these were the wrong drugs, aimed at a downstream node in a cascade already running. Or capacity was never the limiting variable, and the aggregate is a consequence of the upstream lesion rather than its cause, which is the same argument made in Section 06 about whether the inclusion or the nuclear deficit is what kills.

The observation that supports the second reading: the two therapies that have produced unambiguous mechanistic success, tofersen and ulefnersen, both work by reducing production of the misfolding protein rather than by improving the cell's ability to handle it. Turning down the input has worked. Raising the capacity has not.

Section 09Patient biochemistry

ALS is measurable in blood long before it is visible in the clinic, but almost nothing measured is specific to ALS. What follows separates what is reproducible from what is a single study, because that distinction has repeatedly misled the field.

Neurofilaments, the one biomarker that changed practice

Neurofilament light chain is structural debris from dying axons, and in ALS there is a great deal of it.

Pooled across 60 studies and 8,801 participants, CSF NfL distinguishes ALS from healthy controls with sensitivity 0.91 and specificity 0.90. Blood NfL carries a survival hazard ratio of 2.52 (1.84 to 3.47) and correlates with progression rate at r_s of about 0.46. Replicated

Three properties determine how it is used.

It rises before symptoms. A 2026 Nature Medicine proteomics study following 33 presymptomatic carriers to phenoconversion found 92 plasma proteins changing significantly before onset, first detectable at a median of 1.3 years beforehand. NEFL shows the steepest trajectory, rising sharply 6 to 12 months before symptoms. A 19-protein panel achieved AUC 0.945 at six months and 0.897 at five years before conversion. Single study

It plateaus. Longitudinal plasma NfL slopes are flat across fast, intermediate and slow progressors. It is a level marker, not a slope marker: useful as a baseline prognostic and as a pharmacodynamic readout, useless for tracking an individual patient's decline.

It is not specific. NfL indexes axonal injury, not ALS. Against disease mimics it performs substantially worse than against healthy controls, and mimic levels sit about 2.4-fold above normal.

That last caveat notwithstanding, NfL is the basis on which the FDA granted tofersen accelerated approval in April 2023, the first time a biomarker was accepted as reasonably likely to predict clinical benefit in ALS.

Whole-body metabolism: the disease burns fuel

Roughly 50 to 68% of ALS patients are hypermetabolic, consuming more energy at rest than their body size predicts. The wide range is largely definitional, since thresholds vary from 110% to 120% of predicted resting energy expenditure and there is no standardised protocol. Hypermetabolism correlates with faster decline and shorter survival.

This produces one of the disease's genuine paradoxes. Higher BMI and higher cholesterol are protective once disease has begun, with each 5 kg/m² of pre-diagnostic BMI associating with roughly 15% slower progression. Yet lower premorbid BMI associates with higher risk of developing ALS at all. Mendelian randomisation resolves it: genetically predicted BMI has no causal effect on ALS risk. The leanness is prodromal, not causal.

There may be a central driver. Hypothalamic atrophy has been documented in sporadic ALS, with TDP-43 pathology and loss of oxytocin- and orexin-expressing neurons.

Other biofluid findings

Lipids and metabolomics

Lipidomics has replicated better than metabolomics. Consistently reported: elevated very-long-chain fatty acids, reduced phosphatidylcholine and phosphatidylserine, sphingolipid dysregulation, oxysterol alterations with 25-OHC upregulated from presymptomatic stages, and cholesteryl ester accumulation in spinal cord reaching 22-fold for some species. Arachidonic acid pathway activation appears across SOD1, C9orf72 and TDP-43 motor neuron lines alike.

Metabolomics is thinner. Reduced NAA, increased CSF glucose and lactate, reduced TCA intermediates and purine dysregulation recur, but the field's own verdict from a 2025 multiomics review is that "definitive evidence that associates a specific metabolic fingerprint with ALS is still lacking."

Chronology: what appears when

The single most useful structural finding is that there is no evidence any one compartment moves first.

That the presymptomatic plasma panel is dominated by muscle proteins, including nebulin, myosin heavy chain, titin, actinin and troponin C, moving on the same timescale as neurofilament, is direct biochemical support for early peripheral involvement. It bears directly on the dying-forward versus dying-back debate in Section 11.

Section 10Glia and immunity

The motor neuron does not die alone, and in the cleanest experiments it does not even set the pace. Restricting mutant SOD1 expression to neurons alone does not produce ALS-like pathology, while astrocyte-specific deletion of mutant SOD1 substantially slowed progression without changing onset. The working summary: neurons determine when the disease starts, glia determine how fast it moves. Replicated

Astrocytes

Patient-derived astrocytes kill motor neurons in culture, and they do so whether they come from familial or sporadic patients. Haidet-Phillips et al. (Nature Biotechnology, 2011) found astrocytes from post-mortem sporadic ALS tissue similarly toxic to those from familial cases, and that knocking down SOD1 attenuated the toxicity even in sporadic cases. That is one of the few pieces of evidence linking wild-type SOD1 to sporadic disease.

The mechanism has at least two arms. The first is glutamate clearance failure, with EAAT2 downregulated in ALS astrocytes, leaving glutamate in the synapse. The second is release of something directly toxic, since conditioned media alone is sufficient, though what that something is has never been definitively identified.

A caution on terminology: the A1/A2 reactive astrocyte nomenclature is widely used but now contested as an oversimplification. Treat it as a framework, not a validated cell census.

Microglia

Microglia switch sides. In SOD1 models, early disease coincides with an anti-inflammatory, neuroprotective phenotype, and as disease advances there is a switch to a pro-inflammatory phenotype that accelerates degeneration. The M1/M2 framing that describes this is itself disputed, and current work uses disease-associated microglia and lipid-droplet-accumulating subtypes instead, though ALS-specific characterisation of these remains limited.

Microglial activation is measurable in living patients. TSPO PET imaging shows increased binding in motor cortices, corticospinal tracts and precentral gyri, co-localising with regions of cortical thinning, correlating with upper motor neuron burden and inversely with ALSFRS-R. Like NfL, it behaves as a level rather than a slope: patients rescanned at six months showed stable uptake despite clinical progression.

Oligodendrocytes, the energy supply line

This is an underappreciated node. Oligodendrocytes feed axons lactate through the transporter MCT1, and in ALS that supply line fails. Lee et al. (Nature, 2012) found ALS motor cortex contains less than half the MCT1 protein of controls, with normal oligodendrocyte numbers. The cells are present but not doing the job. Knocking down MCT1 in rodent spinal cord killed nearly half the local motor neurons, and mice with 50% of normal MCT1 develop axonal swellings resembling ALS.

The interpretation is that axons are starved of energy. It connects directly to the axonopathy picture from stathmin-2 loss.

The peripheral immune system

Regulatory T cells are the best-quantified peripheral finding, and the relationship is strikingly linear with disease speed.

Suppressive capacity correlated with disease burden at R = 0.529 and with progression rate at R = 0.534. The defect is epigenetic, not merely numerical: methylation of the FOXP3 Treg-specific demethylated region was significantly higher across all 15 CpG islands in rapid progressors.

That finding motivated Treg infusion trials, and the results so far are suggestive but methodologically weak. The first-in-human study of 3 patients, open-label, reported progression slowing during infusion rounds and accelerating between them. A Phase 1 combining abatacept with low-dose IL-2 in 4 patients reported an ALSFRS-R slope of +0.04 points per month over 24 weeks against a retrospectively estimated pre-treatment slope of −1.1. These are uncontrolled studies with single-digit enrolment. A properly powered readout has since appeared in NEJM Evidence (2025) and is the result to judge this approach on. Single study

More broadly, Th1, Th17, NK cells and monocytes all increase with progression, and the neutrophil-to-lymphocyte ratio tracks progression rate.

Section 11Onset and spread

ALS starts in one place and moves. It has no plan, but it has a direction, and understanding that direction is the closest thing the field has to a theory of the disease.

It begins focally

Onset is almost never diffuse. Roughly two-thirds of patients begin with limb symptoms, about 31.9% with bulbar symptoms in the 11,475-patient ENCALS consortium, and roughly 5% with respiratory failure. The first symptom is typically trivial-sounding: a foot that catches on stairs, a hand that drops keys, a voice that slurs at the end of a long day.

The focality is confirmed pathologically. In 14 of 19 autopsy cases, lower motor neuron loss was graded radially outward from the region where symptoms began. A 2026 Nature Genetics study adds a mechanistic candidate for why: rare somatic mutations were found spatially restricted to focal regions of ALS brain while driving widespread degeneration, a local genetic lesion acting as an epicentre.

The split hand is the most distinctive early sign: preferential wasting of the thumb-side muscles with relative sparing of abductor digiti minimi, despite shared innervation. Whether this is cortical or spinal in origin is unsettled. Threshold-tracking TMS shows the intracortical inhibition deficit is far worse for the wasted muscles, arguing cortical. A 2025 study of early hands found motor-unit loss beginning in parallel across all three muscles, with CMAP reduction of 25%, 22% and 7.7%, arguing instead for differential resistance of the spinal motor neurons themselves. Contested

Where it goes

Spread is mostly, but not entirely, to neighbouring regions, and the figure you get depends heavily on whether you measure symptoms or electrophysiology.

The Sekiguchi result is the important dissent. Needle EMG of paraspinal muscles found that in 39% of patients the disease skipped between muscles of similar motor neuron vulnerability rather than between neighbours, and the authors explicitly reject a "single seed and simple propagation" model in favour of multifocal initiation with local propagation.

How fast it spreads predicts how long the patient lives. Fujimura-Kiyono et al. (n = 150) found the interval from onset to involvement of a second region correlated significantly with survival, independent of which regions were involved. No patient whose disease reached two regions within three months survived five years. A prospective cohort found median survival of 71 months for horizontal (contralateral) spread against 25 months for vertical (rostro-caudal) spread.

Spread does not behave like a wave moving at constant speed. In hand-onset patients, the observed pattern matched a constant-velocity propagation model in only 14.8% of cases for a cortical route and 3.7% for a spinal cord route. In 41% of patients, long-distance hand-to-leg spread happened faster than short-distance hand-to-shoulder spread.

The four pathological stages

Brettschneider et al. (Annals of Neurology, 2013) proposed that TDP-43 pathology progresses through a fixed sequence.

1. Stage 1. Agranular motor cortex, brainstem motor nuclei (XII, VII, V), spinal α-motor neurons
2. Stage 2. Prefrontal neocortex, precerebellar (pontine) nuclei, red nucleus
3. Stage 3. Postcentral (sensory) neocortex and striatum
4. Stage 4. Anteromedial temporal lobe and hippocampus

Clinically this maps onto motor symptoms, then executive and behavioural change, then extramotor involvement, then memory failure and ALS-FTD. In vivo diffusion imaging can assign a stage to about 72% of patients, and 27% of those rescanned advanced a stage over six months.

The scheme has real problems. About 12% of cases cannot be staged without modified criteria. Robert Brown's objection is anatomical: ALS is primarily a disease of spinal motor neurons, and the spinal cord is under-sampled in the staging scheme. And a fundamental one: autopsy provides a static snapshot, so direction is inferred, never observed. A data-driven reanalysis of 571 cases classified cases the categorical scheme called unclassifiable, and found two distinct ALS subtypes, subcortical-predominant and corticolimbic-predominant, rather than one ordered sequence. Contested

Four competing models of how it travels

What happens at the neuromuscular junction

The neuromuscular junction is where the disease is doing its work long before anyone notices. Vulnerability follows motor unit type: fast-fatigable units degenerate first, then fast-fatigue-resistant, then slow. Slow units have the greatest sprouting capacity, and they reinnervate the muscle fibres orphaned by their fast neighbours.

That compensation is why weakness lags so far behind pathology. In 222 patients tracked with motor unit number index, the time to lose 50% of motor units was a median of 14.9 months, against 28.4 months to lose 50% of function. Motor unit loss reaches the halfway point in roughly half the time function does. Patients with early motor unit loss had a median survival of 32 months against 74 months for those with late loss.

Fasciculations and cramps, often the very first thing a patient notices and often dismissed, are the electrical signature of this process: denervated and reinnervating motor units firing spontaneously.

Section 12Progression and death

ALS is not one clinical entity, and the variants differ in survival by more than a factor of four. Where the disease starts and which motor neuron population it favours are the two strongest determinants of how long someone lives.

How fast it moves

The standard instrument is the ALSFRS-R, a 48-point functional scale. In the ENCALS cohort of 11,475 patients, the median progression rate was 0.6 points per month (IQR 0.3 to 1.1). The interquartile range is nearly fourfold, which is the whole problem with treating ALS as one disease in a clinical trial.

The model achieves a c-statistic of 0.78 and a three-year time-dependent ROC of 0.86. Its five prognostic groups have median survivals from 17.7 months to 91.0 months, and predicted values match observed closely across all five. Replicated

The diagnostic delay hazard ratio of 0.52 deserves a note, because it looks backwards. Longer delay predicts better survival, not because waiting helps but because slowly progressing disease takes longer to recognise. This is length bias, and it contaminates any survival figure measured from diagnosis rather than onset.

Two staging systems, measuring different things

King's clinical staging counts anatomical regions involved, plus two care milestones. The proportions of total disease course elapsed at each are sobering.

By the time a patient is given a name for what they have, more than a third of their illness is behind them.

MiToS functional staging counts lost independence instead, across four domains: swallowing, walking and self-care, communicating, breathing. Stage 0 is functional involvement without lost independence, stages 1 to 4 count domains in which independence is lost, and stage 5 is death.

The clinical variants

Chiò's population-based classification assigned a phenotype to 1,332 of 1,351 consecutive incident patients, or 98.6%, demonstrating that these are real separable entities rather than clinician impressions.

Primary lateral sclerosis has had formal consensus criteria since 2020. Probable PLS requires absence of significant lower motor neuron degeneration for two to four years from onset, definite PLS for four years or more. It matters clinically because the prognosis is entirely different from ALS.

How patients die

Phenotype changes this substantially. Respiratory failure caused 100% of deaths in respiratory-onset patients and 78% in flail arm. In ALS-FTD, respiratory failure accounted for only 33% of deaths while pneumonia rose to 27%. Aspiration displaces muscle failure as the terminal event when swallowing and judgement are both impaired.

The physiology follows a predictable order. Diaphragmatic weakness manifests first during REM sleep, then across all sleep stages, then in waking life. Nocturnal hypoventilation precedes daytime respiratory failure, which is why morning headache, disturbed sleep, fatigue and orthopnoea are the symptoms to ask about. Forced vital capacity below 1 litre or 30% predicted signals significant risk, and maximum inspiratory pressure may detect insufficiency earlier than FVC does.

Section 13Why it resists treatment

A 2026 network meta-analysis of 109 randomised trials and 16,353 participants found that no intervention class demonstrated a statistically significant survival benefit versus placebo. Thirty years of well-funded effort has produced three approved drugs, one of which applies to 2% of patients and one of which is not recommended by European guidelines. The reasons are specific and mostly structural.

1. The patient arrives too late

This is the deepest problem, and no drug can solve it. Motor unit number estimates reach 50% loss roughly 13 months before fine motor function does, because surviving motor neurons sprout and reinnervate orphaned muscle fibres, masking the loss. By the time weakness is noticeable, that compensatory reserve is exhausted.

Then diagnosis takes a median of 11 months more: three months from onset to first medical contact, six months from that contact to diagnosis. Against a median survival of two to four years, roughly 25 to 40% of the entire disease course is spent undiagnosed. A trial therefore randomises patients whose target cells are already dead. Neuroprotection cannot recover a dead motor neuron.

2. Heterogeneity defeats a single trial design

Progression rates span months to more than a decade. Onset sites differ, cognitive involvement ranges from none to dementia, and roughly 20 genes plus the exposome contribute in different proportions. A drug with a real effect in a subgroup comprising 20% of enrolment is statistically invisible in a 24-week ALSFRS-R trial.

This produces a pattern that recurs often enough to deserve naming: overall miss, subgroup signal. Masitinib, MIROCALS, lenzumestrocel, CNM-Au8, pridopidine, fosigotifator and NurOwn all follow it. Genuine heterogeneity produces exactly this pattern. So does multiplicity artefact. Distinguishing the two requires prospective replication, which is expensive and rarely done.

3. The mouse model does not model the disease

The SOD1-G93A transgenic mouse represents a genotype accounting for about 2% of ALS, and it does not reproduce the defining pathology of the other 97%, since SOD1 is one of the two TDP-43-negative exceptions.

Worse, the model's published results were largely noise. Scott et al. (2008) re-tested a series of compounds previously reported as beneficial under rigorous conditions, with littermate controls, sex balance, transgene copy-number verification, blinding and adequate power, and found essentially none of the reported survival benefits replicated. The original positives traced to transgene copy-number drift, sex differences in survival, litter effects, and severe underpowering. Minocycline, celecoxib and creatine all sit in this bucket. Minocycline went on to a human trial in which patients on drug declined faster.

4. Getting drugs into the central nervous system

Most systemic small molecules never reach therapeutic CNS concentrations. Every therapy with a clear mechanistic success, meaning tofersen, ulefnersen and QRL-201, bypasses the barrier by intrathecal administration. That solves delivery at a cost: 8.7% of tofersen-treated patients had serious neurologic events including myelitis, papilledema and chemical meningitis, and it requires repeated lumbar puncture at a specialist centre in patients who are losing mobility.

5. The mechanism is a network, not a chain

Genetic lesions, TDP-43 proteinopathy, RNA metabolism failure, nuclear transport defects, oxidative stress, mitochondrial failure, excitotoxicity, proteostasis collapse and neuroinflammation are interconnected and mutually reinforcing. Blocking one node in a cascade already in motion is unlikely to be enough. This is the strongest argument for combination therapy and for targeting upstream genetic causes rather than downstream consequences.

6. The endpoint is noisy

The ALSFRS-R sums bulbar, limb and respiratory domains that progress at different rates, with unequal item weighting, ceiling and floor effects, non-linear behaviour over time, and subjective scoring. A 2026 Neurology analysis documented substantial heterogeneity in how it is even analysed across trials. One widely cited editorial called it "an old friend who has overstayed their welcome."

Trials are also too short. VALOR is the object lesson: it missed at 28 weeks, but the open-label extension showed a 3.5-point separation at 52 weeks and a survival separation at five years. A 24-week trial can easily conclude that a working drug does not work.

The failure ledger

Section 14Current landscape

As of September 2026 the US-approved armamentarium is riluzole, edaravone, and tofersen for SOD1 carriers only. No new ALS drug was approved by the FDA in either 2025 or 2026.

Riluzole (1995)

A glutamate modulator, and still the only drug offered to every patient. Pooled across 4 RCTs and 1,477 participants it gives HR 0.84 (95% CI 0.698 to 0.997), about a 9% absolute gain in one-year survival probability, translating to 2 to 3 months of median survival. Small, real, cheap, generic, and available as tablet, oral suspension and dissolving film for patients who can no longer swallow pills. European guidelines grade it at the strongest level: offer it lifelong, to everyone, at diagnosis. Established

Edaravone (2017 IV, 2022 oral)

A free-radical scavenger, and the most contested drug in the field. Its pivotal trial ran only after an earlier trial failed, using inclusion criteria reverse-engineered from a post-hoc subgroup of that failure: definite or probable ALS, onset within two years, FVC at or above 80%, a score of at least 2 on every ALSFRS-R item, and a documented 1 to 4 point decline during a 12-week run-in. That is a small, early, respiratory-intact, fast-but-not-too-fast slice of the real population. Within it, ALSFRS-R declined 5.01 points against 7.50 on placebo over 24 weeks, a 2.49-point difference.

The independent evidence is unkind. A German multicentre propensity-matched cohort across 12 academic centres found ALSFRS-R decline of −0.91 vs −0.85 points per month (P = 0.37): no difference in function, survival, or time to ventilation, including in the subgroup that met the original trial's eligibility criteria. The EMA was not persuaded and the manufacturer withdrew the European application rather than take a formal refusal. European guidelines grade edaravone not recommended outside clinical trials. US list price runs above $150,000 a year. Contested

Tofersen (2023)

An intrathecal antisense oligonucleotide that degrades SOD1 mRNA. It is the field's clearest mechanistic success and also a case study in how hard these trials are to read.

VALOR, the Phase 3, missed its primary endpoint at −1.2 versus −3.0 ALSFRS-R points, p = 0.97 on the prespecified analysis. It hit the biomarkers hard, with CSF SOD1 protein down 38% and plasma NfL down 67%, and the FDA granted accelerated approval in April 2023 on that basis: the first time neurofilament served as a surrogate endpoint in a neurodegenerative disease.

The open-label extension is where the case was made. Comparing early starters to those who began six months later: a 3.5-point ALSFRS-R difference at 52 weeks, +9.2% predicted vital capacity, and HR 0.36 for death or permanent ventilation. Long-term data published in JAMA Neurology in December 2025, at a median follow-up of 4.9 years, reported that 27% of early-start participants gained muscle strength over roughly three years, something with essentially no precedent in SOD1-ALS, and that early initiation in fast progressors associated with a 3.4-year extension in event-free survival.

Relyvrio, withdrawn

The full arc is worth recording because it is the field's canonical regulatory cautionary tale.

Approval rested on a single small Phase 2, at roughly $158,000 a year, and was refuted by an adequately powered trial four times its size.

Elsewhere, masitinib has been refused by the EMA and is running a confirmatory Phase 3 of 495 patients. South Korea retained conditional approval for lenzumestrocel, an autologous mesenchymal stem cell therapy, in May 2026, but only after its Phase 3 failed overall and succeeded in a slow-progressor subgroup, with the label restricted accordingly.

The interventions that are not drugs

This is where the larger effect sizes are.

Non-invasive ventilation is the single best-evidenced intervention in ALS. Bourke et al. (Lancet Neurology, 2006), the only RCT, found that in patients with normal or moderately impaired bulbar function, NIV extended median survival by 205 days, about 7 months, at p = 0.006, with quality of life maintained for most of that period. The authors noted the benefit was "much greater than that from currently available neuroprotective therapy." That was 2006, and it remains true in 2026. In patients with severe bulbar impairment there was no survival benefit, only improved sleep symptoms. Established

Effectiveness depends on how it is delivered, not just whether it is prescribed. Ineffective NIV has been associated with halved survival, at 14 versus 30 months.

Multidisciplinary clinic care, meaning neurology, respiratory, nutrition, speech, physiotherapy, occupational therapy, social work and palliative care in one place, is associated with a mean survival advantage of 141.7 days (95% CI 61.5 to 221.9), about 4.7 months, larger in bulbar-onset patients. Roughly twice riluzole's effect, from organising care differently.

Gastrostomy is standard of care on physiological and quality-of-life grounds. Its survival evidence is observational and confounded by indication, since patients receive a feeding tube when they are deteriorating, and no randomised trial exists or ever will. Guidelines emphasise discussing it early and, if respiratory insufficiency is present, establishing ventilation first.

Exercise is not harmful and probably helps. A network meta-analysis of 25 RCTs and 1,226 patients found exercise the highest-ranked non-pharmacological intervention for ALSFRS-R at SMD 1.01 (95% CI 0.50 to 1.51), with benefits also on vital capacity, strength and quality of life. These are small unblindable trials with likely selection bias toward slower progressors, so the effect sizes should not be read as equivalent in reliability to the NIV trial.

Symptomatic management carries real quality-of-life weight: anticholinergics then botulinum toxin for sialorrhea, baclofen, tizanidine, gabapentin or cannabinoids for spasticity, mexiletine or gabapentin for cramps, dextromethorphan with quinidine for pseudobulbar affect. Pain is frequently under-treated and has no ALS-specific trial evidence base.

One development is worth noting for communication. A 2026 Nature Medicine report described a man with ALS using an intracortical speech brain-computer interface independently at home for nearly two years: over 3,800 hours across 444 of 653 days, nearly two million words at about 56 words per minute with greater than 99% accuracy on prompted tasks, with a care partner needed only to put the hardware on. It is a single case, and it is the first demonstration of durable independent home use. Single study

Section 15What works best

Ranked by demonstrated effect on survival and function, with the quality of evidence stated rather than implied.

The uncomfortable conclusion

For the roughly 98% of patients without a SOD1 or FUS mutation, the two interventions with the largest demonstrated effects on survival are a ventilator and a well-organised clinic, not a drug. Non-invasive ventilation buys roughly three times what riluzole buys. Multidisciplinary care buys roughly twice.

This has a practical implication that is often lost in coverage focused on drug pipelines. The largest available gains for most patients today come from access to specialist multidisciplinary care and timely, well-titrated ventilation. Both are delivery problems, not discovery problems.

What the two genetic successes have in common

Tofersen and ulefnersen are the only two therapies that have produced unambiguous mechanistic success, and they share four features.

• They target a known, single, upstream genetic cause rather than a downstream consequence
• They reduce production of the toxic protein rather than trying to clear or compensate for it
• They are delivered intrathecally, bypassing the blood-brain barrier entirely
• They were tested in genetically defined, homogeneous populations where a real effect is not diluted into statistical invisibility

That is a template, and it is why the cryptic exon programmes matter so much. They are an attempt to apply the same logic, meaning a defined molecular lesion, an antisense drug and intrathecal delivery, to sporadic disease, which is 90% of patients rather than 2.6%.

Section 16The question of regression

No therapy has reversed ALS. Every approved and investigational treatment in this report is measured by how much it slows decline, and the endpoints themselves assume decline: the ALSFRS-R is scored by how far a patient has fallen from 48, and survival models predict time to death or permanent ventilation. The field has no accepted instrument for measuring recovery because until recently there was nothing to measure.

That has changed at the margin, and the change is worth stating precisely because it is easy to overstate.

What has actually been observed

Two findings, both in genetically defined populations, both from 2025 and 2026.

Tofersen in SOD1-ALS. The long-term open-label extension published in JAMA Neurology in December 2025, at a median follow-up of 4.9 years, reported that 27% of early-start participants gained muscle strength over roughly three years. Gained, not retained. This has essentially no precedent in SOD1-ALS natural history, where strength declines monotonically.

Ulefnersen in FUS-ALS. In the Columbia expanded-access programme that preceded the Phase 3, neurofilament reductions reached 82.8%, and at least one participant regained ambulation and independent breathing. That is a single case in an uncontrolled setting, and it is also the most dramatic functional recovery reported in this disease.

Both are non-randomised observations. Neither constitutes proven regression. But both cross a line that thirty years of trials had not crossed, and both occurred in patients whose specific upstream genetic cause was being silenced.

Why partial recovery is mechanically possible

Dead motor neurons do not return, and nothing in these results suggests they do. The explanation lies in the compensation described in Section 11.

Muscle strength in ALS is not a direct readout of motor neuron number, because surviving motor neurons sprout and reinnervate the muscle fibres orphaned by their dead neighbours. That collateral reinnervation is what masks early motor unit loss, and it is the same mechanism running in reverse when a therapy works. Silencing mutant SOD1 does not resurrect lost neurons. It rescues the ones that are sick but alive, and those neurons then extend collaterals into denervated muscle. Strength recovers because the surviving pool reconnects, not because the pool grows.

This has a hard implication. The ceiling on recovery is set by how many motor neurons are still alive when treatment starts, which is why the entire tofersen dataset shows early starters doing better than delayed starters on every measure, and why ATLAS is testing intervention before symptoms at all.

Section 17The pipeline

The productive modality is antisense oligonucleotides, and the strategic question is whether success in genetically defined populations can be carried into sporadic disease.

Ulefnersen, the first unambiguous Phase 3 win

On 22 September 2026, two days before this report, Otsuka and Ionis announced that ulefnersen, formerly jacifusen, met the primary endpoint of the FUSION Phase 3 trial in FUS-ALS. The endpoint was a joint-rank analysis combining ALSFRS-R change, time to rescue, and ventilation-free survival through day 505, in a 72-week double-blind design. p = 0.0005. Serum neurofilament fell significantly, and time to death or permanent ventilation was significantly delayed.

This is the first ALS Phase 3 ever to hit a prespecified primary endpoint on a function-and-survival composite. Three caveats: the announcement is a company press release rather than a peer-reviewed publication, the effect magnitude has not been released, and the drug is not approved.

It is also a very small population. FUS-ALS is about 0.6% of all ALS, but it is a devastating one, accounting for an estimated 43 to 52% of juvenile-onset cases and typically killing within one to two years. The drug originated in Neil Shneider's expanded-access programme at Columbia, where neurofilament reductions reached 82.8% and at least one participant regained the ability to walk and breathe independently.

The cryptic exon programmes, the sporadic-ALS bet

QRL-201, the STMN2 splice-correcting ASO, has the strongest signal in sporadic disease. Interim ANQUR data reported in May 2026, across 52 sporadic and 15 C9orf72 patients: at the 10 mg dose, a 50% slowing of ALSFRS-R decline versus placebo at 8.5 months, a 5.08-point difference at p < 0.05, with 36% of that group stable or improved. CSF pNfH fell 46% and NfL 33% at the 3 mg dose. Target engagement was confirmed as at least a twofold increase in full-length STMN2, strongest in brainstem and lumbar cord. A pivotal study is expected in 2027.

These are small, interim, company-reported numbers from an open-label-risk setting. If they hold, this is the first therapy aimed at the molecular lesion shared by 97% of patients. Single study

UNC13A correction is following. Eli Lilly's LY4256984 is in Phase 1 dose escalation in about 32 sporadic ALS participants, and QurAlis licensed QRL-204 to Lilly for the same target. A dual-targeting snRNA gene therapy correcting both STMN2 and UNC13A cryptic splicing from a single vector has been reported preclinically.

Anti-TDP-43 approaches remain early. VectorY's VTx-002, an AAV-vectorised anti-TDP-43 antibody, dosed its first patient in February 2026.

C9orf72, the graveyard and what remains

Both sense-repeat ASOs failed, and failed instructively. BIIB078 (n = 106) lowered poly-GP with no clinical benefit and higher doses did worse than placebo. WVE-004 lowered poly-GP with no benefit and was discontinued. The lesson, documented in the Emory tissue study and a Brain 2024 post-mortem analysis, is that engaging a biomarker is not the same as reversing pathology.

What remains: TPN-101, a reverse transcriptase inhibitor with a positive Phase 2, now selected for the HEALEY platform; acamprosate in Phase 1; metformin in Phase 2; and PBFT02, an AAV gene therapy in C9orf72-FTD, which showed stabilised neurofilament and reduced brain atrophy at an April 2026 interim.

Other genetically targeted work

Immune and inflammatory approaches

MIROCALS (The Lancet, May 2025) tested low-dose IL-2 plus riluzole in 220 patients over 18 months. The primary survival endpoint was not met overall. In the prespecified subgroup with lower baseline CSF phosphorylated neurofilament heavy chain, about 80% of patients, risk of death was reduced by over 40%. Regulatory T cell expansion was robust and sustained. This is the subgroup-rescue pattern again, though here the subgroup was prespecified and biologically motivated.

COYA 302, combining low-dose IL-2 with abatacept, received FDA Fast Track in 2026 and launched a Phase 2 of about 120 patients. Ibudilast's COMBAT-ALS Phase 2b/3 (n = 234) completed its double-blind phase on 24 September 2026, with topline expected by year end. It is the fourth serious attempt at broad anti-neuroinflammation after minocycline, verdiperstat and ibudilast's own earlier Phase 2.

SAR443820, a RIPK1 inhibitor, reported safety, tolerability and efficacy results in ALS in 2026. RIPK1 sits at the junction of inflammatory signalling and regulated necrotic cell death, which makes it a more mechanistically targeted version of the anti-inflammatory strategy than the broad agents above. Sigma-1 receptor modulation is being pursued on similar logic, though pridopidine, a sigma-1 agonist, already missed as HEALEY regimen D.

The near-term regulatory events

CNM-Au8 is filing an NDA in early Q4 2026 under the accelerated approval pathway, based entirely on neurofilament plus long-term open-label survival. Both its randomised trials missed their primaries. The August 2026 package reported a 74% lower risk of death at 12 months versus concurrent HEALEY controls and HR 0.621 for NfL responders. Read critically: "NfL responder" is a post-randomisation, post-hoc stratification compared against non-randomised controls. The FDA has said the data may be capable of supporting accelerated approval while asking for more evidence linking the magnitude of NfL reduction to clinical benefit.

PrimeC, a fixed-dose ciprofloxacin and celecoxib combination, published PARADIGM Phase 2b in JAMA Neurology in March 2026: a 7.92-point ALSFRS-R advantage at 18 months for the early-treatment group, more than 36% slowing at p = 0.007, with the largest effect in the bulbar domain. It has FDA clearance to begin a pivotal Phase 3. The 18-month figure is an early-versus-delayed-start comparison in 68 patients.

NurOwn received the first Special Protocol Assessment ever granted for an ALS therapeutic in April 2024, for a Phase 3b of about 200 patients. As of August 2026 the trial is still not enrolling, more than two years later.

The HEALEY platform ledger

The platform trial is the field's structural response to heterogeneity and cost: a master protocol with a shared placebo arm, prespecified futility interims, and the ability to add and drop arms.

Seven regimens, seven misses. The platform has not yet produced a winner, but it has produced negative answers faster and more cheaply than seven separate Phase 3 trials would have.

Prevention, the most important unanswered question

Section 18Onset narratives

There are strong patterns in how ALS begins, and there is a strong pattern in how patients explain how it began. These are not the same thing, and separating them is one of the more useful things this literature can do.

What patients actually notice first

The first symptom is almost always trivial-sounding and almost always unilateral and distal. Fasciculations and cramps often precede weakness by months and are commonly dismissed by both patient and clinician. They are, in fact, denervated motor units firing spontaneously. Unexplained weight loss frequently precedes diagnosis, and this is not incidental: it reflects the hypermetabolic state described in Section 09.

Two sub-populations deviate strikingly from the usual onset distribution. Chiò's Italian footballers had 3 of 5 cases with bulbar onset at p = 0.003, and the ALS-plus-CTE brain bank subgroup showed 56% bulbar onset against 14% in ALS alone. Neither has been explained.

The prodrome is now a formal research entity

This is a genuine advance. Benatar et al. (Brain, 2022) followed presymptomatic gene carriers through phenoconversion and found that 11 of 20 converters passed through an identifiable phase of mild motor dysfunction before meeting diagnostic criteria. Its duration is genotype-dependent, as short as 1 to 3.5 months in SOD1 A4V and extending to several years in SOD1 I113T.

They proposed mild motor impairment as a distinct clinical entity, explicitly analogous to mild cognitive impairment in Alzheimer's disease, and formalised research criteria in 2025. Work is underway to extend the construct to sporadic disease and to test whether community neurologists can detect it. This matters because it names the window that every therapeutic strategy now depends on.

The path to diagnosis

Patients typically see three to four clinicians before reaching an ALS specialist. The median delay is 11 months, split roughly three months from onset to first medical contact and six months from that contact to diagnosis, and the delay is similar across countries with very different health systems. That suggests it is intrinsic to the difficulty of the diagnosis rather than a failure of any one system.

The bulk of the delay accrues while the patient is already under neurological care, particularly with non-neuromuscular neurologists. Paradoxically, seeking medical attention earlier predicts a longer time to diagnosis, because very early ALS genuinely does not look like anything.

The stories patients tell, and what the data says

Patients very commonly date the onset of ALS to a specific event: an injury, a surgery, an infection, an episode of extreme physical exertion, an electric shock, a period of severe stress. This is one of the most consistent features of ALS narratives, and the epidemiology gives a consistent answer about it.

The statistical signal is real, short-latency, and almost certainly pointing backwards. Four independent lines of evidence converge.

Read together, these describe a person in a prodromal phase who falls, fractures a wrist, loses weight, quits smoking on doctor's advice, presents to hospital, and is injured, and who afterwards, reasonably, identifies those events as the cause of what followed. The absence of dose-response for head injury is the clinching detail. Causes generally show dose-response. Consequences do not.

Where the origination stories are supported

The distinction that matters is between a single remembered event before diagnosis and a sustained multi-year occupational or environmental exposure. The second category contains genuine signals.

• Professional athletics. Italian footballers at SMR 6.5, rising to 15.2 for careers over five years, and NFL players at SIR 3.59 with OR 1.2 per season. These are prospective-by-construction occupational cohorts with internal dose-response on career length. They are not explicable by recall bias.
• Military service. A consistent 1.5 to 2-fold elevation, though the 2026 branch-and-rank data fits a toxicant-exposure model poorly.
• Selenium in drinking water at Reggio Emilia. Dose-response, plus reversal after exposure ceased. The strongest single-chemical cluster on record.
• Cumulative environmental burden. Pooled risk scores from persistent organic pollutants and metal mixtures predict both risk and survival, independent of polygenic risk.

Summary classification

Established The first-symptom distribution. Contiguous spread. The 11-month diagnostic delay and where it accrues. Genotype-dependent prodromal mild motor impairment. Elevated risk in elite contact and endurance athletes with career-length dose-response.

Emerging Mild motor impairment as a diagnostic entity in sporadic disease. Presymptomatic biomarker trajectories beginning years before onset.

Contested Whether premorbid leanness, high fitness and low resting heart rate are risk factors or prodromal features, where Mendelian randomisation favours prodromal. Whether BMAA caused the Guam epidemic.

Not supported That a specific remote injury, surgery, illness or stressful episode caused any individual's ALS. The latency structure of every dataset argues against it.

Section 19Open questions

The honest measure of a field is what its own members disagree about. These are live, falsifiable disputes, not gaps in the literature review.

Seven questions where the answer would change practice

1. Is ALS a multistep process at all? The six-step model has organised thinking about ALS genesis for a decade and is replicated across four continents. Foffani et al. (January 2026) show the age-incidence curve is exponential rather than power-law, with Bayes factors up to 2.3 × 10⁴ against the model, while correctly identifying cancer data as power-law. If they are right, the intuition that ALS requires a specific number of accumulated hits, and the derived claim that a mutation pre-pays some of them, needs rebuilding on a different mathematical footing.

2. Which comes first, nuclear loss or cytoplasmic aggregation? Seeding experiments say cytoplasmic aggregation precedes nuclear depletion. The dissociation argument says inclusion burden does not track degeneration, so the soluble nuclear deficit must be what kills. This determines whether a therapy should clear aggregates or restore nuclear function, which are opposite strategies.

3. Is nuclear pore injury upstream of TDP-43? The CHMP7 work places nuclear pore complex damage before TDP-43 mislocalisation in sporadic and familial ALS. Nearly every other framework in this report treats TDP-43 as the initiating lesion. Both cannot be right.

4. Dying forward or dying back? Cortical hyperexcitability is measurable before symptoms and now has causal mouse evidence. Neuromuscular junction denervation precedes cell body loss in rodents and now has human biochemical support from presymptomatic muscle proteins in plasma. The field has not converged, and the two are not mutually exclusive, but which is the therapeutic target differs entirely.

5. Where does the missing heritability live? Family and twin studies put heritability near 50 to 60%. Common variants account for 2.8%. Known genes explain about 15% of sporadic cases. The gap is not explained, and candidate answers including rare variants, somatic mutation, structural variation and gene-environment interaction have not been systematically tested against each other.

6. Does presymptomatic intervention work? ATLAS is the only trial that can answer this. If treating a presymptomatic carrier whose neurofilament has just risen prevents or meaningfully delays ALS, everything about how the disease is approached changes: screening, genetic testing uptake, trial design, regulatory pathways. If it does not, the window may be narrower than anyone hopes.

7. Is neurofilament a valid surrogate endpoint? It carried tofersen's approval and underpins Clene's pending NDA. But BIIB078 and BIIB105 both lowered their biomarkers and failed clinically, and post-mortem tissue analysis found that CSF biomarker change did not track CNS pathology. The field has made a large regulatory bet on a proposition it has not proven.

Things widely believed that the evidence does not support

• That the blood-brain barrier is broadly broken in ALS. The largest human dataset, n = 1,059, found albumin and immunoglobulin quotients indistinguishable from controls.
• That a remembered injury or stressful event caused an individual's disease. Every latency-stratified analysis argues against it.
• That premorbid leanness or athleticism is causal. Mendelian randomisation is null for BMI, and the observational signal is best explained by prodromal hypermetabolism.
• That the Guam epidemic is solved. BMAA remains a hypothesis with unresolved dose-plausibility and analytical-reliability objections, and competing explanations were never excluded.
• That the gut microbiome has an established role. There is no replicated taxon-level signature.
• That ALS is "the athlete's disease" generally. Elite professional athletics carries real elevated risk. Recreational fitness does not.

Where the next answer most likely comes from

1. ATLAS, because prevention would reframe the entire disease.
2. The cryptic exon programmes, QRL-201 and the UNC13A correctors, because they are the first attempt to apply the tofersen template to the 90% of patients who are sporadic.
3. Earlier detection, whether through presymptomatic plasma proteomic panels, the formalisation of mild motor impairment, or models reading electronic health records to flag patients before a neurologist would. Every therapeutic strategy in this report is limited by how late patients arrive.
4. Combination therapy, because if the mechanism really is a redundant network, single agents will keep producing the overall-miss-subgroup-signal pattern indefinitely.

How to read this literature

Three recurring patterns are worth carrying forward when assessing any new ALS claim.

The subgroup rescue. A trial misses its primary endpoint, and a subgroup analysis finds a significant effect. Masitinib, MIROCALS, lenzumestrocel, CNM-Au8, pridopidine, fosigotifator and NurOwn all did this. Genuine heterogeneity produces exactly this pattern, and so does multiplicity artefact. Only prospective replication distinguishes them, and it is rarely done.

The window artefact. Any exposure measured in the two to five years before diagnosis will show an inflated association, because that window is the prodrome. Effect sizes that shrink to null at longer latency are describing consequences, not causes.

Target engagement without benefit. A drug can demonstrably do the molecular thing it was designed to do and change nothing clinically. BIIB078, WVE-004 and BIIB105 all did. Biomarker movement is a necessary condition for a mechanism-based therapy, never a sufficient one.

Section 20Sources and method

All figures in this report come from sources retrieved and read on 24 September 2026. Where a number could not be verified at its primary source, it has been omitted rather than reproduced from memory. Several publisher sites, including PubMed Central, Wiley, Lancet, Neurology.org and Taylor & Francis, blocked automated retrieval. In those cases institutional repositories, author-accepted manuscripts or regulatory documents were used instead.

Known limitations

• Company press releases are labelled as such. The FUSION Phase 3 result, the QRL-201 interim data, the CNM-Au8 biomarker package and several 2026 pipeline updates are not peer-reviewed publications.
• Three widely quoted figures could not be verified and are flagged in text: the "50 to 80% of motor neurons lost before weakness" statistic, which traces to poliomyelitis research rather than ALS; the commonly cited 60 to 70% EAAT2 loss figure; and the exact compound counts in Scott et al.'s 2008 mouse model critique.
• Geographic coverage is uneven. There are essentially no population-based ALS genetic or incidence data from Sub-Saharan Africa, the Middle East or most of Asia-Pacific. Conclusions about ancestry-related variation rest on European and East Asian data.

On the registry evidence

Sections of this report draw on the trial and paper indexes in the Luminary Health platform, principally the posted-outcome data in Section 14 and the active NIV trials in the same section. Two observations about that source are worth recording, because they bear on how registry evidence should be read here and elsewhere.

Registry density is not therapeutic significance. A query for the most successful ALS trials returns edaravone, because one sponsor ran a long series of safety and extension studies and those studies dominate the record by volume. Trial count measures regulatory activity, not clinical benefit. The interventions with the largest demonstrated effects in this report, non-invasive ventilation and multidisciplinary care, generate almost no registry footprint at all, because nobody files an IND for a clinic.

Retrieval precision degrades on broad queries. The source set returned for the same query included trials in Alzheimer's disease, metastatic cancer and hereditary angioedema. Those were not used. Any registry-derived claim in this report was checked against the trial record itself.

Key sources by section