The research, led by Meixi Wang, Sheng Bi, and colleagues at Xuanwu Hospital of Capital Medical University in Beijing and published in the European Journal of Nuclear Medicine and Molecular Imaging, examined 359 amyloid-positive individuals imaged with the widely used tracer 18F-Florbetapir. Among them, 52 showed visible amyloid-beta deposition in the cerebellar cortex itself, while 307 did not. The team combined PET imaging, structural MRI, and cerebrospinal fluid biomarker analysis to ask a deceptively simple question: what happens to standard amyloid measurements when the reference region is itself contaminated with amyloid?

The answer was striking. Participants with cerebellar cortical amyloid deposition had significantly lower Centiloid values than those without, despite all being amyloid-positive. The mechanism involves a dual bias: tracer uptake in the whole cerebellum was elevated by direct amyloid binding in the cerebellar cortex, while cortical tracer uptake was paradoxically reduced, possibly reflecting advanced-stage processes such as synaptic loss and declining soluble amyloid in cerebrospinal fluid. Dividing a lower cortical signal by a higher cerebellar signal systematically underestimates the true cerebral amyloid burden, potentially misleading clinicians about how much plaque a patient actually carries.

The quantitative details matter. Whole-cerebellum standardized uptake values were significantly higher in the cerebellar-positive group, cortical uptake values were significantly lower, and Centiloid values dropped dramatically, with bootstrap analysis confirming the differences were far from chance. By contrast, uptake in the pons did not differ significantly between groups, and when cerebral white matter was used as the denominator, the resulting SUVR showed no significant between-group difference at all. White matter, long considered a secondary option, emerged as the only reference region that remained stable in the face of cerebellar pathology.

TheCentiloid scale, anchored at zero for young healthy controls and 100 for typical Alzheimer’s patients, has become the lingua franca of amyloid quantification. It underpins threshold decisions in clinical practice and serves as a primary endpoint in trials of anti-amyloid antibodies such as lecanemab, donanemab, and aducanumab. The new findings suggest that in patients with cerebellar involvement, Centiloid values cannot be taken at face value. A patient whose Centiloid score appears to fall below the conventional amyloid-positivity threshold of 30 may still carry substantial cortical plaque burden, simply because the denominator of the ratio has been inflated by cerebellar amyloid.

To help clinicians detect this problem, the team established and validated a quantitative diagnostic threshold. Using cerebellar SUVR normalized to white matter, receiver operating characteristic analysis yielded an optimal cutoff of 0.428, with an area under the curve of 0.931, sensitivity of 86.5 percent, and specificity of 86.3 percent. In an internal validation subset of 30 patients with equivocal visual readings, agreement between the quantitative threshold and expert visual assessment was substantial, with a Cohen’s kappa of 0.856. The negative predictive value of 97.4 percent was particularly strong, positioning the metric as an effective screening tool for ruling out cerebellar amyloid involvement.

The longitudinal component of the study adds a provocative therapeutic dimension. Two amyloid-positive patients with cerebellar cortical deposition were treated with lecanemab and followed with repeat PET at seven and eleven months. In both, Centiloid values decreased substantially over time, yet cerebellar SUVR normalized to white matter rose, remaining above the 0.428 threshold throughout follow-up. In one patient, the Centiloid value fell from 70.9 to 22.4, crossing below the recommended positivity threshold even though visual assessment confirmed persistent cortical amyloid and unchanged cerebellar deposition. Reading such scans through a Centiloid lens alone could falsely suggest near-complete amyloid clearance.

The persistence of cerebellar amyloid despite antibody therapy has a plausible biological explanation. Unlike the cerebral cortex, where compact cored plaques predominate, the cerebellum is characterized by diffuse amyloid deposits. Autopsy findings from the historic AN-1792 immunization trial showed that cerebellar diffuse plaques resisted clearance even when neocortical plaques were removed. Because passive antibodies such as lecanemab engage microglial clearance pathways that depend on Fc-gamma receptor-mediated phagocytosis, pathways more effective against compact plaques than diffuse ones, the cerebellar signal may simply not respond to treatment in the same way as the cortex. The new data are consistent with that framework.

Cerebellar amyloid is not a rarity confined to exotic cases. According to the Thal staging system, cerebellar cortical involvement marks phase five of Alzheimer’s pathology, the most advanced stage. In autosomal dominant Alzheimer’s disease caused by Presenilin-1 mutations, PET studies have detected cerebellar amyloid roughly a decade before symptom onset. And a recent autopsy study in the oldest-old suggested cerebellar cortical deposition may be more common in aged clinical populations than previously appreciated, precisely the group now being considered for anti-amyloid therapy. As treatment eligibility increasingly hinges on quantitative amyloid thresholds, accurate measurement in these patients becomes a pressing clinical concern.

The authors caution that the 0.428 threshold was derived in a single-center cohort, the longitudinal analysis included only two patients, and the findings require replication with other amyloid tracers and in independent multicenter samples. Nevertheless, the practical implications are immediate. Where cerebellar amyloid is suspected, white matter offers a robust alternative reference region, and the cerebellar SUVR threshold provides a standardized, readily implementable check on cerebellar status. The study also redefines what makes a reference region valid: freedom from amyloid cannot be assumed globally, but must be verified for each patient and disease stage. As quantitative amyloid PET becomes central to the era of disease-modifying Alzheimer’s therapy, ensuring that the yardstick itself is clean may prove just as important as the drug being measured.

Subject of Research: Bias in 18F-Florbetapir amyloid PET quantification caused by cerebellar cortical amyloid-beta deposition

Article Title: Cerebellar cortical amyloid deposition biases 18F-Florbetapir PET quantification with whole cerebellum as reference region

Article References: Wang, M., Bi, S., Xue, H., Guan, L., Wang, Y., Zhang, X., Liu, X., Xu, X., Zhang, C., Qi, Z., Zhao, Z., Yan, S., & Lu, J. (2026). Cerebellar cortical amyloid deposition biases 18F-Florbetapir PET quantification with whole cerebellum as reference region. European Journal of Nuclear Medicine and Molecular Imaging. https://doi.org/10.1007/s00259-026-08140-6

Image Credits: AI Generated

DOI: 10.1007/s00259-026-08140-6

Keywords: Alzheimer's disease, amyloid-beta PET, 18F-Florbetapir, cerebellar amyloid, Centiloid scale, reference region, white matter, SUVR, lecanemab, quantification bias, nuclear medicine, diagnostic threshold

A-to-I RNA editing is catalyzed by the ADAR family of enzymes, primarily ADAR1 and ADAR2, which chemically convert adenosine bases to inosine within RNA transcripts. Because the cellular machinery reads inosine as guanosine, this single change can rewrite the meaning of a gene: it can swap amino acids in proteins, alter splice sites, create or destroy start and stop codons, reshape RNA folding, and modulate microRNA targeting. In humans, millions of editing sites dot the transcriptome, with the vast majority clustered in repetitive elements such as Alu sequences, where paired inverted repeats form the double-stranded RNA substrates that ADAR enzymes prefer. Editing in these repetitive regions also helps cells distinguish their own RNA from foreign double-stranded RNA, preventing inappropriate immune activation.

The functional stakes are high. Aberrant A-to-I editing has been implicated in amyotrophic lateral sclerosis, autism spectrum disorder, schizophrenia, epilepsy, depression and multiple cancers, and dynamic editing levels are known to vary across tissues, developmental stages and species. Because editing levels, defined as the fraction of adenosines converted to inosine at a given site, are the currency of nearly every study in this field, accurate measurement is foundational. Yet while considerable effort has gone into developing methods to identify editing sites from sequencing data, far less attention has been paid to whether the quantification of editing levels itself is trustworthy.

To test this, the team, led by Shi Cheng and Yulong Song of Guangzhou Medical University alongside colleagues at Sun Yat-sen University, performed matched short-read and long-read cDNA RNA sequencing on two human cell lines, HEK293T and U2OS. The short-read data were generated on an Illumina NovaSeq 6000 platform using both polyA-selected and rRNA-depleted library strategies, producing reads of roughly 150 nucleotides. The long-read data were produced on Oxford Nanopore Technologies PromethION sequencers, which capture full-length cDNA molecules with median clean read lengths of 812 to 878 nucleotides. Both approaches were anchored to a curated reference list of 2,858,028 human A-to-I editing sites compiled from the group’s previous work.

The results were striking. In HEK293T cells, long-read sequencing identified 3,897 editing sites measured significantly higher than by polyA-selected short-read sequencing, against only 445 sites measured higher by the short-read approach, an approximately 8.8-fold difference in the number of sites. Globally, the median editing level estimated by long reads was about 2.3 times that from short reads. The pattern held for sites in Alu elements, in repetitive non-Alu regions and in nonrepetitive regions alike. Most dramatically, at nonrepetitive sites the median short-read estimate was 0 percent, implying no editing at all, whereas long-read sequencing revealed a true median level of about 2.1 percent. Parallel comparisons using rRNA-depleted short-read libraries, and independent experiments in U2OS cells, reproduced the same directional bias, with long reads detecting roughly 6.6 to 9.0 times more significantly higher-measured sites.

To rule out artifacts, the researchers ran an extensive battery of controls. They raised coverage thresholds from 30 to 40, 50 and 60 reads per site, tightened the base-quality cutoff for long reads, and removed PCR duplicate reads; the underestimation by short reads persisted in every case. As expected, no meaningful difference emerged between polyA-selected and rRNA-depleted short-read libraries, or between the two cell lines, confirming that the bias tracks with sequencing technology rather than biology. A full-length DNA and RNA amplicon sequencing experiment provided direct validation: DNA amplicons, which cannot carry RNA edits, showed editing levels near zero, while RNA amplicons and long-read cDNA sequencing agreed with each other and both substantially exceeded the short-read estimates.

The mechanistic explanation converged on read length and its effect on genome alignment. Because short reads carry less sequence context, they are more likely to map to the wrong place in the reference genome, particularly within repetitive regions where many loci look alike. By computationally constructing artificial short-read datasets of 50, 75, 100 and 150 nucleotides from their own 150-nucleotide data, the team showed a strongly positive correlation between read length and measured editing level: longer reads yielded higher estimates, with fold differences reaching 12.1 between 150-nt and 50-nt reads. Crucially, when they truncated reads at the alignment stage rather than the sequence stage, recalculating mapping coordinates so that no re-alignment occurred, the correlation vanished. This pinpointed misalignment, not any property of the sequences themselves, as the culprit.

The alignment analysis went further. As read lengths decreased, millions of reads shifted from uniquely mapped to multiply mapped or unmapped statuses, and the counts of these misalignment categories correlated negatively with measured editing levels and with the Alu editing index, a widely used global metric computed with the RNAEditingIndexer tool. The same positive relationship between read length and editing quantification appeared in public short-read data from HEK293T, U2OS, HeLa, HepG2, K562, A549 and U87 cells, and in artificial datasets derived from long-read data segmented to shorter lengths, indicating the effect is universal across platforms and laboratories.

The generality of the finding was then tested against independent public resources. A paired short-read and long-read dataset covering eight human lung cancer cell lines showed higher editing quantification by long reads in every line, with nonrepetitive sites again registering zero percent under short reads versus roughly two percent under long reads. Cross-study comparisons using long-read data from GM12878 and HeLa cells against ENCODE and other short-read datasets revealed 23.8-fold and 6.4-fold excesses, respectively, of sites measured higher by long reads. The authors conclude that long-read RNA sequencing is preferable for accurate quantification of A-to-I editing levels, while acknowledging its higher cost; for laboratories constrained to short reads, they recommend raising coverage depth, restricting analyses to uniquely mapped reads, or validating key sites with full-length amplicon or Sanger sequencing.

Beyond editing, the study raises a caution for transcriptomics at large. Since read-length-dependent misalignment degrades any measurement that depends on precise genomic placement, splicing and gene expression estimates from short-read data may also harbor systematic errors that warrant closer scrutiny. Given that A-to-I editing shapes nearly 85 percent of the human transcriptome and is an emerging target for RNA-based therapeutics, the message is clear: the technology chosen to read RNA does not merely observe biology, it shapes what biology appears to say.

Subject of Research: Comparison of short-read and long-read RNA sequencing for accurate quantification of A-to-I RNA editing levels

Article Title: Short-read RNA-seq yields lower estimates of A-to-I RNA editing levels than long-read cDNA sequencing

Article References: Cheng, S., Qi, Y., Ya, J., Xia, L., Zhang, W., Xiong, Q., Liu, Q., Zhang, J., & Song, Y. (2026). Short-read RNA-seq yields lower estimates of A-to-I RNA editing levels than long-read cDNA sequencing. Advanced Biotechnology, 4(3), Article 27. https://doi.org/10.1007/s44307-026-00123-w

Image Credits: AI Generated

DOI: 10.1007/s44307-026-00123-w

Keywords: A-to-I RNA editing, RNA-seq, short-read sequencing, long-read sequencing, ADAR, alignment accuracy, read length, transcriptome, Alu repeats, cancer cell lines, Oxford Nanopore, quantification bias