NMD Scanner — scanning selection model for NMD susceptibility

Isoform Scanner

Search both catalogues at once, or paste any sequence you like, including a transcript nobody has sequenced. The model needs the mature spliced transcript, 5′→3′.

1 · Find a transcript optional

Look in NMD Study Ensembl 115

Each source answers a different question. NMD Study is the study's own isoforms, whose sequence and exon structure this site already holds. Some of them carry an observed NMD label to check a prediction against; the rest are marked not measured. Ensembl returns sequence and exon structure together, so a transcript ID is enough on its own. NMD Study matches on partial words; Ensembl has no fuzzy search, so it wants an exact gene symbol or an ENST/ENSG identifier. Each source shows its five best matches; Show all on a source gives it the whole list and a box to search within it.

2 · Exon-exon junctions optional

1-based positions of the last base of each exon but the last. Filled in for you by either lookup. Leave it empty for a single-exon transcript, but note that the model leans heavily on junctions, so an unspliced guess will read as NMD-escaping.

3 · Transcript sequence required

Only A, C, G, T and U are read as bases. An N, an IUPAC ambiguity code or any other character matches no codon, so no start or stop is read across it and it scores nothing in the Kozak window. The position itself is still kept, and it holds its reading frame, any junction that falls on it, and its place in the GC average, where it counts as a non-GC base. The model was trained that way, so that is how it is shown.

Average all five trained members

Bulk Scan

Scores every transcript in an annotation and ranks them by P(NMD). Nothing is uploaded: both files are opened locally and the model runs in this browser, one worker per processor core.

Please upload two files: The model reads bases, since it has to find ATGs, score initiation context and encode 1000-nt windows, so exon coordinates alone cannot be scanned. Give it the GTF and the matching transcript FASTA and it joins them on transcript_id.

1 · Annotation GTF required

no file chosen

Needs exon rows carrying transcript_id and, where known, gene_name. Junctions are the cumulative exon lengths, reversed on the minus strand. .gz is read directly.

2 · Transcript FASTA required

no file chosen

One record per transcript, headers matching the GTF's transcript_id, spliced and already 5′→3′. Leave it uncompressed if you want to open transcripts outside the top twenty, since that needs a seek.

Most likely first ?Changes the order transcripts are scored in, never the scores themselves. It first makes one cheap pass, finding each transcript's ORFs and counting exon junctions more than 50 nt past the stop codon at about 0.3 ms each against roughly a second for the model, then scores in descending order of that count. Substrates tend to arrive early, so the list below settles sooner and stopping the scan costs you less. On a test of — real transcripts the true top twenty was complete after — of the scan rather than —. A scan run to completion gives exactly the same answer either way; if you do intend to score everything, leaving this off is about — quicker. Needs an uncompressed FASTA, since it has to jump around the file. Average all five members ?The model was trained five times from different random starts, giving five members. Unticked, every transcript is scored by member seed 100 alone. Ticked, all five score it and the results are averaged, which is the more reliable number, and it also tells you how much the members disagree. A wide spread means the transcript sits somewhere they do not agree about. It costs five times the work, so a scan that would take an hour takes five. One member is usually enough for ranking a whole annotation; tick this when you care about an individual transcript's number. Scan at most transcripts

Splice Junction Analysis

Splits a gene's isoforms into the ones that carry a given splice junction and the ones that do not, where carrying it means splicing that intron out rather than retaining it. It then compares how much RNA each group holds with NMD running (DMSO) against NMD blocked (SMG1i). RNA being destroyed by NMD is scarcer when NMD works, so a low ratio is the signature, and the difference between the two groups says whether the splice event changes the gene's NMD fate. Expression is the study's own NMD-anchored table, so the ratio is read as it stands: 1.0 is untouched, 0.5 is half destroyed. You pick one cell type at a time below. Nothing is pooled or averaged, and where the other cell types disagree the verdict says so by name.

1 · Gene required

Symbols as the study's long-read annotation spells them, over the study's — indexed isoforms. Nothing here is fetched from Ensembl.

2 · Cell type required

One at a time. Every isoform, ratio and share below belongs to that cell type alone, because the direction of a splice event's effect is not constant across cell types.

3 · Junction or exon required

One or the other, never both, so choosing in either menu clears the other. Junctions are quoted by the intron they remove, exons by their own first and last base, 1-based and inclusive as the GTF counts them, with how many isoforms have each. Or pick either straight off the models below: the cursor takes whatever it is over, an exon block or the intron between two.

Isoform models

Expression floor CPM DMSO SMG1i OR AND ?Which condition an isoform has to clear the floor in. OR keeps it if it clears the floor in either ticked condition, so an isoform that is silent under DMSO but accumulates under SMG1i stays, which is what most NMD targets look like. AND requires it to clear the floor in both, which is stricter and drops exactly those. With only one condition ticked the two settings are identical, because there is nothing to combine. With neither ticked there is no filter at all. The floor is a strict cut, so even its lowest setting means expressed at all rather than no filter. Untick both boxes for no filter. Drives the tables as well as the figure: each one says which floor and which rule it was computed under.

Every isoform of this gene that clears the filter, in genomic coordinates for the selected cell type, ordered by expression under DMSO with its CPM in both conditions on the right. Boxes are exons, the thin line between them is intron, chevrons give the direction of transcription. An orange dot marks an isoform the study calls unproductive in this cell type. Move over the figure to aim: the cursor takes whatever it is over, an exon block or the intron between two, and a click chooses it. Once a junction or an exon is chosen the tracks regroup into the ones that have it and the ones that do not, it is shaded down the figure and marked in blue on each isoform that has it, and the rest go pale. That shading is group membership, not a coding annotation. No exon here is shaded for being coding, because the study's long-read annotation carries no coding-region records to say which ones are.

Model card

The v6 scanning selection model developed for the study. Everything on this page is read out of the trained checkpoints that ship with this site, or copied from the study's own recorded results. The model and its training code will be released with the paper.

What it computes

A transcript carries an ordered list of candidate reading frames, 5′→3′. Each gets two numbers from sequence: capture p_k, the probability a scanning ribosome initiates here, read from the start codon window alone; and decay d_k, the probability the transcript is degraded given that translation begins here, read from both windows and the structural block. Selection is stick-breaking, so a ribosome reaches candidate k only by passing every earlier one:

P(select k) = p_k ∏_j<k (1 − p_j) · P(NMD) = ∑_k P(select k) d_k

The mass left over, where the ribosome passes every candidate, contributes nothing. That is why the bars in the prediction panel do not sum to one.

Performance

Held-out test AUROC — ± — across seeds, on — transcripts of which — are NMD susceptible, on chromosomes 1, 3, 5 and 7 with paralogs excluded. Taken from the study's input ablation analysis, using the model with every input retained.

Where the signal comes from

Each cell keeps only the named input groups and removes the rest: SEQ is the four base channels, JUNC the exon-junction channel, and STRUCT the structural block, which for this variant is the downstream exon-junction count alone. The floor is measured rather than assumed to be 0.5, because P(NMD) is a product over a candidate queue whose length correlates with the label, so a model given nothing informative still discriminates.

Verification

The model on this page is a JavaScript reimplementation of the study's own. It was checked against PyTorch 2.13 running the real checkpoint on identical inputs, stage by stage: the convolutions, the batch normalization, the binned max-pool, the fully connected embedding, the per-candidate capture and decay, the stick-breaking, and both output paths. Worst relative disagreement across — comparisons: —, which is float32 rounding rather than a difference.

Known limits, from the model's own authors

Three things worth knowing before you trust a number here.

The AUG window's fill boundary leaks ORF length. The window is filled only as far as the midpoint between start and stop, so the downstream fill extent is min(100, ORF length ÷ 2). For any ORF shorter than 200 nt, which is — of candidates, where the fill stops encodes the ORF's length exactly with no sequence involved. The study measures this fill saturation as a ~—× odds marker for "this is the real ORF". It is reproduced faithfully here because the model was trained with it and it is computable at prediction time, but it does mean capture should not be read as pure initiation context.
The decay head is handed the answer. Its only non-sequence input is the downstream exon-junction count, which is the premature stop indicator.
Every candidate has a start and a stop by construction, so the model never saw a negative example for either landmark and had no gradient from which to learn to check them.

Assumptions this page makes

About

A browser-resident implementation of the scanning selection model, so you can score a transcript that has never been sequenced.

The three tabs

Isoform Scanner takes one transcript at a time. Paste a sequence with its exon-exon junctions, or look one up from the study or Ensembl, and the model returns P(NMD) along with the candidate ORFs behind it.

Bulk Scan takes a whole annotation at once. Give it a GTF and its matching FASTA, and every transcript is scored in your browser and ranked by P(NMD).

Splice Junction Analysis asks what a splice event does to a gene's NMD fate. Pick a gene, a cell type, and a junction or exon, and the tab splits that gene's isoforms by whether they have it, then compares how much RNA each group holds with NMD running against NMD blocked. These numbers are measured rather than predicted, and the model has no part in this tab.

What NMD is, and what the model predicts

Nonsense-mediated decay is the cell's system for destroying faulty messenger RNAs. The model answers one question per transcript: would NMD destroy this? Its output is a probability. It reads only sequence and ORF structure, with no expression data and no cell type, so it can be applied to transcripts nobody has measured.

Where the training labels came from

Lung cells were treated with an SMG1 inhibitor, which switches NMD off, so transcripts that accumulate under SMG1i treatment are likely to have been destroyed by NMD. An isoform counts as NMD susceptible when its mashr local false sign rate is below 0.05 with a positive posterior mean log₂ fold-change in at least one of four cell types.

Reading the output honestly

Junctions matter more than sequence. The study's input ablation analysis puts the junction channel alone at — AUROC and the four base channels alone at —. If you paste a sequence without junction positions, you have removed the model's strongest single input and it will lean toward "escapes NMD". The page says so when that happens.
Check the split badge on study isoforms. A transcript in the training split was used to fit the model, so agreement there is not evidence.
The reference-start rescue cannot fire here. That rescue is the study's rule that a transcript's annotated start codon is always admitted, even when it scores below the Kozak floor. The annotation is not public, so a pasted transcript can get a slightly smaller candidate pool than the study built for the same sequence. It affected — of — candidates.
Five members, not one. —

Scope

— isoforms indexed from the study, across — genes, of which — are observed to be NMD susceptible.
Exon structures come from the study's long-read transcript annotation. The sequences themselves run to 1.6 GB and do not ship with this site, so an isoform lookup fills in the junctions and then either reads the sequence from a sequence store this deployment has been given, or leaves you to supply it. checking…
Any transcript in Ensembl release 115 can be scored, by exact gene symbol or by ENST/ENSG identifier. Release 115 is the annotation behind GENCODE v49 on GRCh38.p14, the gene set this site's Kozak floor was calibrated against, which is why it is pinned rather than following whatever is current. Sequence and exon structure arrive together, and the page checks the exon lengths sum to the length of the cDNA before pairing them. Ensembl's API has no fuzzy search, so a partial symbol finds nothing there while still matching in the study index.

What leaves your machine

The model, its weights and every calculation on this page are local, so a sequence you paste is never uploaded. That is the point of shipping the model rather than an API. Two things are outbound by construction, and both only when you use them.

The Ensembl lookup sends the same term to this site's own server, which forwards it to e115.rest.ensembl.org and returns the answer. It goes the long way round because Ensembl's archived releases send no CORS header, so a browser cannot call them directly, so pinning release 115 costs one hop. Your search term reaches Ensembl either way; what differs is that this site's server sees it too.
The study isoform lookup, on a deployment with a sequence store configured, sends the isoform ID you selected to this site's own server and gets that transcript's sequence back. It sends nothing you typed. Where no store is configured it makes no request at all. checking…

Each is behind its own checkbox above the search box, and an unticked source is never contacted. Untick them and the page makes no outbound request at all, though a pasted sequence still scores and everything else on the page still works.

Who made this

This site, meaning the browser port of the model, the isoform index and the Bulk Scan and Splice Junction Analysis tabs, is the work of Mateo Martinez, Yul Leshem and Peter Castaldi. The model itself and the study whose data it is trained and indexed on are cited separately, under How to cite.

Reusing it

The code behind this site is released under the MIT license, so take it, change it and build on it. That covers the browser implementation of the model, the four tabs and the build scripts. It does not cover the trained weights or the study data shipped with them, which come from the study's own sequencing and training, and whose terms the study sets. To reuse the weights or the data rather than the code, cite the study and check its terms first.

Data version

This build: loading….

Methods summary

What happens between pressing the button and seeing a number. Every step is the study's own, ported rather than reinvented.

Which set of transcripts a number is about

Numbers on this site count four different things, and they are not the same size. The largest is the indexed universe of — isoforms, everything the study measured as expressed, which is what the search covers. Of those, — carry an observed NMD call and made up the model's own universe, split into training, validation and test sets. The test AUROC on this page is measured on the — transcripts held out on chromosomes 1, 3, 5 and 7. A candidate is one reading frame rather than one transcript, so those counts run much larger, at — across the model's universe. The Kozak floor was calibrated on a fifth set that is none of these, the MANE Select transcripts of GENCODE v49.

1 · Enumerate

Every ATG with an in-frame stop codon downstream becomes a candidate. No minimum length, so an ATG immediately followed by a stop counts. Two ATGs in the same frame sharing one stop are separate candidates, as are ORFs in different frames. In the study this averaged — candidates per transcript before any filtering.

2 · Score initiation context

Each start codon is scored with the Cavener–Ray position weight matrix at labeled positions −6…−1, +4, +5 relative to the A of the ATG. Positions running off the transcript are skipped and the remainder summed.

3 · Admit

A candidate is admitted when it scores at or above the MANE calibrated Kozak floor (—, the 5th percentile of scores at MANE Select start codons, — of — scored under GENCODE v49) and starts in the first half of the transcript. If nothing is admitted, the five highest-scoring candidates are taken instead. Average admitted pool in the study: — candidates, median —, maximum —.

4 · Encode

Two 1000-nt windows per candidate: one anchored on the AUG, 900 bases upstream and 100 downstream; one anchored on the stop codon, 500 either side. There are nine channels per position: one-hot A/C/G/T, an exon-junction indicator, a 50-base centered rolling GC fraction, and three reading-frame channels relative to that candidate's own start codon. Each window is filled only up to the midpoint between start and stop, and zero-padded beyond without the anchor moving.

A position whose base is not A, C, G or T leaves the four base channels empty, and only those. The junction indicator, the three reading-frame channels and the GC average are written there as they are anywhere else, and the position counts in the GC denominator as a non-GC base. This is what the training encoder does, so it is what the site does. Leaving the position blank would read better, but it would move every prediction on a sequence containing an N away from what the weights were fitted to.

5 · Score and aggregate

Three convolutional encoders, one for capture and one each for the AUG and stop windows on the decay path, reduce each window to 32 numbers via two convolutions and a maximum within each of eight bins along the length. Binning rather than taking a global maximum is what lets position survive, since a global maximum records only that a pattern occurred somewhere and not how far along it was, and the distance from a stop codon to the next junction is the whole question here. The structural block enters through a linear layer, the three embeddings fuse to 64 dimensions, and stick-breaking over the candidate queue produces the transcript's NMD logit.

Expression · how the productive fraction is calculated

Everything above is the model. This is the expression table the Splice tab reads, which the model has no part in: it is measured, not predicted.

Two compartments per gene, per cell type

Within each cell type a gene's isoforms are split by NMD status: unproductive ones are the ones NMD degrades, productive ones are not. The call is the study's own and is made per cell type, so an isoform is unproductive where mashr gives a local false sign rate below 0.05 with a positive posterior mean log₂ fold change, meaning it rose measurably when NMD was blocked. A gene's unproductive compartment in a sample is the summed expression of those isoforms, its productive compartment the rest.

Why the SMG1i arm has to be adjusted

CPM divides by the library total. Under SMG1i the unproductive isoforms accumulate massively, which inflates that denominator and makes every productive isoform read smaller than it is. That is an artifact of composition rather than a fact about the gene, and the adjustment below removes it.

The adjustment, in three steps

For a gene g with productive isoforms prod(g) and unproductive isoforms unprod(g). First, the productive compartment is summed under DMSO:

P_g^DMSO = Σ_{i ∈ prod(g)} CPM_i^DMSO

Second, each unproductive isoform is given the proportion it occupied within that gene under DMSO, and that share is scaled to the gene's productive compartment in the SMG1i sample s:

share_i = CPM_i^DMSO ⁄ P_g^DMSO, i ∈ unprod(g)

CPM_i,s = share_i × Σ_{j ∈ prod(g)} CPM_j,s^SMG1i

Third, each adjusted library is rescaled back to one million, with the sum in the denominator running over the reassigned values from the step above rather than the originals:

CPM_i,s^adj = CPM_i,s × 10⁶ ⁄ Σ_k CPM_k,s

An unproductive isoform whose gene has no positive DMSO productive compartment has no share to take, and is held at its DMSO mean instead. Summing CPM^adj over prod(g) gives the gene's adjusted productive CPM, in which the unproductive surge's contribution is gone.

The productive fraction

prodCPM and unprodCPM are the sums of CPM over a gene's productive and unproductive isoforms and totalCPM is the two together, so the productive fraction is that gene's productive share of its own output:

f = prodCPM ⁄ totalCPM

What was checked

Control genes, the ones with no unproductive isoform to surge, come out essentially unchanged between DMSO and adjusted SMG1i, at Spearman ≥ — with a slope near 1. Gene rankings also survive, at Spearman ≥ — between adjusted and raw fold changes. Together those say the adjustment removes a compositional artifact rather than reordering the biology.

How this relates to the share the Splice Junction Analysis tab prints

Two differences, both deliberate, and worth knowing before the tab's number is quoted against f above.

It is per group, not per gene. That tab splits one gene's isoforms by whether they carry a junction and computes a productive share within each side, so it answers "how much of this half of the gene is productive" rather than the gene's own f.

Its SMG1i arm is hybrid. The shipped per-condition table carries anchored values on productive rows and observed values on unproductive ones, so it sums to 1.03–1.09 million rather than to one million. The tab divides that column as it stands: an anchored numerator over a total whose unproductive part is still raw. The DMSO arm has no such issue, being a proper CPM library summing to one million. Undoing the anchoring on the productive arm instead, one factor per cell type (—), moves the SMG1i share about 0.5–1.3 points lower. The tab says so under its own tables.

What is not done here

The study always admits a transcript's annotated start codon, which it calls the reference-start rescue. That annotation is not public, so the rescue never fires on this page. The Kozak floor is also recorded to four decimal places rather than at full precision; exactly one achievable score sits inside that rounding interval, so only starts scoring precisely −1.2507921 could be admitted differently.

Isoform Scanner

Bulk Scan

Progress

Most likely to be destroyed by NMD

Starred

Find any transcript

Splice Junction Analysis

Isoform models

Groups

Productive share

Model card

What it computes

Performance

Where the signal comes from

Verification

Known limits, from the model's own authors

Assumptions this page makes

About

The three tabs

What NMD is, and what the model predicts

Where the training labels came from

Reading the output honestly

Scope

What leaves your machine

Who made this

Reusing it

Data version

Methods summary

Which set of transcripts a number is about

1 · Enumerate

2 · Score initiation context

3 · Admit

4 · Encode

5 · Score and aggregate

Expression · how the productive fraction is calculated

Two compartments per gene, per cell type

Why the SMG1i arm has to be adjusted

The adjustment, in three steps

The productive fraction

What was checked

How this relates to the share the Splice Junction Analysis tab prints

What is not done here

How to cite

Sources behind this build

Contact