The Live Variant Playground

The objection. logSED is a log ratio, and those three genes have by far the lowest baseline coverage — 77, 213 and 502 against up to 48,547. A given absolute change in predicted coverage moves logSED 161× more at GAL3 than at TDH3, so low-baseline genes should show large log effects whatever the biology.

The control, from the data itself. Baseline does not separate the two groups cleanly, and where it overlaps the sign still tracks regulation: DTD1 (baseline 299) and MMS2 (556) sit between GAL3 (213) and GAL1 (502), and both are firmly negative (−0.45, −0.42). Two pairs at matched baselines with opposite signs — so the sign is about regulation, not scale.

The magnitude is a different matter, and the orderings invert. As a fraction of its own baseline the largest lever on the page is HOP2's +160%. In absolute predicted coverage HOP2's best substitution is worth +125 while PDC1's costs −8,841, seventy times more. Neither reading is wrong; they answer different questions. logSED is deliberately a ratio — that is what makes a silent promoter and a maximal one comparable — so "a stronger effect" is meaningless here without saying stronger relative to what.

What the old 500 bp window would have missed. It covers 3.1% of the sequence and holds a median 17.6% of the window's total |logSED| — genuinely well chosen, 5.7× enriched. But the strongest substitution falls outside it in 4 of the 14 windows, including HOP2's +1.38, the largest effect in the whole set.

One fold, and that is the paper's choice. Figure 4's ISM reads .../f0c0/part<N>/scores.h5 — a single model, fold 0 of the cross-validation, which is the checkpoint this page runs. The repository does ship an 8-fold ensemble loader, but it belongs to Figure 7's eQTL analysis, not to these logos. Averaging eight folds here would move this page away from the figure it reproduces, so it does not.

The score is a different formula from the paper's, and they agree. Figure 4 averages per-track logSED across the 384 _T0_ tracks; this page computes logSED of the track-averaged coverage. Those are not the same quantity — a mean of logarithms is not the logarithm of a mean — so it was measured rather than assumed. Over 240 real substitutions in TDH3's promoter the two agree to r = 0.999987, with a largest difference of 0.00033, the same strongest substitution, and a scale ratio of 1.005. That is far below what the shipped uint8 packs can even represent. The reason is that the T0 tracks are strongly correlated; for a heterogeneous track set the two would genuinely diverge.

What is not checked. The paper's own ISM arrays live on the authors' cluster and are not in the public repository, so nothing here is a numeric comparison against published values. The claim is that the recipe matches — the fold, the track subset, the gene-body slice, the rc-averaging, the mean-centring and the one-hot projection — and that the drawing reproduces the published figure.

A solid box is a motif; a hollow box is a database region. Curated binding-site calls come from a PWM scan under a conservation filter, not from matching a consensus string — so most of them do not contain their factor's canonical motif. Measured across these windows, only 22.8% of 670 curated calls do. Crucially their coordinates are not wrong: where the consensus is present it sits at offset 0 of the call. So a box is drawn on the match where one exists, tight to the motif and with the matched bases printed beneath it — the reference-DB convention of the paper's own Figure 4F — and on the database's region otherwise, hollow and labelled as such. Two different claims, drawn differently.

Splice and codon landmarks are spans, not a fixed window. They used to be boxed as six bases centred on the junction, which framed AAGGTA where DTD1's donor motif is GTATGT. Each landmark now carries the span of the motif it names — donor 6 bases, acceptor 2, codons 3, branch point 7 — mirrored on the minus strand, and a test decodes the sequence inside every drawn box to confirm it.

The consensuses come from Figure 4H, the paper's motif dictionary. Recording them explicitly caught one already: this page had Fhl1 as GTAAACA, which is not the paper's Fhl1 motif, and was boxing a position 27 bp from where Figure S19B puts it.

The geometry is the paper's, not an approximation of it. Real DejaVu Sans Bold outlines for A/C/G/T with the paper's own hand-tuned per-letter x-offsets (A −0.350, C −0.366, G −0.384, T −0.305 — not half-advances), scaled by globscale = 1.35 on both axes, which is what makes a stack summing to S fill height S and what makes adjacent letters touch. Colours are the paper's saturated X11 values #008000 / #0000FF / #FFA500 / #FF0000, fixed in every theme because a figure's base colours are part of the figure. Letters are stacked descending by magnitude, positives up from zero and negatives mirrored below, with y-limits at the data min and max each padded by 5% of max|v|.

Mutagenesis is the paper's Figure 4 quantity, in three steps (fig4_common.py:227-230, and identically in two other files): average logSED over the T0 tracks → mean-centre across the four bases → project on the reference one-hot. Step two is what turns "what this substitution did" into "how much the base that is there matters"; step three is why exactly one letter survives per position. Given the shipped plane P with the reference cell zero, that value is exactly −(Σ P)/4.

Gradient × input also draws one letter per position, but for a different reason: the input is one-hot, so gradient × input is identically zero at the three bases that are not there. That is a property of the method, not a rendering choice.

Occlusion resolves 64 bp, so every base inside a window carries that window's value and the logo reads as blocks. Its sign is flipped to match the others: occlusion measures what is lost, so a base that matters has a negative logSED, and the logo convention is that up means "raises the prediction".

All four answer the same scalar, so they can be laid on one axis: f = log2( Σgene bins mean384 T0 tracks coverage + 1 ). That is the paper's own logSED target (ensemble.py:97-104): mean over tracks, sum over bins, +1 pseudocount, log base 2. Being a log ratio is what makes a silent promoter and a maximal one comparable.

1. Gradient × input. One backward pass of f w.r.t. the one-hot input gives [16384 × 4]. Mean-centre across the four bases (g − meanACGT g) — the Borzoi convention — then multiply by the input and sum the base axis. Because the input is one-hot, three of every four values are exactly zero, so this is one number per position. Cost: 1 backward pass. Reads as: if I nudged this base up a little, how would the prediction move. Fails when: the model is locally flat but globally sensitive — a saturated promoter can show near-zero gradient at a base that matters enormously.
2. Integrated gradients. The same gradient, accumulated over 32 steps along a straight path from an all-zero-DNA baseline (species channel kept, because a baseline with no species is not a sequence this model was trained to see), then multiplied by x − baseline. Not mean-centred, deliberately: completeness — that the values sum to f(x) − f(baseline) — is a telescoping integral of the raw gradient, and centring destroys it. Measured, centring pushed the error from ~0.05 to 8–650%. Cost: 32 forward+backward per region. Reads as: this base's share of the whole difference between the real sequence and an empty one. Unique property: it is the only method here you can check against its own total, and the panel prints that check.
3. Attention rollout. The eight attention matrices, each mixed half-and-half with the identity to account for the residual stream, renormalised and composed: ∏ (½Aℓ + ½I). Row i then reads as where position i's representation came from. Cost: free — the maps already ship. Reads as: what the transformer can see for this region. Not an attribution: it says what could influence what, never what changed the prediction, and it is unsigned by construction.
4. Occlusion. Zero the four DNA channels across a 64 bp window and re-run; the change in f per output bin is that window's effect. Cost: 256 real forward passes × 2 strands. Reads as: what the model loses without this stretch. Fails when: two stretches are redundant — remove either alone and nothing happens, which occlusion reports as "neither matters".

Both strands, averaged, for all four, matching Borzoi and the --rc every published Shorkie ISM run passes. Worth knowing what that is: this model was not trained with reverse-complement augmentation (augment_rc: false in all four params.json), so it is not rc-equivariant — measured on TDH3 the target reads 15.60 forward against 14.23 reversed and the two gradients correlate at 0.31. Averaging them is a test-time augmentation, not a free symmetry.

How much do they actually agree? Measured on TDH3's own gene body, at the 64 bp grid occlusion resolves: gradient × input against integrated gradients r = 0.874, gradient × input against occlusion r = 0.901, integrated gradients against occlusion r = 0.856. The three full-window methods agree strongly, which is the reassuring part. Against mutagenesis, per base over its own 500 bp, gradient × input falls to r = 0.658 — and that is expected rather than alarming: a gradient is a local linear sensitivity and a substitution is a finite jump to a different base. Where a promoter is saturated the two genuinely differ.

One thing the curve above does not share. The drawn coverage is the 3,053-track RNA-seq group mean; every attribution scores the 384 _T0_ subset the paper uses. Measured, the two correlate at r = 1.0000 and differ by 1% at the peak, so the distinction does not affect reading — but it is a distinction.

What full-window mutagenesis costs, and why it was once refused. 16,384 × 3 substitutions on both strands is 98,304 forward passes a locus and 1,376,256 for all fourteen. An earlier round priced that at 39.6 h and dropped it — but that was measured through onnxruntime on the CPU, on a graph whose batch axis is pinned at 1, so it baked in both the slowest engine available and the impossibility of batching. Neither limit belongs to the model. Re-measured on the same machine's GPU at batch 32, with the output head sliced to the 384 tracks actually scored: 10.47 ms a forward pass — and a substitution costs two of them, because every published Shorkie run averages both strands. Measured over the run that produced the packs on this page: 23.8 ms a substitution, 19.5 minutes a locus, 4.6 h for the fourteen loci that run covered. The numbers on this page are that run. The engine change is not a precision compromise either — against the CPU the GPU agrees to 6.6 × 10⁻⁷ relative, three orders of magnitude tighter than the fp16 graph the earlier packs were built from.

For each of 256 windows of 64 bp: zero the four DNA channels across that window, run the model, and record log2(alt + 1) − log2(ref + 1) for every one of the 896 output bins. One forward pass answers a whole row, which is why a complete [256 × 896] map costs 256 passes — 27 s a locus, 54 s with both strands — and why this is the cheapest exact method here.

Per cell the unit is logSED, a log2 fold change, so bins of wildly different expression are comparable within a row. The default colour scale is shared across the whole matrix for that reason; the checkbox rescales each output bin to its own range, which trades the between-bin comparison for the within-bin one and is worth it when a few loud bins dominate.

The margins are plain sums of |logSED|: the top strip is each input window summed over all 896 bins, the right strip each output bin summed over all 256 windows.

The diagonal is not the identity line. The two axes cover different spans — the input is the whole 16,384 bp window, the output only the cropped 1,024–15,360 interior — so it is a line of slope <1 that reaches neither corner. It is drawn dashed where it actually falls.

Zeroing is not shuffling. Zeroing the DNA channels is how the paper's language model masks a position and is indistinguishable from a run of N: it asks whether the stretch carries information at all. The motif knockouts above shuffle, which preserves base composition and asks whether the arrangement matters. Different questions, and they can disagree.

What occlusion cannot see: redundancy. If two stretches carry the same information, removing either alone changes nothing, and this map reports both as unimportant.

Every view in this panel starts from the same tensor and collapses it differently. For one stage, relevance = |∂f/∂a ⊙ a| over its activations a: [channels × positions] — how much each neuron, at each place, moved the prediction over the selected bins. That array is far too large to ship for every region, so what is precomputed is its two margins:

1. Channel margin — sum over position, giving [channels]. "Which neurons mattered, wherever they fired."
2. Position margin — sum over channel, then sum-pooled to a common 128, giving [128]. "Where in the window this stage drew from, whichever neuron did it."

Both are exact, and both superpose: gradients are linear in which outputs you select, so an arbitrary contiguous region is the row-sum of the precomputed 112 groups of 8 bins — no model run, no interpolation. That is why dragging is instant and still exact.

The map's interior is not exact. Reconstructing [channels × positions] from two margins requires an assumption, and the one used is independence: the outer product, normalised to sum to 1. It is the unique reconstruction consistent with both margins, and it is still an assumption. The neuron traces below the map exist precisely because they are not: each is one channel's real activation, straight from the shipped stage maps.

The stage stack is the position margin, one row per stage, each row scaled by its own 99th percentile and then painted against its own mean — red above average, blue below, neutral where the stage has no positional preference. Per-row scaling because stages differ by orders of magnitude; against the mean because the early residual blocks are genuinely near-uniform once pooled to 128, and a from-zero ramp draws that truth as a saturated bar reading "maximally relevant everywhere".

Three stages are missing from all of this, and that is not an oversight: the input one-hot, the conv stem and the output head live on their own tensors in the exported graph and have no per-layer relevance in the pack. They show their activations and say so.

The enrichment number on each neuron trace is the share of that channel's Σ|activation| falling inside the traced region, divided by the share of the window the region occupies. 1.0× means the neuron fires there no more than anywhere else. Measured on block 4, the top-relevance channels come out 0.85–1.32× — so a channel's relevance to a region generally comes from what it computes there, not from firing only there.

Nothing here is new data. The per-class enrichment is the same statistic the panel below computes, run once per window offline and the medians taken; the knockout winners are read from the same -ko.json packs the sweep panel reads. It is a second view of numbers that already exist, not a second copy of them — verify_pipeline re-derives it from the packs and fails if the two disagree.

The TSS profile aligns on the direction of transcription, not on coordinate order: the start is txStart on the plus strand and txEnd on the minus, and minus-strand profiles are reversed so "upstream" is the same side of the plot for every gene. Without that flip the average puts promoters against terminators and produces a flat curve that looks exactly like a real null result.

Each gene is normalised to its own total before averaging, so a highly expressed gene does not dominate. The band is the spread across genes, not a confidence interval on the mean — these genes differ by orders of magnitude in expression and the spread is the honest statement of that.

What it cannot tell you. 158 genes from nineteen windows is a small, non-random sample: they were chosen as classic high expressers, the galactose regulon and the paper's own figure panels. A metaprofile from them describes those genes, not the yeast genome.

Per site: Fisher–Yates shuffle the bases inside the site, re-run the model, and record log2(Σ coverage + 1) over the gene body minus the same quantity on the unedited sequence. Repeat with k = 5 independent permutations and report mean ± sd. Cost is one forward pass per shuffle — 104 ms — so the whole ChIP-supported tier across all fourteen windows is a few thousand passes.

Why not one shuffle. A single permutation is one draw from the distribution of "sequences with this composition". Two shuffles of the same site can differ by more than the effect being reported, so the single number the page used to show was a measurement with an unstated error bar of unknown size. The sd here is that error bar, measured.

Why the gene body, not the peak. A 16 kb yeast window holds a dozen genes and the tallest is rarely the one whose promoter was edited. Measuring globally reports a number about an unrelated gene — on one window the global peak is 114.3 while the named gene's own body peaks at 7.8.

What it cannot tell you. Shuffling a site does not remove the possibility that the model reads a different, overlapping element; sites here overlap frequently. And a null effect at a real binding site can mean the model is robust to it, that a redundant site elsewhere covers it, or that the site is not used in the condition the T₀ tracks represent.

The statistic. Take the per-base signal s over the 16,384 bp window and a binary mask m for the class. The ratio is (Σ|s|·m / Σm) / (Σ|s| / L) — mean magnitude inside over mean magnitude everywhere. It is scale-free, so two methods with wildly different units are directly comparable, and it is a magnitude, so a class the model suppresses still reads as enriched. The mean signed value inside is reported alongside so that direction is not lost.

The null is a circular shift, not a resample. The mask is rotated by each of 256 evenly spaced offsets and the ratio recomputed. Rotation preserves the feature count, every feature's length, and the spacing between them — so the null compares against an annotation that looks exactly like the real one but is not aligned to the signal. Resampling positions instead would compare against a feature set that does not resemble the real one at all, and would call almost everything significant. The offsets are deterministic (evenly spaced, zero excluded), so a published number is reproducible rather than a draw.

The p is empirical, (#{null ≥ observed} + 1) / (256 + 1), so it is never exactly zero and its floor is 1/257 ≈ 0.0039. A cell at that floor means "no shift out of 256 reached this", not "p = 0".

What it cannot tell you. Circular shifts are a weak null against broad positional trends: if a method's signal is concentrated near the window centre and a class happens to be too, the shift will not fully break that. It is also a per-class test with no multiple-testing correction across the table — read a single marginal cell with that in mind.