The Live Variant Playground

The window, and the one number every method differentiates

Question

What does the model take in, and what single scalar does every method on this page differentiate?

Why this matters

Every panel below differentiates, perturbs or ranks against one number. If that number is not fixed and stated, two panels can disagree because they were quietly explaining different quantities — the commonest way an interpretability page contradicts itself.

Method

A 16,384 bp window centred on one gene. f(x) = log2( Σ over that gene's bins of the mean over the 384 T0 RNA-seq tracks, + 1 ) — the paper's own logSED target, ensemble.py:97-104.

Intuition

A model with 5,215 outputs has no single 'prediction'. Scalarising to one gene's own bins over one declared track subset is what makes 'which bases drove it' a well-posed question at all.

Cost

Free. The sequence and the gene models ship with the page.

What would refute it

Nothing here is a claim; it is the definition everything below is measured against.

What the answer tells us Nothing is claimed here; this is the definition. Its value is that every number below is a change in this one quantity, so where two panels disagree they are disagreeing about the model and not about what was measured.

Each locus is a 16,384 bp window — the model's full context — centred on the named gene. Six of them are the windows Figure 4 of the paper prints, at the published coordinates; the rest are classic high-expressers and the galactose regulon. In goes 16,384 × 170 channels — 4 DNA, 1 unused, 165 species — and out come 896 bins of 16 bp across 5,215 tracks, with 1,024 bp cropped from each end.

Every method below differentiates the same scalar, and that is the only reason they can share an axis or be compared at all:

f(x) = log₂( Σgene bins mean384 T₀ tracks coverage + 1 )

That is the paper's own logSED target (ensemble.py:97-104): the mean is over tracks, the sum is over bins, the pseudocount sits inside the log, and the base is 2. The order matters — a mean of logarithms is not the logarithm of a mean. Being a log ratio is what makes a silent promoter and a maximal one comparable, and it is why “a stronger effect” on this page is meaningless without saying stronger relative to what. The 384 _T0_ tracks are the uninduced baseline subset the paper's Figure 4 uses, not the whole 3,053-track RNA-seq block.

Before asking which bases drove it — how many bases can? effective receptive field

Question

How much of the 16,384 bp of context does the prediction actually depend on?

Why this matters

Every attribution below is scoped by what the model actually uses. If a prediction is settled by the nearest kilobase, an attribution peak six kilobases away is noise — and a method that reports one is wrong rather than interesting.

Method

Keep a centred core of real sequence, dinucleotide-shuffle everything outside it, and grow the core until the prediction settles inside ±5%.

Intuition

Architectural reach and dependence are different claims: the transformer connects the whole window, but 'can see' is not 'uses'. Shuffling a flank while preserving its dinucleotide composition removes the information without removing the sequence.

Cost

5 shuffles × 8 radii = 40 forward passes a locus, 23 windows.

What would refute it

Two controls. At the largest radius the window is entirely real, so that point must reproduce the full-context prediction — and convergence requires every larger radius to stay inside the band too.

What the answer tells us This is the practical scope of every attribution on the page. A peak outside the radius that settles a locus is not a long-range dependency — it is a base the prediction does not use, and reading it as regulatory would be reading the method.

Every attribution method on this page distributes credit across all 16,384 bases, and the architecture really can reach all of them — dilated convolutions into full self-attention at the bottleneck. But can see and depends on are different claims, and only the second is measurable. Keep a centred core of real sequence, replace everything outside it with dinucleotide-shuffled DNA, and grow the core until the prediction stops moving.

The effective context is a property of the gene, not of the model. The median window is settled by ±2,048 bp of the ±8,192 available, and the six constitutive glycolytic genes — TDH3, PGK1, ACT1, ADH1, FBA1, PDC1 — are all finished by ±1 kb: a housekeeping promoter determines its own gene. GAL1 needs the entire window, and GAL3, FUN12 and KRE33 need ±4 kb. That is the practical scope of every attribution on this page — outside a window's own radius there is little left to attribute.

Predicted coverage — the full model this window, 16,384 bp

Question

What does the model predict across this window, what drove it, and what is actually annotated there?

Why this matters

This view replaced four stacked panels that each computed their own left inset, so one base pair sat at four different x positions and no two lanes could be compared by eye. A shared axis is what makes 'these methods disagree here' a statement about the methods.

Method

One forward pass for the curve (896 bins × 16 bp); every attribution lane rc-averaged and drawn at its own real step; annotation from SGD, ORegAnno, Harbison/MacIsaac and a JASPAR scan.

Intuition

Every lane answers a different question about the same coordinates. Drawing them on one axis at each method's own real resolution lets a divergence be read directly instead of inferred across two figures at different scales.

Cost

Free. Every locus was run offline and the packs ship with the page.

What would refute it

This is a drawing, not a claim. Its correctness condition is that one base pair lands at one x in every lane — which one canvas makes true by construction rather than by assertion.

What the answer tells us Where the lanes part company is the finding, and on one axis it is readable directly rather than inferred. A method drawing a peak no other method draws is either resolving something the others cannot, or reporting its own estimator.

Everything on this page is about ONE window at a time — the locus in the bar above. This is what the model predicts across it, what drove that prediction, and what is actually annotated there, on one shared axis so a peak can be read against a gene. For the same quantities across all 12,157,105 bases, and for the language model beside them, the genome browser is the other half of this lab.

Drag to pan, drag the ruler to select, scroll to zoom — and keep going. Past about seven pixels a base the three per-base attribution lanes stop being bars and become DNA logos — mutagenesis, gradient × input and integrated gradients, each in place, each still the same lane — and the sequence appears beneath them. Nothing new pops up: the same three measurements are simply drawn where every base has room for its own letter. Occlusion and attention rollout stay as bands, because 64 bp and 128 bp a value cannot be letters and stretching them into some would claim a resolution neither has. Every lane is the same base pair at the same x, so a peak can be read straight down against a motif and a gene. Click anywhere to trace the 16 bp bin under the pointer; the shaded band is what the region-conditioned panels below are keyed on — gradient x input, integrated gradients, occlusion and rollout, but NOT mutagenesis, which keeps its own focal-gene target and is unconditional. The flanks are shaded because the head crops 1,024 bp from each end — the curve covers 896 bins of 16 bp and the model still reads the whole 16,384 bp to produce them.

Every attribution lane is signed, and that is the whole reading. A bar above the rule is a base whose presence RAISES the prediction, below it one that lowers it. Each lane is drawn at its own real step and labelled with it: mutagenesis and the two gradient methods resolve a single base, occlusion 64 bp, attention rollout 128 bp. Nothing is stretched into a resolution it does not have.

What is actually there, from the databases. Genes, CDS, introns, uORFs, tRNA and snoRNA genes, replication origins, LTRs and transposons come from SGD; regulatory regions from ORegAnno; binding sites from the Harbison 2004 / MacIsaac 2006 conserved regulatory code and from a JASPAR PWM scan. Provenance is drawn, not just recorded: a ChIP-supported site is solid, a conserved-only call is hollow, a PWM match is a hairline — three different strengths of evidence that must not be read as one. A grey cap marks a feature clipped by the window edge, so that edge is never mistaken for a real boundary.

Nothing here waits on a model. Every preset locus was run offline at the full 16,384 bp context and its predictions ship with the page. Loading the model — a 28.6 MB download and about 17 s of WebAssembly a run — buys the live layer activations, sequence editing and motif knockouts, not a curve that already exists. A dashed line, once you edit, is the unedited reference.

Every output track

Question

Is the curve above representative, or an artefact of averaging thousands of experiments?

Why this matters

The headline curve is a mean over thousands of experiments, and a mean can be shaped by a handful of loud tracks. If it is, every attribution built on it is explaining an artefact of averaging.

Method

All 5,215 predicted tracks as one row each, in the released targets sheet's order, with any single one plotted on the shared axis.

Intuition

A summary statistic cannot report on its own adequacy. Drawing all 5,215 rows shows structure within a block if the group mean is unrepresentative, and a uniform block if it is not.

Cost

Free. The 5,215 × 896 plane ships per locus, 2.15 MB.

What would refute it

If the four assay blocks did not read as bands the track ordering would be wrong — which it was once, and the curve labelled RNA-seq was mostly ChIP-exo.

What the answer tells us Where a block is uniform the group mean is a fair summary and every attribution built on it inherits that. Where it is not, the mean is the wrong object and the single-track view is the honest one.

One row per predicted experiment, in the order the released targets sheet lists them, so the four assay blocks read as bands. Click a row to plot that single track below — the overview above averages thousands of experiments, which is a curve no instrument would produce. Where rows outnumber pixels each drawn row is the maximum over the tracks it covers.

Four attributions, and one that is not

Question

Which bases drove the prediction over the region you traced?

Why this matters

Attribution methods are routinely presented as interchangeable views of one truth. They estimate different mathematical objects, and treating a disagreement as noise hides that one of them may be answering a different question.

Method

Four estimators of the same scalar, all rc-averaged: a finite difference (mutagenesis), a local derivative (gradient × input), a path integral (integrated gradients) and a window ablation (occlusion).

Intuition

Any single map is a choice of estimator presented as a fact. Four that must coincide wherever the model is locally linear turn that choice into a measurement: their agreement bounds how much of the map belongs to the model rather than to the method, and their divergence localises where the model is doing something no first-order summary can carry.

Cost

Mutagenesis 23.8 ms a substitution → 19.5 min a locus; IG 32 forward+backward passes; occlusion 512 passes; gradient × input one backward pass.

What would refute it

They are not expected to agree. On TDH3's gene body grad × input against IG is r = 0.874, against occlusion 0.901, IG against occlusion 0.856 — and against mutagenesis per base, 0.658, which is the gap between a slope and a finite jump.

What the answer tells us Agreement between estimators of different mathematical type is evidence about the model; disagreement is evidence about where the model stops being locally linear. Neither is noise, which is why the page draws all four rather than averaging them.

The lanes themselves are in the viewport above — every method on one zoomable axis, becoming letters once a base is wide enough to be a letter. What is here is what the drawing cannot say: what each method measures, what it costs, and where they disagree.

Three of them draw letters, and they do not all draw the same number. A logo of the whole 16,384 bp window is not a drawing that exists — one base is 0.078 px — so letters appear only in a zoom. Mutagenesis ships every substitution, so it can stack up to four letters in a column; the checkbox above switches between that and the paper's projection onto the base that is actually there. Gradient × input and integrated gradients both multiply by the input, and the input is one-hot, so both are identically zero at the three bases that are not there and draw exactly one letter a column. That asymmetry is a fact about the methods, not a simplification of the drawing. Attention rollout resolves 128 bp and occlusion 64 bp, so neither becomes letters at all.

And they are not all scored on the same thing. Mutagenesis is logSED on this window's own gene body and needs no selection — the plane is a property of the locus. The other three are conditioned on whatever region you traced. So the mutagenesis lane routinely peaks over a different part of the axis than the lanes below it, and reading that as disagreement between methods is reading a difference in the question. Each lane names its own scoring target on its right.

Four methods, one axis, one annotation. They do not measure the same thing and are not expected to agree everywhere: mutagenesis substitutes one base and re-runs the model; gradient × input is a local linear sensitivity at the reference base, mean-centred across the four bases as Borzoi does; integrated gradients accumulates that gradient along a path from an all-zero baseline and is the only one here whose values sum to the prediction difference — a property it would lose if it were mean-centred, so it is not; occlusion ablates a whole 64 bp stretch and measures what the model loses. Where they agree the result is robust; where they diverge, the divergence is the finding. A method that covers only part of the window is drawn in place with the rest left blank, which is what the paper does for its own partial windows.

Traceback. Gradient × input: how much the prediction over the selected bins moves per unit of each input base, times the base that is actually there. It is a local sensitivity, not a decomposition — the values do not sum to the prediction. Dragging is exact rather than interpolated, because gradients superpose: the attribution for a set of bins is the sum of their individual gradients. Dragging resolves to 128 bp; picking a gene gives single-base resolution.

Two sources, and they are not the same method. Mutagenesis is the paper's Figure 4 quantity: every substitution actually run, then mean-centred across the four bases and projected on the reference, so the height reads as “how much the base that is there matters” rather than “what this one substitution did”. It now covers every base of the window — all 49,152 substitutions per locus, on both strands — so it is the primary track here and needs no caveat about where it stops. Gradient × input is a local linear sensitivity — it is identically zero at the three bases not present, because the input is one-hot, so it draws one letter per position. The paper's published figures use no gradients at all; its single gradient routine is unreached scaffolding that still declares a 4-channel input the 170-channel model cannot accept. Gradient × input is here because it is instant, not because it is what the paper did.

What the whole window shows that a promoter window cannot. Running all 16,384 bp on both strands for every locus makes one result visible that no single window does: the sign of the strongest substitution follows the gene's expression state in 22 of 23 windows. The three genes repressed in this baseline — HOP2 (meiosis-specific), GAL3 and GAL1 (galactose regulon, in glucose) — all have a largest substitution that raises predicted expression (+1.38, +1.12, +0.78 logSED, with 20–33 of their top 40 positive), and nineteen of the twenty expressed genes are negative. For a gene already being transcribed the model's biggest lever is breaking what drives it; for a silent one it is breaking what holds it down.

The exception is worth more than the rule. PWP1 is expressed here and its largest substitution is nonetheless +0.223. It is a ribosome-biogenesis gene of the RRB regulon, and Supplemental Figure S20B labels this very window with both of that regulon's repressor motifs — the PAC motif and RRPE. A positive lever on a gene carrying two repressor elements is mechanistically reasonable, but it was classified as expressed before the number was computed, so it stands as an exception to the rule as stated rather than as a confirmation of it.

The five lanes, side by side

Question

What is each method actually measuring, and where can it not be trusted?

Why this matters

A reader comparing lanes needs to know which differences are findings and which are properties of the estimator. Resolution, sign convention and scoring target all differ, and any of them can masquerade as disagreement.

Method

A side-by-side of mathematical nature, resolution, cost, completeness, mean-centring, sign and scoring target — every entry measured elsewhere on this page.

Intuition

Tabulating each estimator's own properties separates 'these methods disagree about this locus' from 'these methods cannot agree, because one resolves 64 bp and the other resolves one base'.

Cost

Free. Nothing here is a new run.

What would refute it

A row with a blank failure mode would be a method with no known failure mode, which does not exist.

What the answer tells us Read any divergence against this table before reading it as biology. Two lanes that cannot resolve the same feature are not in conflict; two that can and still differ are.

Four of the five differentiate the same scalar, which is what lets them share an axis. Attention rollout is the exception, and the table says so on every row: unsigned, summing to nothing, reporting what the transformer can read rather than what moved the prediction. It sits beside the other four because that contrast is the point, not because it is a fifth estimator. The four that do share the scalar differ in every other respect, and the differences are not stylistic — a method's resolution decides whether it can draw letters, its mean-centring decides whether its values sum to anything, and its scoring target decides which part of the window it peaks over.

The bottom row is the one to read first. Every method here has a failure mode, they are different failure modes, and that is precisely why the page draws five lanes instead of picking a winner. Where four of them agree the result is robust; where one dissents, its own row usually says why.

What drives what

Question

Which input stretches does which output region depend on?

Why this matters

Every other method here collapses the output axis to one scalar, which discards where in the gene an input stretch acts. Local and long-range effects then look identical.

Method

Zero the four DNA channels across a 64 bp window, re-run, and read all 896 output bins at once — the complete 256 × 896 matrix.

Intuition

One ablation answers all 896 output bins at once, so the complete input-region by output-region matrix costs no more than a single-scalar sweep. Two dimensions are affordable here and nowhere else on this page.

Cost

256 ablation windows × 2 strands = 512 forward passes a locus. The only two-dimensional method here — one pass answers every output bin at once.

What would refute it

The diagonal dominates per cell — but it is one window in 256, so summed over a row the local footprint is only 0.2–5.6% of the most damaging window's total. Either half alone is the wrong answer.

What the answer tells us A narrow diagonal says the model is local, and the mass off it says how much of an effect travels. Reporting only one would make Shorkie look either purely local or purely long-range, and it is neither.

Every input window against every output bin. Each column is a 64 bp stretch of the input, ablated; each row is one of the 896 output bins; the colour is what that bin loses when that stretch is gone. Blue is a loss, red a gain. This is the only genuinely two-dimensional view on the page — every other method collapses to a single profile over the input — and it is measured, not inferred: 256 ablation windows on both strands, 512 real forward passes a locus.

Read it as a map, and read both parts. The dark diagonal is local effect — a window damaging the output directly above it — and per cell it is by far the strongest thing here. But it is also narrow: one window of 256. Summed across the row, the local footprint accounts for only 0.2–5.6% of the most damaging window's total effect, so the diagonal is where the model is most intense and the off-diagonal is where most of what it does actually lives. Neither statement alone is the truth. Off the diagonal, a vertical stripe is an input stretch many outputs depend on and a horizontal stripe is an output bin that reads widely.

The edges are the margins. The strip along the top is the column sum — how much each input window matters summed over every output — and the strip down the right is the row sum, how much each output bin depends on the sequence at all. Click a row to read that one output bin's input profile, or a column to see which outputs that window drives; click again to clear. Bins differ in expression by orders of magnitude, so the checkbox rescales each row to its own range when the raw map is dominated by the loudest ones.

Ablation zeroes the four DNA channels, which is exactly how the paper's language model masks a position and is indistinguishable from a run of N. That is a different question from the motif knockouts above, which shuffle: a shuffle keeps base composition and asks whether the arrangement matters, while zeroing asks whether the stretch carries information at all.

Does the attribution land on real biology?

Question

Does the attribution land on features the model was never shown?

Why this matters

An attribution map is a claim about the model, not about biology. Without an external reference it cannot be distinguished from a smooth function of position, and every peak looks meaningful.

Method

Mean |attribution| inside a feature class over the mean across the window, against a null of 256 deterministic circular shifts of the annotation — which preserves feature count, length and gap structure and destroys only alignment.

Intuition

The annotation was never shown to the model, so alignment with it is evidence the model found something real. The null must destroy alignment and nothing else — a circular shift keeps the feature count, every length and every gap, so it cannot be beaten by geometry alone.

Cost

Free. Both the annotation and the packs ship.

What would refute it

A class at 1.00× says the model puts no more weight there than anywhere else. Every class it can measure is shown, not the ones that came out well.

What the answer tells us Alignment with an annotation the model never saw is the first evidence here that an attribution is about sequence function rather than about the estimator. It remains an association under one null, not a demonstration that the model uses the feature.

One number per method, per annotation class. For each class, the mean |attribution| on the bases that class covers, divided by the mean across the whole window — so 1.00× means no preference. The null comes from circularly shifting the annotation against the signal, which keeps the number of features, their lengths and their spacing exactly and destroys only their alignment; the shaded figure beside each ratio is that null's spread. Cells are shaded by how far the ratio sits from its own null, not by its size. The table measures every evidence tier whatever the lanes above are drawing — comparing the three tiers against each other is the point, and a drawing toggle should not silently change a statistic.

This table is computed from one checkpoint. The attribution underneath it is gradient × input on fold f0, whose per-base signal survives retraining for 85% of the strongest bases and 54% of matched controls — act 9 measures it. An enrichment ratio inherits that: it is a statement about where this training run puts its weight.

Read the negatives too. A class at 1.0× is a real answer: it says the model puts no more weight there than anywhere else. And a method cannot resolve a feature finer than its own resolution — occlusion moves in 64 bp steps and attention rollout in 128 bp, so neither can say anything about a 7 bp binding site, which is why their columns are marked. This is a correlation against a stated null, not evidence the model uses the feature: a site that happens to sit inside a region the model attends to for unrelated reasons scores exactly the same.

Every curated site, knocked out

Question

Is an attribution peak a real motif, or base composition?

Why this matters

A peak on a binding site could be its arrangement or merely its base composition — poly-A tracts and GC-rich stretches draw attention for reasons that are not regulatory.

Method

Permute the exact bases inside each curated site — same length, same composition, only the arrangement destroyed — and measure the change over the window's own gene body, as a mean over 5 shuffles with its spread.

Intuition

Permuting bases inside the site holds length and composition exactly and destroys only the order. An effect that survives the permutation was never about the motif.

Cost

5 shuffles a site, two forward passes each — forward and reverse complement, averaged per bin before the gene body is summed — around 50 sites a locus. Same rc-averaging as every attribution on this page, so the numbers are directly comparable.

What would refute it

A near-zero result is an answer about that site. A low-complexity site cannot be shuffled at all, so the sweep records how many DISTINCT permutations it achieved rather than reporting a zero spread as a measurement.

What the answer tells us An effect that survives a composition-preserving permutation was never about the motif's arrangement. A near-zero result is therefore an answer about that site rather than a failed measurement — provided the site had distinct permutations to draw from at all.

Not six hand-picked motifs — all of them. Every ChIP-supported binding site in this window was shuffled and the model re-run, five times per site, and what is reported is the mean logSED with its spread. One shuffle is one sample; a single draw presented as a measurement is exactly what this replaces. Scored over the window's own gene body on the paper's 384 T₀ tracks, both strands — the same quantity the interactive knockout reports, using the same seeded shuffle, so a swept value and a clicked one agree exactly rather than approximately.

Where a gene has a well-known regulator, the sweep tends to find it, unprompted. Ranked by the size of the effect, the strongest site is RAP1 at the glycolytic genes TDH3 and PDC1 and TYE7 — also a glycolytic activator — at PGK1; FHL1 at the ribosomal protein gene RPL26A; and a GAL-regulon factor at both galactose genes, GAL80 at GAL1 and GAL4 at GAL3. At GAL1 the effect is unusually clean: the six largest sites in the window are all GAL4 or GAL80, the activator and the repressor of that regulon. Nothing in this pipeline knows which factor is supposed to matter — the sites come from a database and the ranking comes from the model.

It is a tendency and not a law, and the table is ranked by magnitude so you can check it: at HOP2 and ACT1 the winner is STE12, which is a mating and filamentation factor rather than the regulator either gene is known for, and at KRE33 and DTD1 the winners are not characterised regulators of those genes at all. Sign is not reliable either — knocking out the GAL80 repressor site raises the prediction, which is the right direction for a repressor, but several activator sites also come out positive

A site whose bar is inside its own spread did nothing measurable, and that is a result about that site rather than a broken button. Shuffling preserves base composition and destroys only the arrangement, so this asks whether the order of those bases matters — a different question from the occlusion map, which zeroes a stretch and asks whether it carries information at all. The two can disagree. A low-complexity site cannot be shuffled at all — one site across the 23 windows is five Cs and an A, which has almost no distinct permutations — so those are labelled rather than reported as a null result. “Unshuffleable” and “the model ignores it” are different findings that produce the same number.

What all twenty-three windows say

Question

Do the single-locus results hold across the whole set?

Why this matters

The page's strongest claims — that evidence tier predicts attribution, that introns are enriched — would be anecdotes if they rested on whichever window happens to be on screen.

Method

Every enrichment and every knockout re-measured on all 23 windows and reported as medians, plus attribution aligned on the direction of transcription across 198 genes.

Intuition

A median across every window is robust to one locus behaving oddly, and reporting the spread beside it shows whether a claim is a tendency or something that holds everywhere.

Cost

Free. Derived from the shipped packs.

What would refute it

A median over 23 can contradict any single window, and does: the 3.26× ChIP-supported enrichment quoted from one locus is 1.73× across the set.

What the answer tells us The medians here are the page's cross-locus claims; any single-locus figure beside them is an illustration. Where the two disagree, the median is the claim and the locus is the anecdote.

Every other panel in this act shows one window at a time. This one is the pattern across all twenty-three, computed once from the same packs and the same null those panels use — so it cannot drift from them.

Each row is one annotation class; the bar is the median across the windows that contain it, and the dots are the individual windows. A class present in two windows and a class present in 23 are drawn the same width, so the count beside each row is what says how much to trust it.

This layer, in detail

Question

What is one stage actually computing?

Why this matters

A stage summary — a norm, a mean — cannot tell a layer where one channel does everything from one where the work is spread over hundreds. Those are different mechanisms with the same summary.

Method

Every channel of the selected stage as one row, on a diverging scale centred at zero, normalised by percentile rather than by min–max.

Intuition

A residual stream is signed, so 'idle' and 'strongly negative' are different states that a one-sided ramp draws identically — reading sign as activity once made 90% of cells look busy on most stages. Centring at zero is what lets an idle neuron leave its cell empty, and percentile bounds keep a handful of outliers from setting the range for everything else.

Cost

Free. Precomputed per locus.

What would refute it

A uniformly inked raster encodes sign rather than activity — which the p1–p99 ramp did, putting 90–96% of cells above the ink floor on 15 of the 20 stages.

What the answer tells us A stage where a few channels carry the signal is one a sparse decomposition might improve on; a stage where it is spread evenly is one where per-channel readings mean little. The raster says which kind each stage is.

Trace a region — layer by layer, back to the sequence

Question

For the region you traced: which stages, which neurons, and back to which bases?

Why this matters

Knowing that a region matters does not say how. Without a path from the region back through the stages to the sequence, an internal panel is a picture of activity rather than an account of the computation.

Method

Per-channel and per-position relevance margins, both exact and both superposing, with the interior reconstructed as their outer product under independence.

Intuition

Both margins are exact and superpose, because gradients are linear in which outputs you select — so an arbitrary region needs no new model run. Only the interior is reconstructed, and it says so.

Cost

Free for any region — gradients are linear in which outputs you select, so an arbitrary contiguous region is exact with no model run.

What would refute it

The earlier estimate weighted the channel margin by activation and is essentially uncorrelated with the exact answer (r −0.08 to 0.23, argmax positions disagreeing entirely). The reconstruction is checked against its own margins rather than assumed.

What the answer tells us The exact margins say which stages and channels carry a region; the reconstructed interior says roughly where within them. Reading the interior as a measured neuron-by-position map would claim more than the computation supports.

Five questions, in order. Each one narrows the last: which stages carry this region, then which neurons inside the strongest of them, where those neurons fire, what they respond to in the curated annotation, and finally what the transformer could have read regardless of what moved the prediction.

Which attention layer reads what?

Question

Do the transformer's attention layers specialise on kinds of sequence?

Why this matters

Attention maps are the most over-read object in interpretability. Establishing what this one can and cannot support is worth more than another visualisation of it.

Method

Each layer's attention mass — its four heads averaged, which is how the pack ships — against annotation classes, with the mask pooled by MEAN (a 7 bp site is 5% of a 128 bp cell) and a circular-shift null.

Intuition

Attention is a distribution over keys, so a row's mass on annotated positions asks what fraction of what a layer reads is annotated — a quantity a merely loud layer cannot inflate.

Cost

Free. The 8 × 128 × 128 maps ship in every pack — no forward passes at all.

What would refute it

Every class records its ceiling, 1/coverage. Genes cover 91.5% of a window so nothing can exceed 1.09× on them, and reading their 1.03 as flat would be wrong.

What the answer tells us Enrichment says where a layer's attention mass falls under one null. It does not say the layer uses what it attends to, and with the four heads averaged before export it cannot speak about specialisation within a layer at all.

The transformer at the bottleneck is eight layers of four heads over 128 positions, and every locus pack already ships their maps — so this asks a mechanistic question with no forward passes at all. For each layer, how much of its attention lands on each class of curated annotation, against a null that rotates the annotation and destroys only its alignment.

Layers, not heads, and the distinction is forced by the pack. The export averages the four heads within each layer before writing the map — attention.mean(dim=2) in build_onnx.py — so what ships is eight layer matrices, not thirty-two head matrices. Every number below is therefore a statement about a layer with its heads already collapsed. Per-head specialisation is a different and finer question, and answering it would need a pack this page does not have. Not a live run either: the export averages the heads before the graph is written, and the browser consumes that collapsed output, so recovering the four heads needs a newly instrumented pass against the raw checkpoint — four times the attention payload. This panel does not answer it, and an earlier version of this text claimed it did.

The layers place their attention mass the way a biologist would divide the genome. Layers 0 and 3 are enriched on regulatory DNA against the shifted null — conserved binding sites at 1.75× and 1.88×, ORegAnno regions at 1.82× and 1.72×, ChIP-supported sites at 1.65× and 1.52× — and depleted on coding sequence at 0.73 and 0.69. Layer 4 is the mirror image: CDS 1.10× and every regulatory class below 0.90. Layer 6 alone is enriched on tRNA genes, at 1.50×. Nothing in training named any of these categories. Enrichment is where attention mass lands, not evidence that the layer uses it: this is a descriptive alignment under one null, and it is unsigned token mixing rather than attribution. And because the heads are averaged, a layer reading “flat” here could still hold one specialised head and three that cancel it.

Is the model's own basis the right one? sparse dictionary · 6,144 features

Question

The layer panel draws one row per channel and the traceback ranks channels by relevance. Both assume a channel means something. Does it?

Why this matters

Two panels above rank channels and draw one row per channel. Both assume a channel is a meaningful unit; if the model's features are spread across channels, those panels are reading an arbitrary basis.

Method

A TopK sparse autoencoder on the attn_out8 residual stream: 384 dimensions expanded to 6,144 features with exactly k = 32 active, decoder columns unit-norm, trained on 94,970 bottleneck vectors swept from the genome.

Intuition

If a sparse overcomplete dictionary reconstructs the same activations far better than the identical architecture trained on column-shuffled activations, the structure it found was in the co-activation and not in the marginals.

Cost

One forward pass a genome window to collect, then 4,000 Adam steps. Minutes, not hours — the bottleneck is only 384 wide.

What would refute it

Two controls, both run. Training on column-shuffled activations keeps every channel's marginal distribution and destroys only the co-activation structure: it reconstructs at FVU 0.373 against 0.019. And the 384 raw channels are scored against the annotation on identical terms — if they grounded as well, the dictionary bought nothing. They reach 4.715x against the features' 5.977x.

What the answer tells us If the dictionary beats a composition-matched control, the structure is in the co-activation rather than in the model's own channel basis — and the two panels above, which rank and draw channels, are reading a basis rather than a set of features.

A channel is a direction the optimiser happened to land on. In a network carrying more concepts than it has channels, the concepts sit in superposition — spread across non-orthogonal directions, so one channel fires for a binding site and a splice donor with nothing connecting them. A sparse dictionary looks for a better basis: overcomplete, so there is room for one direction per concept, and hard-sparse, so each direction is under pressure to mean one thing.

The reconstruction gap is the first finding. With 32 of 6,144 features active at a time the dictionary reconstructs the bottleneck to within 1.9% of its variance. The same architecture, sparsity and step count on column-shuffled activations reaches only 37.3% — a gap of 35.4 points. Shuffling each channel independently keeps every channel's own distribution and destroys only which channels fire together, so that gap is exactly the co-activation structure the model's own basis was spreading out.

Does a feature mean anything?

Reconstructing well is necessary and not sufficient. Every feature is scored against the genome-wide annotation with the same circular-shift null the enrichment table uses, and the 384 raw channels are put through the identical test — same number of candidates, same null — because if the basis the dictionary replaced grounds just as well, the dictionary bought nothing.

The dictionary grounds better than the basis it replaced. Matched at 384 candidates each and the same null, the sparse features reach a median best enrichment of 5.977× against the raw channels' 4.715×, and a maximum of 204.042× against 13.642×. The top features spread over 19 annotation classes rather than piling onto one.

Read the classes before the ratios. The largest numbers land on tRNA, LTR and snRNA genes — classes with distinctive base composition, which a feature can find without learning anything about regulation, and the 6-mer column shows exactly that: the strongest tRNA features are GC-rich where the genome is not. The harder claim is the quieter one: 5 of the interpreted features ground best on a binding-site tier, the strongest at 9.7× on ChIP-supported sites. That is a real result at a much lower ratio, and it is the one worth trusting least casually.

At what depth is the answer already decided? causal tracing · 22 windows

Question

Where in the network — at what depth, and at which positions — does the promoter's information become sufficient to reconstruct the prediction?

Why this matters

Attribution says which bases matter, not where in the network their information becomes sufficient. In a skip-connected model 'the bottleneck decides it' and 'the skips carry it' predict different things.

Method

Causal tracing (Meng et al. 2022). Corrupt the promoter by dinucleotide shuffle, then restore the CLEAN activations at one stage and one 512 bp band, and measure how much of the clean prediction returns.

Intuition

Restoring a whole stage is degenerate — everything downstream simply inherits the clean run. Restoring one spatial band at one stage is what makes depth and position separable at all.

Cost

19 stages × 32 bands × 22 windows — 608 forward passes a locus, 346 s for the set.

What would refute it

Three controls, asserted per locus. Restoring every position of the first stage must recover exactly 1.0, restoring every position of the last stage must too by an independent route, and restoring nothing must recover exactly 0.0. Together they catch a patch written into the channel axis, which broadcasts cleanly and produces entirely plausible numbers.

What the answer tells us Where recovery concentrates is where the promoter's information becomes sufficient. A bottleneck that does not fully recover is not a defect but the skips measured, and it bounds how much of this model any bottleneck-only story can explain.

Every attribution above answers which bases mattered. None of them answers where in the network that mattering became decisive — and those are different questions, because a base can matter enormously while the representation carrying it only forms several layers in.

Why 22 windows and not 23. Recovery is a fraction of the gap the corruption opened, so a window whose gap is ~0 has no denominator. HOP2's promoter shuffle moves the prediction by less than 0.001 — it is meiosis-specific and silent in this baseline, so scrambling its promoter changes essentially nothing — and the generator skips it rather than dividing by noise and reporting a recovery of forty. An exclusion, not a missing measurement.

Recovery rises through the encoder and decays through the transformer. Averaged over the 22 windows it climbs from 0.434 at the stem to 0.546 at the last residual block, is essentially tied there with the first attention layer (0.543), and then falls monotonically across the remaining seven to 0.349. Per window the peak cell sits in the encoder in 18 of 22, at the bottleneck in 2 and in the decoder in 2 — and it is nearly total, a median of 96.5%.

So by the time a promoter reaches the last residual block, one 512 bp band of its representation is already almost the whole answer — and every attention layer after that makes a single band less sufficient, not more. Restoring one band at the bottleneck recovers a median of only 3.9%. That is what self-attention is for: the information is not lost, it is no longer in one place. Read the two claims together — the encoder localises and the transformer distributes — rather than either alone, and note that the encoder-versus-bottleneck margin at the peak is 0.003, which is a tie and is drawn as one.

The three controls, and what each one pins. Restoring every position of the first stage makes everything downstream the clean run's, so it must recover exactly 1.0 — and does, at every window. Restoring every position of the last stage must too, by a completely different route: the head is a deterministic function of it. Restoring nothing must recover exactly 0.0. Together they fix both ends of the scale and catch the one failure that otherwise passes silently — a patch written into the channel axis rather than the position axis, which broadcasts cleanly, completes without error, and produces entirely plausible numbers.

Is the motif enough on its own?

Question

Is a motif enough on its own to move the prediction?

Why this matters

The knockout sweep can only report necessity in the context a motif already sits in. A site can be necessary because of what surrounds it and carry nothing on its own.

Method

Implant the consensus into 200 dinucleotide-shuffled backgrounds and compare with its own scramble in the SAME background — paired, so the error bar is the sd of per-background differences rather than sd/√n.

Intuition

Necessity and sufficiency can come apart, and a neutral background isolates the second: whatever the motif does there, it does without help. The backgrounds preserve dinucleotides because yeast promoters are heavily biased toward poly(dA:dT) — a mononucleotide shuffle would destroy that too, turning a composition control into a composition change.

Cost

11,600 forward passes, minutes.

What would refute it

CACGTG is its own reverse complement, so the forward and reverse arms must score identically and the script raises if they do not. Poly(dA:dT) has no scramble at all and is reported as degenerate rather than as a measurement.

What the answer tells us Sufficiency in a neutral background and necessity in context are different properties, and a motif can have either without the other. What this arm alone establishes is what a motif carries when nothing else is present.

Each consensus from the paper's dictionary implanted at the centre of 200 dinucleotide-shuffled backgrounds, paired against the same background without it (Global Importance Analysis, Koo & Ploenzke 2021). The tick on each bar is that motif's own scramble — the same bases in a shuffled order, so same composition and no motif. A bar that does not clear its own tick is measuring composition.

Two controls make this a measurement rather than a ranking. CACGTG is its own reverse complement, so implanting it forward and backward must score identically — the generator raises if the two arms differ, and it is the only check that catches a mis-wired implantation path, because a wrong one still produces entirely plausible numbers. And poly(dA:dT) has no scramble: every permutation of a homopolymer is itself, so its composition control does not exist and it is reported as degenerate rather than as a measured zero.

The model recovers known regulatory logic from sequence alone. Among the transcription-factor motifs the three strongest activators are Cbf1, Phd1, Tye7p — and all three are E-box motifs containing CACGTG. The three strongest repressors are Ume6.2, PAC motif (Dot6), Dot6p: Ume6's URS1 site, and Dot6 twice over, since the PAC motif is what Dot6 binds — together the ribosome-biogenesis repressor module. Nothing told the model any of that.

The two splice signals are held out of that ranking, and they split. A branch point implanted alone is the most repressing element in the whole set (-0.13277, z -13.834) while a 5′ splice site raises predicted coverage (0.03504, z 4.209). Neither is a regulatory motif and the quantity being predicted is RNA-seq coverage, so they are reported separately rather than ranked against Rap1 and Ume6 — but the split is real, and it is the direction a spliceable intron would push each of them.

Necessary and sufficient are different claims

Question

Are the motifs the knockout sweep calls necessary the same ones implantation calls sufficient?

Why this matters

Necessity and sufficiency are routinely collapsed into one word, 'important', and the motifs where they part are the informative ones.

Method

The two-by-two of in-context necessity against in-background sufficiency, from the two arms measured above.

Intuition

Crossing the two arms puts every motif in one of four cells. A general regulatory factor should come out necessary and NOT sufficient — it works through the promoter around it — which makes that expectation testable rather than assertable.

Cost

Free. Both arms are already run.

What would refute it

The answer is largely no, and that IS the finding: Rap1 and Reb1 are necessary and not sufficient at z ≈ 0.2, and Ume6 is the mirror.

What the answer tells us The off-diagonal cells are the informative ones. A factor that is necessary and not sufficient is working through the promoter around it — which is what a general regulatory factor should look like, and what a single importance score would have hidden.

Crossing the knockout sweep in act 5 against the implantation above. Horizontal is necessity — what the prediction loses when a curated site of that factor is shuffled in its own promoter, over every swept site. Vertical is sufficiency — its margin over its own scramble in an empty background.

Reb1 and Rap1 are necessary and not sufficient, and they are the two most-swept factors on the page — 14 and 23 curated sites, among the largest knockout effects anywhere in act 5 — yet implanted into a neutral background neither is distinguishable from its own scramble. Both are yeast's general regulatory factors, and the model has them working through the promoter around them rather than alone. Ume6 is the mirror image: strongly sufficient, barely necessary in these 23 windows, which are mostly constitutive genes where a meiotic repressor has little to do.

Where does a motif work?

Question

Does a motif work anywhere, or only in a particular place?

Why this matters

A motif's effect is usually quoted as a single number, which presumes position does not matter. For a promoter element that presumption is precisely what is in question.

Method

Implant at every position across the window and subtract a SCRAMBLE implanted at the same position.

Intuition

Overwriting real sequence destroys whatever was there, and in a promoter that is exactly where the real sites are — so the artefact would look like the signal. Subtracting a scramble implanted at the same position cancels it.

Cost

47,104 forward passes, 21 minutes.

What would refute it

Overwriting 8 bp of a real promoter destroys what was there — in a promoter, exactly where the real sites are. Without the scramble at every position that artefact would look like the signal.

What the answer tells us A motif whose effect depends on where it sits is a positional element and not merely a sequence one; a flat curve says the opposite. Both are answers, and the same-position scramble subtraction is what makes either trustworthy.

Panel 1 asked whether a motif does anything at all, dropped into empty DNA. This asks the question a promoter actually poses: the same motif walked across a real window 64 bp at a time, scored on that window's own gene, and plotted against distance to the transcription start site.

The motifs that were sufficient are the motifs with a position. Cbf1 reaches +0.09398 log₂ in the 500 bp before the TSS and -0.00188 inside the gene body — it works upstream and nowhere else, which is the same asymmetry the attribution metaprofile found by a completely different route (1.40× a gene's mean base upstream against 0.94× inside). Ume6 represses from both sides, more strongly inside. And Rap1 and TATA box — the two that panel 2 showed are necessary at their own promoters but not sufficient in empty DNA — are flat everywhere here too. Three methods, one answer.

And every curve goes flat beyond about ±2 kb, which is not something this measurement was set up to show. It is where act 2's receptive-field sweep independently put the median window's effective context — ±2,048 bp of the ±8,192 available. Implanting a strong activator outside that radius moves the prediction by nothing, because outside it there is nothing left to move.

Does the arrangement matter?

Question

Does the arrangement of two motifs matter, and is there helical phasing?

Why this matters

Grammar claims — that two sites cooperate at a helical spacing — are among the most attractive and least tested in regulatory genomics, and a periodogram will always return periods.

Method

Walk a motif pair 4–200 bp apart in four orientations, re-measuring the moving motif's SOLO effect at every separation so its positional dependence is not folded into the interaction.

Intuition

Helical phasing predicts something specific: cooperation recurring near every 10.5 bp, the turn of the double helix, because that is when two sites face the same way. A periodogram always returns periods, so the discriminating test is the explicit in-phase against anti-phase contrast — and any period near a simple fraction of the scan window is the ruler rather than the DNA.

Cost

25,448 forward passes.

What would refute it

The explicit in-phase against anti-phase contrast is the primary reading, because a 61-point periodogram cannot separate 10.5 bp from its own sixth harmonic (n/6 = 10.17 bp). Measured ratio 1.094 — no helical grammar.

What the answer tells us A helical grammar would show as a periodicity surviving the solo-effect subtraction. Its absence, with short-range unphased interaction present, says this model has learned proximity rather than phasing.

One motif anchored at the centre, a second walked away from it one base at a time in all four orientations. Plotted is the interaction — the pair's effect minus each motif's own effect at its own position — so a flat line means the model simply adds the two up. Grey rules mark multiples of 10.5 bp, one turn of the double helix.

There is no helical grammar, and the constructive test agrees with the Hessian below. Across the six pairs the in-phase / anti-phase ratio has a median of 1.0938 and runs from 0.7423 to 1.3676 — both sides of 1.0, which is what noise looks like rather than a suppressed signal. Orientation barely matters either. What the sweep does find is short-range and unphased: the strongest interaction anywhere is Cbf1 × Cbf1 at 6 bp (0.20858 log₂), and Reb1 reaches TATA best at 31 bp.

Which bases interact with a binding site

Question

Which bases interact non-linearly with a binding site?

Why this matters

Everything else on this page is first order, scoring each base on its own. A first-order method cannot express a pair of bases that only matter together.

Method

A Hessian-vector product, H·v = ∇⟨∇f, v⟩ — one extra backward pass per site, mapping every base that interacts with that motif.

Intuition

The second derivative at the reference is the cheapest object that can carry an interaction — one extra backward pass. Its symmetry is the only available check that the double-backward is wired correctly, since a mis-wired one has the right shape and magnitude.

Cost

425 ms a site — 431 sites across 23 windows, 18.7 a locus on average.

What would refute it

H is symmetric, so ⟨H·v_A, v_B⟩ must equal ⟨H·v_B, v_A⟩; the worst residual over 431 sites is 4.2e-04. Nothing else catches a mis-wired double-backward — it has the right shape and magnitude and is simply the wrong quantity.

What the answer tells us A base that interacts with a site is one no first-order map can score correctly, which is the gap this panel exists to fill. Whether these particular second-order values predict real double edits is a separate question, and act 9 measures it.

Regulation is not additive — factors cooperate, compete, and care about spacing — and a first-order method cannot see any of that by construction. For a motif spanning [a, b) with indicator v, H·v = ∇⟨∇f, v⟩ is one vector over all 16,384 bases whose entry i says how much that motif's own sensitivity changes when base i changes. One extra backward pass — 425 ms — rather than the 16,384² Hessian.

A second derivative is not a double substitution, and act 9 now measures the gap. Every pair of positions in a locked panel was mutated together — 39,330 exact double mutants — and the residual left after subtracting both single effects is what this panel is trying to predict. The Hessian does not call it: median r = -0.0843 across 23 loci, against 0.3227 for separation alone. Read this panel as the local curvature map it computes, not as a prediction about what two edits do together. The calibration is in act 9.

The helical-periodicity prediction is refuted, and that is the result. If the model had learned B-DNA's 10.5 bp pitch, |H·v| would repeat with it. Pooled over all 431 ChIP-supported sites in the 23 windows it does not: dividing out the decay and comparing distances in phase with 10.5 bp against anti-phase gives a ratio of 0.948358 — if anything slightly lower in phase. A falsifiable prediction, tested, and answered no.

What would the model build?

Question

What would the model build, given a promoter to improve?

Why this matters

Reading what a model finds important is a weaker test than asking it to construct. A model can attribute sensibly and still hold no usable notion of what makes a promoter strong.

Method

Greedy ascent on the prediction, one substitution at a time, with the motifs it invents matched against the database afterwards.

Intuition

Greedy ascent verified by real forward passes turns the question into a construction. The control decides what it means: the same ascent on shuffled DNA says whether the motifs it builds are about yeast or about the search.

Cost

5.5 minutes — 12 greedy substitutions a window chosen from 20 candidates, and the same ascent again on shuffled DNA as the control.

What would refute it

The same ascent on dinucleotide-shuffled DNA builds 54 motifs in 23 of 23 windows against the real arm's 18 in 11 — the control wins 3.0× raw, and the panel says so. What survives is that the achievable gain is set by headroom, r = −0.873 against starting expression.

What the answer tells us What the model builds is a statement about what it takes to make a promoter strong. The control decides whether that statement is about yeast: if shuffled DNA yields as many motifs, the ascent is finding the structure of the search rather than of the genome.

The generative question, and the only one here that lets the model choose the sequence. Starting from each real window, the 12 single-base edits that most raise its own gene's predicted expression — proposed by the gradient, then verified by real forward passes, because a gradient is a local slope on a saturating function and this page has already measured where that fails.

What it finds is headroom, not promoter quality. The gain runs against starting expression at r = -0.8726: the five quietest genes gain 2.7958 log₂ from twelve bases while the five loudest gain 0.3906. HOP2 and GAL3, both repressed in this baseline, move most; TDH3 and PDC1, already near the top of the model's range, barely move at all. A promoter the model thinks is maximal cannot be improved by editing it.

And the interesting claim does not survive its control. It is tempting to report that the ascent invents recognisable regulatory motifs — it does, 18 of them across 11 windows. But running the identical ascent on dinucleotide-shuffled DNA of the same composition builds 54 in 23 of 23. Raw, the control wins 3×; normalised by how much expression each arm actually gained, the two are within 25% (0.6913 against 0.5498 motifs per log₂). Either way, building a Reb1 site is a property of the ascent and the base composition — not evidence about promoters.

What does the model think a motif is? discovered, not matched

Question

Left to itself, does the model's mutagenesis contain repeated sequence patterns — and are they the motifs a database would name?

Why this matters

Every motif on this page so far was supplied — from a database or from the paper's figure. Whether the model's own attributions contain repeated patterns is a different and harder question.

Method

TF-MoDISco in the small (Shrikumar et al. 2018): pull high-|saliency| windows out of the mutagenesis planes at each pre-registered width (11 and 15 bp), cluster them on their mean-centred contribution blocks over both strands, build each cluster's PWM from the bases its seqlets contain, and only THEN match against JASPAR.

Intuition

Any clustering returns clusters and a lower threshold returns more, so the threshold has to come from the control before either arm is clustered — and the control has to perturb the model's input, not only the projection.

Cost

Free. No model and no GPU — the raw mutagenesis planes are already on disk.

What would refute it

The identical pipeline on dinucleotide-shuffled sequence, and the shuffled arm sets the clustering threshold. It found almost as much: the control is close enough that the headline has to be stated carefully, and it is.

What the answer tells us Repeated patterns in the model's own attributions would be motifs it discovered rather than motifs it was handed. Whether they clear a null that recomputes the model on shuffled input is settled below in this same panel, and until they do the clusters describe the clustering rather than the sequence.

Every other panel on this page starts from an annotation and asks whether the model agrees with it. That is the right way round for validation and the wrong way round for discovery — it can only ever find motifs already in a database, and it cannot say what the model thinks a motif is.

Both seqlet widths were declared before the run, and both are here. At 11 bp the real planes give 3 patterns and the shuffled control 2; at 15 bp the real planes give 3 patterns and the shuffled control 3. At the wider width the control ties exactly — three against three, all matched. Running a second width after a weak first result is only defensible because the grid was fixed in advance and neither cell is dropped; no third width was added afterwards.

Read the control before the table. Of 1,047 seqlets pulled from the real planes, 41 fall into 3 groups large enough to be called a pattern — the rest are singletons or pairs below the size threshold, which is what a strict similarity cut is for. All 3 of the patterns match a JASPAR matrix. The same pipeline on dinucleotide-shuffled sequence yields 2 patterns, and 2 of those match too. Three against two is not a result. What does separate them is sharpness — the best real pattern carries 7.00 bits against the control's 3.91 — but every pattern in both arms is AT-rich, and a dinucleotide shuffle preserves AT-richness by construction.

So the honest statement is narrow, and the second width narrows it further. At a threshold set by the control arm's own 99.9th percentile, what the model's highest-attribution 11-mers have in common is base composition rather than motif syntax — and at 15 bp, where the control matches the real arm pattern for pattern, there is no separation left to interpret at all. That is a real finding about where a fixed-width, single-pass clustering of 1,000 seqlets can and cannot reach — not evidence that the model has no motif vocabulary, which the knockout sweep and the sufficiency panel both show it does.

Tell the model it is a different fungus

Question

Does the species channel carry anything, with the DNA held byte for byte?

Why this matters

The model takes a 165-way species one-hot the paper never explains. Whether it encodes anything, or is a nuisance input the network learned to ignore, changes how every other result here should be read.

Method

Sweep all 165 species one-hots on a fixed S. cerevisiae promoter and rank the prediction.

Intuition

Holding the DNA byte for byte and changing only the label isolates the channel. The control that decides whether it means anything is between-locus rank correlation — a per-species bias would produce the same ordering everywhere.

Cost

165 forward passes a locus, 1.6 min for the set.

What would refute it

The between-locus rank correlation is the control: median 0.33, so the orderings genuinely differ and this is not a per-species bias. The tidy story that the exceptions are the repressed genes is NOT supported — rank against log baseline coverage is only ρ = −0.34.

What the answer tells us A channel that reorders predictions carries information; one that only scales them is a gain. The between-locus control separates the two, and only the first would license reading the channel as phylogeny.

Shorkie's input carries a 165-dimensional one-hot naming which Saccharomycetales genome a sequence came from — an architecture no other sequence-to-function model has, and a counterfactual available here and nowhere else. The DNA below never changes; only the declared species does.

Told the truth about what it is reading, the model predicts the most expression. S. cerevisiae ranks first of 165 at 19 of 23 windows — probability 6.37e-39 under a null where the ordering is unrelated to the sequence — and sits above the median species at 21 of 23, by a median of 0.912 log₂ units. The exceptions are GAL1, GAL3, HOP2, GLK1, and all of them top out on Yarrowia lipolytica, the most distant lineage in the set.

The tidy explanation of those exceptions does not survive checking. Three of the four — GAL1, GAL3 and HOP2 — are the genes repressed in this baseline, which invites the story that the species channel only helps where a gene is being transcribed. Across all 23 windows that story is not supported: rank against log baseline coverage is only ρ = −0.34, a weak tendency rather than a rule, and the fourth exception breaks it outright. POP4 is among the quietest windows in the set and ranks first; GLK1 is far louder and ranks 162nd of 165. The pattern is suggestive and it is reported as suggestive.

Do the drivers change as the response unfolds?

Question

Do the sequence drivers change as an induction response unfolds?

Why this matters

The model predicts a time course, and it is natural to read that axis as mechanism. Whether the sequence drivers actually move along it has to be measured rather than assumed.

Method

Attribution at early and late timepoints, per REGULATOR rather than on the timepoint means.

Intuition

Timepoint means average over hundreds of regulators and are near-identical by construction, so a per-regulator contrast is the only level at which a real change could show.

Cost

Free. Precomputed per locus.

What would refute it

On the timepoint means early-against-late is r ≥ 0.9995 with 99% of the top 500 bases shared — no effect at all. Per regulator it is real, MSN2 at GAL1 reading 0.215 with 50.2% shared.

What the answer tells us A driver that moves between timepoints is a sequence element whose relevance is condition-dependent. On the timepoint means the question is unanswerable because those means are near-identical, so the per-regulator level is where any real dynamics has to appear.

The 3,053 induction tracks resolve into 335 regulators over 13 timepoints — a sparse grid, 3,053 of 4,355 possible combinations, so a timepoint does not carry every regulator. Taking one regulator's own tracks early and late and attributing each over this gene body asks whether the model reads a different part of the sequence as the response develops.

It does, and only where the pairing is real. Across 26 regulators × 23 windows the median early-to-late correlation is 0.9973 — but the floor is 0.2149, at MSN2 on GAL1, where only half the strongest 500 bases survive. MSN2 and MSN4 are the paralogous general-stress factors that bind the same STRE motif, and they rank one and two independently; GAL1 and GLK1, where they move most, are both glucose-repressed.

Does selection agree with the model?

Question

Does natural selection avoid the bases the model says matter?

Why this matters

Every result above is about the model. Whether any of it corresponds to selection in real yeast populations needs evidence from outside it.

Method

Paired at the base: the observed allele against the two alternates nature did not choose at the SAME position, so position, context, gene and expression are all held fixed and no null model is needed.

Intuition

If the model has learned what constrains a sequence, the bases where it predicts a large effect should be bases where surviving variation is milder, because selection has already removed the costly alternatives. The prediction is directional — the observed allele is the gentler one — so a sign test is the right instrument, and the skew in the per-site ratios is why a mean is the wrong one.

Cost

Free. No model run — the 2,319 UCSC variants that fall inside the windows mutagenesis covers, scored against the shipped planes.

What would refute it

A sign test, because the ratios are skewed: for missense the ratio of means is 0.99 while the median per-site ratio is 0.90. The reference base matched sacCer3 at 2,319 of 2,319, which is what makes the coordinates trustworthy.

What the answer tells us This is the only evidence on the page from outside the model. It is an alignment with selection, not a demonstration that the model has captured the mechanism selection acts on.

2,319 real segregating variants from UCSC's evaSnp8 fall inside the windows mutagenesis covers — and those packs hold all three substitutions at every base. So for each variant the allele that actually segregates can be compared with the two alternates at the same base that nature did not choose. Paired, so position, context, gene and expression level are held fixed by construction and no null model is needed.

They agree, modestly and consistently. The observed allele is the milder one 55.8% of the time for missense (sign test z = 2.54), 55.4% for synonymous (z = 3.3) and 54.6% for non-coding (z = 2.74), with a median per-site ratio near 0.90 in all three. The model has never been shown a variant; this is purifying selection on expression, visible through a network trained only to predict assays.

Does the signal survive retraining? 85% of strong bases

Question

Would a different training run of the same architecture call the same bases important?

Why this matters

Every headline on this page comes from one checkpoint. If a different training run of the same architecture ranks different bases, the page is reporting a property of one optimisation rather than of the architecture or the data.

Method

Every exact single-base effect in a locked panel per locus — the 128 strongest by mutagenesis plus 128 matched on distance to TSS — recomputed under all 8 released folds and on both strands. Fold and strand are kept as separate axes and never pooled.

Intuition

Eight released folds are eight draws from one training procedure, so their disagreement estimates training variance directly. The panel must be selected on one fold and scored on the others, or the selection is being reported back to itself.

Cost

17,664 substitutions × 8 checkpoints × 2 strands, about an hour on an M1 Pro, plus 385 MB of checkpoints that are public and were simply never downloaded.

What would refute it

The panel is chosen from f0, so f0's own agreement with it is a tautology: f0 selects and is reported separately, and every cross-fold statistic below is computed over f1, f2, f3, f4, f5, f6, f7 only. The matched controls are what "how many survive" is measured against.

What the answer tells us A base that keeps its sign across eight independent training runs is a property of the architecture and the data; one that does not is a property of a single optimisation. The matched controls set the scale — 54% survive by construction, so 85% is what selecting on mutagenesis actually bought. The pairwise overlap adds the caveat that the headline cannot: direction is stable, ranking much less so.

The eight folds are eight different models, and this was measurable all along. Both audit documents record training-fold uncertainty as unestimated. It is not unestimable — all eight checkpoints sit at a public URL, and on one locus they predict g across a range of a factor of 1.68 in coverage. What follows is what happens to the page's own top-ranked bases when the training run changes.

All eight checkpoints, one row each

The summary above is a median over these. Each row is one released training run and each column one locus, coloured by how far that fold's own prediction sits from the across-fold mean for that locus. Deviation rather than the raw score because the loci span 5.8 to 14.9 log₂ units — on an absolute scale every column would be one flat colour and the differences between checkpoints, which are the whole subject, would be invisible.

How much do two checkpoints agree?

Sign is stable; ranking is much less so, and the two are easy to conflate. The headline above says 85% of the strongest bases keep their direction across folds. But asking which bases each checkpoint puts in its top 1% is a different question, and the answer is 41–55%. Two checkpoints trained on the same data agree about roughly half of what matters most. A single-fold ranking is therefore a weaker object than a single-fold sign.

Strand and fold disagreement are comparable in size, and neither is noise. The median spread across folds is 0.0085 against a median strand deviation of 0.0080, a ratio of 0.943×. The model was trained with augment_rc: false, so forward and reverse are genuinely different states and averaging them is a test-time choice, not a symmetry being exploited. Pooling the two axes into one error bar would hide both.

Which attribution methods survive an exact edit? 3 of 4

Question

Do the page's attribution maps rank bases the way exact mutagenesis does — and do they beat a baseline that knows nothing about the model?

Why this matters

The site shipped five attribution surfaces and had never scored one against an exact edit. Without that, choosing between them is an aesthetic judgement.

Method

Rank agreement with the exact ISM planes, each method scored at its OWN native resolution with the ground truth pooled to match, plus a deletion curve: rank the bases, substitute the top-k to their worst alternative, and measure the real fall in g.

Intuition

Exhaustive mutagenesis is the finite-difference contrast every other method approximates, so it is the natural reference. Rank agreement alone is not enough: a map can correlate well and still order the wrong bases first, which only an intervention reveals.

Cost

Minutes. The exhaustive mutagenesis ground truth for all twenty-three windows was already on disk; what was missing was the interventions that score against it.

What would refute it

A method that cannot beat GC content, distance to TSS or a random ranking is demoted on this page rather than defended. Attention rollout is included knowing it is unsigned token mixing at 128 bp and has no reason to pass.

What the answer tells us A method that clears the baselines has earned the right to be read as an attribution; one that does not is a picture. This is the verdict act 6's copy now defers to, and it is why attention rollout is described there as token mixing rather than importance — it finds 4.0% of the achievable damage against 15.3% for a random ordering.

Scoring a 64 bp method per base measures its resolution, not its faithfulness. Occlusion and rollout resolve 64 and 128 bp, so the mutagenesis truth is pooled to each method's own grid before the ranks are compared. The deletion curve gives them no such allowance — within a block it can only order bases by position, which is exactly the handicap their resolution imposes on a reader.

The two columns measure different things, and the ceiling row shows it. The oracle is ordered by how far g falls when each base is substituted to its worst alternative — the quantity the deletion curve applies — so it bounds that column at 1.000. Its rank agreement is only 0.415, below gradient × input, because the ground truth for that column is the largest effect a base can have in either direction. The two orderings part company wherever a base's biggest effect is an increase. An earlier version ranked the oracle by |effect| and integrated gradients scored 1.0053 against it — a ceiling a method can exceed is a mislabelled competitor, not a ceiling.

1.000 is ranking by the answer. Every curve is divided by what the exact mutagenesis ranking itself achieves, so the column reads as the fraction of achievable damage a method found. The bar is the strongest baseline, at 0.166 — GC content is a real predictor of where a promoter is, and distance to TSS still better, so a method that merely finds promoters clears a random ranking without demonstrating anything about the model. The conv stem is filed with them deliberately: it is linear and unnormalised, so its filters are a basis any invertible recombination leaves unchanged.

attention rollout does not clear the baselines. attention rollout finds 4.0% of the achievable damage, against 15.3% for a random ranking. Attention rollout is unsigned token mixing at 128 bp — an architectural quantity describing what the transformer can read, not an attribution — so this is the expected result rather than a surprise, and the panel in act 6 says so on its face. It is reported here because a method nobody scores is a method nobody can demote.

Every fold is scored against fold f0's exact mutagenesis, not its own. That is the only exhaustive mutagenesis that exists — per-fold planes would be about 52 GPU-hours — so the fold arm asks a specific question: do the page's f0-derived conclusions transfer to another checkpoint? The page's claims are f0-derived, so that is the right question, but it is not "each checkpoint against its own ground truth", and one consequence shows in the numbers: the oracle is a strict ceiling only for f0, and another fold's ranking can edge past it.

The verdicts hold across the eight training folds, which is what makes them verdicts. The whole benchmark was re-run under every released checkpoint: integrated gradients, gradient × input, occlusion clear the baselines in 8 of 8; attention rollout in none. No method is split across folds, so no verdict here rests on which checkpoint was used. The spread across checkpoints differs by method: gradient × input ranges 0.742–0.914 while attention rollout ranges 0.026–0.040, so how much a method's score depends on the training run is itself a property of the method.

Does the map read the model, or the input?

An attribution that survives having the network destroyed is not reading the network. This is the Adebayo sanity check: randomize the weights from the head backwards, stage by stage — permuting each tensor's own values, so every parameter keeps its exact distribution and only the arrangement is lost — and watch the map decorrelate from its intact self. Gradient × input falls from 0.96 with only the head destroyed to 0.25 with the whole network destroyed, over 6 loci.

The branch column is why one number would have misled. Destroying the entire transformer takes the map only from 0.60 to 0.34 — not to zero — because the three decoder skips are fed by block5, block6 and block7 and carry real signal around the bottleneck. A pooled score would have read that architectural fact as the check failing. It does not settle at zero either: gradient × input on a fully randomized network still shares 0.25 with the intact map, which is what the one-hot input geometry contributes before any learning.

Is the IG baseline doing the storytelling? best: dinucleotide shuffle

Question

The site ships one integrated-gradients reference — all four DNA channels zeroed. Is that choice carrying the result?

Why this matters

A path method integrates from a reference, and that reference is a modelling choice usually made for convenience. If the choice carries the result, the map describes the baseline rather than the sequence.

Method

The same 32-step path integral under five legal reference families on the same scalar and the same loci, every one preserving the species one-hot and none assuming channel 4 is a mask. Scored by agreement with exact mutagenesis and by real deletion passes.

Intuition

Every family is scored on the same scalar and the same loci, so none gets an easier target. A reference the sequence model finds impossible is one the expression model was never trained near, which makes plausibility measurable instead of arguable.

Cost

23 loci × 5 families × 2 strands × 32 steps, about ten minutes.

What would refute it

Shorkie_LM's own masked negative log-likelihood scores each reference: a baseline the sequence model finds impossible is one the expression model was never trained near. And strand-wise maps are compared BEFORE averaging — a family that only looks stable once the two arms are averaged has failed, not passed.

What the answer tells us A path attribution inherits its reference, so this decides how the integrated-gradients lane may be read. The shipped default does not win and has the weakest strand agreement in the table, so integrated gradients ships here labelled reference-sensitive rather than as the map.

The shipped baseline does not win — dinucleotide shuffle does. All-zero DNA reaches 0.899 of the achievable damage against 0.955, and its forward and reverse maps agree at only r = 0.36 — the weakest in the table. A family that needs the two arms averaged before it looks stable has failed the transparency requirement, not passed it. Integrated gradients therefore ships labelled reference-sensitive, with the menu above exposed, rather than quietly defaulting to one path.

Across all 8 checkpoints, one thing is robust and one is not. The shipped all-zero default wins in 0 of 8 folds — every shuffle-based family beats it under every checkpoint, so "the default is not the best" is a property of the reference rather than of one training run. Which family is best is not robust: the winner moves between dinucleotide shuffle, mononucleotide shuffle, expected gradients across folds (dinuc, mono, expected, mono, expected, mono, mono, expected), and their medians sit within 0.031 of each other. So this panel recommends moving OFF the all-zero baseline, and does not recommend a specific replacement. The single-fold result named a different winner from the eight-fold median, which is the clearest argument on this page for why the fold axis was worth the compute.

Rank agreement and deletion damage part company here, and the shipped baseline is where they part hardest. All-zero DNA has the best rank correlation with exact mutagenesis in the table — 0.469 — and nearly the worst deletion AUC. Correlating with the ground truth and ordering the bases whose mutation actually hurts are different achievements, which is the same dissociation attention rollout shows one panel up. The most in-distribution reference does worst of all: a real window shares so much with the sequence being explained that the path between them carries little.

Real yeast sequence costs Shorkie_LM 1.7648 bits a base under its own masked schedule, which is the number every reference in the table should be read against. An all-zero window is not merely unusual sequence — zeroing the four DNA channels is literally how this model family masks a position, so that reference asks the expression model to integrate a path from a fully masked window.

Does the second derivative predict a real double edit? no — r = -0.0843

Question

The epistasis panel in act 7 is a Hessian — curvature at the reference. Does it predict what two actual substitutions do together?

Why this matters

Act 7's interaction panel is a second derivative. Whether curvature at a point predicts what two discrete edits do together is an assumption this page had never tested.

Method

Every double substitution over 20 positions a locus — nine per pair — evaluated exactly and rc-averaged, with the inclusion–exclusion residual as the measured interaction. Singles are recomputed in the same run so both terms of the subtraction share a convention.

Intuition

A Hessian is the limit of an infinitesimal perturbation; a substitution is a finite jump to another vertex of the simplex. Recomputing the singles in the same run is what makes the subtraction meaningful, since both terms must share one convention.

Cost

39,330 exact double mutants across 23 loci, about half an hour.

What would refute it

Additive predicts exactly zero and is the null the Hessian has to beat. Separation alone is scored too, in case proximity is all the second derivative is tracking. The recomputed singles are checked against the shipped mutagenesis planes; the worst drift is a free correctness check on the whole path.

What the answer tells us Whether curvature at the reference predicts real double edits decides how act 7's interaction panel may be read. At r = -0.0843 it does not, so that panel is a local curvature map. The interactions are real — a median residual of 0.051× a single-base effect — but which pairs interact has to be measured, not inferred from a second derivative.

A second derivative is not a double substitution. The Hessian is curvature at the reference one-hot; a real double edit is a finite jump to another vertex of the simplex, and nothing guarantees the first predicts the second. Half the positions are the strongest bases and half are matched on distance to TSS, so "how much interaction is there" has something to be compared against.

The Hessian does not call the measured interaction, and two trivial predictors do better. Over 39,330 exact double substitutions the Hessian's median correlation with the measured residual is -0.0843, ranging -0.3223 to 0.6826 and positive at only 8 of 23 loci. Separation alone reaches 0.3227 and is positive at every locus, and the summed magnitude of the two single effects does better still. The pipeline is not what failed: the singles recomputed for this benchmark reproduce the shipped mutagenesis planes to 0, so both terms of the subtraction share a convention exactly.

Re-measured under every checkpoint, the failure is not a property of one training run. The Hessian's correlation with the measured residual is positive in 0 of 8 folds (-0.11, -0.17, -0.17, -0.03, -0.06, -0.08, -0.02, -0.00), and separation alone beats it in 8 of 8. A predictor that a trivial baseline outperforms under most checkpoints is not a predictor of this quantity. The fold arm uses a smaller panel than the headline — 10 positions a locus against 20, so 9,315 pairs a fold rather than 1,710 — because the fold question is whether the calibration's sign survives retraining, which a smaller panel answers at a quarter of the cost. The headline above is the full panel on f0.

A curvature is not an edit, and that is the whole finding. The Hessian is the second derivative at the reference one-hot; a double substitution is a finite jump to two other vertices of the simplex. Interactions here are real but small — a median residual of 0.051× a median single-base effect — and the second-order term at the reference does not point at them. The act 7 panel should be read as a local curvature map, which is what it computes, and not as a prediction about combined edits. An earlier single-locus probe of this benchmark gave r = +0.79 on 90 pairs; the full grid is 39,330, and a correlation from one locus is not the correlation.

This is a calibration, not a spacing scan. The separations here are wherever the strong bases happen to be, so they are not uniform and no periodicity claim can be read off them. The constructive spacing sweep in act 7 is where that question belongs — and it already reports no helical grammar.

Every dimension this page collapses

Nothing here is drawable at its real size. A stage is [channels × positions] with up to 16,384 positions; the output is [896 × 5,215]. Every panel therefore collapses at least one axis, and — this is the part worth knowing — they do not all collapse the same axis the same way. Position is max-pooled for display, because the question there is “did any neuron fire here”, and sum-pooled for relevance, because relevance is additive and a mean would make a coarse stage look quiet for being coarse. Both are right; the difference is invisible unless it is written down.

The last five rows are new, and the second is the one most easily got wrong: a 7 bp binding site covers 5% of a 128 bp cell, so pooling the annotation by max would mark that whole cell as annotated and make every class look identical — producing numbers that mean nothing while looking exactly like numbers that do.

Two rows are worth reading against each other. The layer relevance map is the only place on this page where a number is reconstructed rather than measured, and the neuron traces exist directly beneath it as the un-collapsed answer to the same question. Where they disagree, trust the traces.

Every layer, with its tensor shapes

Paper vs checkpoint