Genome browser

Shorkie_LM constraint + Shorkie expression · sacCer3 / R64 ·12,157,105 bp · 17 sequences← Language model Expression model Lab Paper

—

Tap a chromosome in the top strip to jump to it, pinch to zoom, drag to pan, tap the strip below to jump, and tracks to choose what is drawn. Every score lane keeps a fixed axis at every zoom unless you press autoscale, and prints its own units.The top strip is all 17 chromosomes, drawn strictly to scale — chrM really is 5.6% of chrIV, and nothing is widened to a minimum, because a floor would distort the one channel a bar is read by. It carries no bands: yeast has no Giemsa banding, so this is a position indicator and not an ideogram in the cytogenetic sense. Click anywhere along it to jump — the nearest chromosome answers, so the 8 px chrM is reachable. Drag the ruler to select a region (or shift-drag anywhere); drag a track to pan, scroll to zoom, click the overview strip to jump. Every score lane keeps a fixed axis at every zoom, so any two positions are read against the same ruler — and each lane prints its own range, because bits and a 0–1 posterior are not the same units. autoscale suspends that rule for the visible window; it is off by default and every rescaled lane says so on its own face, because an axis that was rescaled without announcing it is the same trick as a bar chart drawn from a non-zero baseline. Above a lane's own resolution a pixel column is a summary bin: the filled bar is its mean, the pale mark its maximum and the dark tick its minimum, so one determined base inside a wide bin stays visible. Signed lanes — the attribution — grow both ways from a zero rule instead. At the deepest zoom the per-base lanes become a logo, one letter a base; a lane whose data is coarser than that stops at its own floor and the readout names it. The nuclear genome averages 0.197 bits masked against 0.685 unmasked, and chrM 0.457 masked — so most of the track is genuinely low and the tall columns are the exception.

This view, in numbers

How the whole genome was scored

The model reads a 16,384 bp window, so a genome-wide track has to be assembled from overlapping windows — and a base near a window edge has less context than one in the middle, which makes its score a different measurement. Only the central 8,192 bp of each window is kept, with 4,096 bp of flank discarded on each side: 1,502 windows over 17 sequences.

The two passes, and why only one of them is a prediction

A masked language model only predicts where it is masked. The browser can draw the model two ways, and the difference between them is the paper's Figure 2A point:

pass	what it asks	mean IC, nuclear genome
masked	positions are split into 7 disjoint strided sets, each masked in turn, and each position is read back only from the pass that masked it — so the model never sees the base it is scoring	0.197 bits
unmasked	one forward pass with nothing masked, so the model can see the base it is scoring and is largely reading its own input	0.685 bits

Only the masked pass is a prediction. The unmasked pass scores 3.47× higher for that reason alone, not because it knows more — and the browser labels its lane not a prediction rather than leaving the caveat to a paragraph you might not read. It is nonetheless the quantity the paper's Figure 2A logo is built on, which is exactly why it is worth being able to draw beside the other.

They are not two views of one number. Across the 23 primary regions the two correlate at r ≈ 0.62, and per chromosome here at 0.59–0.62 — related, and far from redundant. Worth noticing too: the unmasked pass picks the right base 97.5% of the time while averaging only 0.68 of a possible 2 bits. It is confident about which base, not sharply peaked.

Turn both on in the track panel to see them stacked on the same 0–2 axis.

Does the model agree with evolution? Partly — and the way it does not is the interesting part

The model's constraint is alignment-free: it comes from pretraining on 165 Saccharomycetales genomes and says what the model finds predictable. phastCons is alignment-based — a phylo-HMM posterior over a 7-yeast whole-genome alignment — and says what evolution has held still. Two independent measurements of related things, which is what makes laying them in adjacent lanes a check rather than a second opinion from the same source.

over	bases	Pearson r	Spearman ρ	mean IC	mean phastCons	phastCons ≥ 0.99
whole genome	12,078,291	0.121	0.147	0.199	0.601	—
coding sequence	8,583,108	0.045	0.073	0.219	0.713	40.1%
intergenic	3,495,183	0.148	0.152	0.150	0.327	13.1%

Read the split before the headline. Both measures rank coding sequence above intergenic — IC 0.219 against 0.150, phastCons 0.713 against 0.327 — and that shared regional agreement is most of the overall r = 0.121. Within coding sequence they agree far less (r = 0.045 over 8,583,108 bases).

That is partly a real disagreement and partly a property of phastCons: 40.1% of coding bases sit at ≥ 0.99, median 0.966, so inside a gene it is a near-constant 1 and there is very little variation left for the model's per-base signal to track. A correlation is bounded by the range of both variables, and reporting r = 0.045 without that is a wrong interpretation of a right number.

phastCons has no value where the alignment has none — 0.65% of the genome, concentrated on chrM and at telomeres. Those bases are drawn as a gap, never as zero: a zero would read as "completely unconserved" where the truth is "not aligned", and the lane says what fraction of the view is missing.

Why the flank is 4,096 bp, and how the first answer was wrong

How much flank is needed is measurable: score the same 8,192 bases twice, once centred in a window and once against its edge, and compare information content per base. The answer depends entirely on which region you measure.

distance from edge	quiet chrIV:400k	promoter chrI:76k
0–64 bp	0.0321	0.1049
512–1,024	0.0112	0.0245
1,024–2,048	0.0063	0.0224
2,048–3,072	~0.007	0.0186
3,072–4,096	~0.007	0.0109
4,096–6,144	~0.007	0.0060

On a quiet stretch the effect looks finished inside 1 kb, and a 2,048 bp flank looks like a 2× margin — which is what this track first shipped with. On a gene promoter, where the model is actually resolving motifs and where anyone will look, it is still 0.022 bits at 2 kb and only settles around 4 kb. Measure the error on the case that matters, not on the first one to hand.

The pooling-grid phase matters 20× more than the flank does

The encoder pools to 128 bp, so a window's start position modulo 128 decides which bases share a pooled cell. Scoring the same bases from windows at different phases changes information content by 0.0395 bits on average; from windows at the same phase, by 0.0020. That is an order of magnitude larger than the edge effect the flank was chosen for.

Every window in this track therefore starts on a multiple of 128, so every base in the genome is scored on one grid and any two positions are comparable. The 23 primary regions on the language-model page start at their own coordinates and sit on whatever phase that gives, which is why the two agree in shape (r = 0.95–0.99) rather than to the last decimal. Neither is more correct; they are two phases of the same model.

69,632 bases — 0.57% of the genome, all of them within 4,096 bp of a chromosome end — could not be given a full flank and were scored with the largest one available.

What the pyramid stores, and why a summary bin keeps its minimum and maximum

A summary bin that reports only its average hides exactly what a constraint track exists to show: one strongly determined base inside a 4,096 bp bin disappears into the surrounding noise. Every level above base resolution stores three rows, so the drawing can show the envelope and a spike stays visible at genome zoom. A pyramid that smooths is a pyramid that lies.

level	bp / bin	bins	rows	used at
L0	1	12,157,105	value	under 8 bp a pixel
L1	8	1,519,646	min / max / mean	8 bp a pixel and coarser
L2	16	759,826	min / max / mean	16 bp a pixel and coarser
L3	64	189,964	min / max / mean	64 bp a pixel and coarser
L4	512	23,752	min / max / mean	512 bp a pixel and coarser
L5	4,096	2,977	min / max / mean	4,096 bp a pixel and coarser

5,251 tiles of 65,536 bins across 48 score tracks, as greyscale PNG — the same encoding every other pack in this lab uses, because createImageBitmap decodes it natively and no JavaScript inflate has to ship. A viewport fetches only the handful of tiles it covers, at only the resolution it can draw, and the cache is bounded — and the bound scales with how many score tracks are on, or enabling all three would make every pan evict tiles it is about to need again. Alongside them: the reference sequence one byte a base for the letter view, 6,613 gene models, and the annotation layer.

Byte 0 means "no data", in every track. phastCons has no value where the 7-yeast alignment has none, so values occupy 1–255 and 0 is reserved. Quantised naively that base would become 0, which draws as completely unconserved exactly where the truth is not aligned. It costs 0.4% of the precision and it is the same rule for the model tracks, which need it nowhere — one decode path rather than two.

track	units	axis	what it is
Shorkie_LM · masked	bits	0–2	information content, 2 − H(p), from the K = 7 iterative masked pass. A prediction: every position is read back only from the pass that masked it, so the model never sees the base it is scoring.
Shorkie_LM · unmasked	bits	0–2	information content from one forward pass with nothing masked. NOT a prediction. The model can see the base it is scoring and is largely reading its own input, which is why its information content runs ~3.4× the masked pass's. It is nonetheless the quantity the paper's Figure 2A logo is built on.
phastCons · 7 yeasts	posterior	0–1	posterior probability that a base lies in a conserved element. A DIFFERENT UNIT from the two above — a 0–1 probability, not bits — so it is drawn on its own axis and must never be read against theirs. It is also alignment-based where the model is alignment-free, which is what makes it an independent check.
GC content · 50 bp	fraction	0–1	fraction of A/C/G/T in a centred 50 bp window that are G or C. A CONTROL rather than a finding. If the model's information content were mostly base composition it would show here; measured, r = -0.020 genome-wide.
Shorkie · predicted RNA-seq	a.u.	0–1097.560669	mean predicted coverage over the 384 T0 RNA-seq tracks. The quantity every attribution on this site is scored on, so this lane and the gradient lane below it describe the same 384 tracks. Drawn on a LOG axis: the median 16 bp bin reads 2.07 against a maximum of 1,097.6.
Shorkie · predicted ChIP-exo	a.u.	0–15.910071	mean predicted coverage over the 1,128 ChIP-exo tracks. A different assay, not a rescaled copy of the RNA-seq lane: genome-wide the two correlate at r = 0.38.
Shorkie · predicted ChIP-MNase	a.u.	0–8.468615	mean predicted coverage over the 20 ChIP-MNase tracks. Nearly independent of the expression lane -- genome-wide r = 0.08 -- which makes it the most different thing the model predicts, not a redundant one.
Shorkie · predicted 1,000-strain RNA-seq	a.u.	0–271.634308	mean predicted coverage over the 1,014 natural-isolate RNA-seq tracks. A different RNA-seq corpus from the lane above -- natural isolates rather than TF-induction timepoints -- and only r = 0.49 with it.
Shorkie · gradient x input	d log2 cov	-1.342147–1.342147	signed contribution of each base to the predicted RNA-seq of its window. SIGNED: bars above the zero rule are bases whose presence RAISES the prediction, below it bases that lower it. Genome-wide this differentiates each window's whole cropped interior, not a chosen gene, so it does NOT reproduce the per-locus figures on /shorkie-lab/shorkie/.
Shorkie · fold disagreement	sd / mean\|g\|	0–2	spread across the 8 released checkpoints, over the mean effect. Mid-lane is the reference: at 1.0 the across-fold spread EQUALS the mean effect. Blank on the quieter half of the genome by construction, not by omission.
Shorkie · condition-dependence	fraction	0–2	spread of predicted expression across the 13 induction timepoints. Where the model expects expression to DEPEND on the condition, rather than how much of it there is. High over regulated genes (HOP2 0.455, GAL3 0.297) and low over the constitutive glycolytic enzymes (PDC1 0.066, TDH3 0.083).
Shorkie · mutagenesis (ISM)	logSED	-0.647629–0.647629	the paper's own attribution: every substitution actually run. The paper's Figure 4 quantity, and the only method here that changes a base and looks. It exists on 3.10% of the genome — the 23 analysed windows — because genome-wide it is 1,231 hours. Everywhere else this lane is blank, which is the honest rendering of a measurement that was not made.
Shorkie · integrated gradients	d log2 cov	-0.172005–0.172005	signed contribution of each base, integrated from an all-zero-DNA baseline. SIGNED, and NOT a smoother version of the lane above: measured on chrI the two correlate at only r = 0.60 per base and share 27% of their strongest bases. A gradient is the slope AT the sequence; this is the integral along the path to it, so it still reports where the local slope has saturated.
Shorkie · occlusion	d log2 cov	-1.671213–1.671213	measured effect on the prediction of ablating each 64 bp of input. The only EXACT attribution here: every value is the model actually re-run with that stretch removed, not a derivative of it. 64 bp is the ablation window, so no finer level exists.
Shorkie - predicted H2B	a.u.	0–16.98889	predicted ChIP-MNase coverage for H2B. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted H2BK123ub	a.u.	0–15.305061	predicted ChIP-MNase coverage for H2BK123ub. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted H3	a.u.	0–19.805643	predicted ChIP-MNase coverage for H3. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted H3K27ac	a.u.	0–5.607394	predicted ChIP-MNase coverage for H3K27ac. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted H3K36me3	a.u.	0–9.756899	predicted ChIP-MNase coverage for H3K36me3. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted H3K4me3	a.u.	0–8.234179	predicted ChIP-MNase coverage for H3K4me3. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted H3K79me3	a.u.	0–2.880923	predicted ChIP-MNase coverage for H3K79me3. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted H3K9ac	a.u.	0–7.274961	predicted ChIP-MNase coverage for H3K9ac. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted H4K12ac	a.u.	0–6.985842	predicted ChIP-MNase coverage for H4K12ac. One histone mark, not the 9-mark average. The marks disagree strongly - pairwise shape correlation 0.60 and peaks 174 bins apart - so the pooled lane is the least representative average on this page.
Shorkie - predicted SUA7 ChIP-exo	a.u.	0–473.241211	predicted ChIP-exo coverage for SUA7 - TFIIB - pre-initiation complex. TFIIB - pre-initiation complex. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted SPT15 ChIP-exo	a.u.	0–48.441666	predicted ChIP-exo coverage for SPT15 - TBP - pre-initiation complex. TBP - pre-initiation complex. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted TFA1 ChIP-exo	a.u.	0–122.933655	predicted ChIP-exo coverage for TFA1 - TFIIE - pre-initiation complex. TFIIE - pre-initiation complex. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted TFB4 ChIP-exo	a.u.	0–353.792877	predicted ChIP-exo coverage for TFB4 - TFIIH - pre-initiation complex. TFIIH - pre-initiation complex. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted RAD3 ChIP-exo	a.u.	0–157.646637	predicted ChIP-exo coverage for RAD3 - TFIIH helicase. TFIIH helicase. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted CET1 ChIP-exo	a.u.	0–132.492691	predicted ChIP-exo coverage for CET1 - mRNA capping enzyme. mRNA capping enzyme. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted HTZ1 ChIP-exo	a.u.	0–54.234638	predicted ChIP-exo coverage for HTZ1 - H2A.Z - promoter-flanking variant. H2A.Z - promoter-flanking variant. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted SPT7 ChIP-exo	a.u.	0–9.891804	predicted ChIP-exo coverage for SPT7 - SAGA coactivator. SAGA coactivator. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted HHF1 ChIP-exo	a.u.	0–36.146435	predicted ChIP-exo coverage for HHF1 - histone H4. histone H4. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted RAP1 ChIP-exo	a.u.	0–31.419741	predicted ChIP-exo coverage for RAP1 - general activator - ribosomal and glycolytic promoters. general activator - ribosomal and glycolytic promoters. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted ABF1 ChIP-exo	a.u.	0–17.251659	predicted ChIP-exo coverage for ABF1 - general regulator - ARS-binding. general regulator - ARS-binding. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted REB1 ChIP-exo	a.u.	0–25.190792	predicted ChIP-exo coverage for REB1 - general regulator - terminator. general regulator - terminator. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted TBF1 ChIP-exo	a.u.	0–37.983974	predicted ChIP-exo coverage for TBF1 - general regulator - subtelomeric. general regulator - subtelomeric. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted CBF1 ChIP-exo	a.u.	0–16.522972	predicted ChIP-exo coverage for CBF1 - centromere and MET regulon. centromere and MET regulon. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted FHL1 ChIP-exo	a.u.	0–40.18956	predicted ChIP-exo coverage for FHL1 - ribosomal protein regulon. ribosomal protein regulon. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted SFP1 ChIP-exo	a.u.	0–10.059575	predicted ChIP-exo coverage for SFP1 - ribosome biogenesis. ribosome biogenesis. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted GAL4 ChIP-exo	a.u.	0–8.878935	predicted ChIP-exo coverage for GAL4 - galactose regulon activator. galactose regulon activator. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted MSN2 ChIP-exo	a.u.	0–10.913806	predicted ChIP-exo coverage for MSN2 - general stress response (STRE). general stress response (STRE). One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted HSF1 ChIP-exo	a.u.	0–10.297993	predicted ChIP-exo coverage for HSF1 - heat-shock factor. heat-shock factor. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted PHO4 ChIP-exo	a.u.	0–9.413221	predicted ChIP-exo coverage for PHO4 - phosphate starvation. phosphate starvation. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted INO4 ChIP-exo	a.u.	0–4.88596	predicted ChIP-exo coverage for INO4 - inositol / phospholipid. inositol / phospholipid. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted UME6 ChIP-exo	a.u.	0–9.924767	predicted ChIP-exo coverage for UME6 - early meiotic repressor. early meiotic repressor. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted GCN4 ChIP-exo	a.u.	0–5.031454	predicted ChIP-exo coverage for GCN4 - amino-acid starvation. amino-acid starvation. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted SWI4 ChIP-exo	a.u.	0–15.02518	predicted ChIP-exo coverage for SWI4 - SBF - cell-cycle G1/S. SBF - cell-cycle G1/S. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.
Shorkie - predicted STE12 ChIP-exo	a.u.	0–10.389616	predicted ChIP-exo coverage for STE12 - mating and filamentation. mating and filamentation. One target of 765, not their average: individual targets peak a median 38 bins apart, so the pooled ChIP-exo lane averages experiments that disagree about where the signal is.

Is the model just reading base composition? No — and here is the number

The first objection to any claim that a language model measures constraint is that it might only be measuring composition: AT-rich sequence is easier to predict than GC-balanced sequence for reasons that have nothing to do with function. The GC lane exists so that objection has an answer on the page rather than in a reader's suspicion.

over	bases	Pearson r	mean IC	mean GC
whole genome	12,157,105	-0.020	0.199	38.1%
coding sequence	8,598,345	0.004	0.220	39.6%
intergenic	3,558,760	-0.221	0.150	34.7%

Genome-wide, r = -0.020 — composition explains about 0.04% of the variance in the model's certainty. That is the reassuring answer, and it is the one a control is supposed to give.

But small overall is not zero everywhere. Inside coding sequence the correlation is 0.004; in intergenic sequence it is -0.221 — AT-rich intergenic sequence really is more predictable to the model. The same effect shows at chromosome scale: chrM is the most AT-rich sequence in the genome at 17.1% GC and also the most predictable, at 0.457 bits against a nuclear 0.197. Read a peak in an AT-rich region with that in mind.

The window is 50 bp, centred. A 5 bp window — UCSC's choice for its own GC track — takes only six values and is not a composition; the model's own 128 bp pooling grid would build the thing being controlled for into the control. As a check that the computation is right, the genome-wide figure comes out at 38.15% against the published 38.1%.

The annotation layer, and why the binding sites are in three tiers

Every feature comes from a flat file rather than a per-window API call: the SGD annotation for tRNA, snoRNA, ncRNA, snRNA and rRNA genes, replication origins, centromeres, telomeres, LTRs, transposons, pseudogenes and uORFs; the Harbison/MacIsaac regulatory code for transcription-factor binding; ORegAnno for literature-curated regulatory regions; phastCons element calls; and UCSC's tandem repeats.

The binding sites are split three ways and never merged, because they are three different claims. A ChIP-supported call is a measurement that the factor binds there. A conserved-only call is an argument that it might. A motif-only call says just that the letters look right. The expression playground measures attribution enriching 3.26× on the ChIP tier against 1.25× and 1.49× on the other two — and there are 122,225 motif-only calls genome-wide against 15,979 ChIP-supported ones, so one merged "TFBS" lane would bury the strong result under the weak one. Turn the tiers on separately in the track panel and the difference is visible directly.

JASPAR is deliberately absent. Unfiltered it is 16.7 million hits genome-wide — 1.4 per base, which is a scan rather than an annotation.

Above about 60,000 bp a feature lane draws a density profile instead of individual features, and the readout says which it is showing. That is what IGV does, and it is the difference between a lane that carries information and a solid bar.

Two models on one axis, and they predict different things

The lanes here come from two different networks that share an encoder and answer opposite questions, and reading one as the other is the mistake this page most invites.

Shorkie_LM (2 lanes) predicts the sequence itself: mask a base, ask what belongs there, and report how concentrated the answer is. Its units are bits, and a tall column means constrained — the surrounding sequence determines what goes here.

Shorkie (44 lanes) predicts what an assay would measure on that sequence — RNA-seq, ChIP-exo, ChIP-MNase — plus a signed per-base attribution of its own RNA-seq prediction. Its units are arbitrary coverage, and a tall column means expressed, which is a completely different claim. A gene body is high on both lanes for unrelated reasons.

Both were run over the same 8,192 bp cores in the same 16,384 bp windows on the same 128 bp pooling phase, so the two are aligned base for base and can be read against each other. Turn on exactly two lanes and the header prints their correlation over the visible window, which is the fastest way to see that a genome-wide number and a local one are rarely the same: model constraint against phastCons is 0.121 genome-wide and about 0.47 over a strongly expressed gene.

The coverage lanes are 16 bp, and no finer level exists. Shorkie's head emits 896 bins of 16 bp; storing a per-base version would be 12,157,105 numbers carrying 759,819 values of real information, drawn as though the model resolved single bases. The attribution is per base, because it is a derivative with respect to the one-hot input. When a lane is pinned at its own floor the resolution readout names it rather than letting the headline speak for every lane.

What every score track is — source, meaning, and how to read it

Four fields per track, and the fourth is the one that matters: every track here invites a specific misreading, and naming it is more useful than another sentence about what the track shows. These come from the generator's own track table, so a track cannot be added to the browser without them.

Shorkie_LM · masked bits · 0–2

Source. Shorkie_LM (Chao et al. 2025, bioRxiv 2025.09.19.677475), the fold-f0 checkpoint, run over sacCer3 in 16,384 bp windows. Not a published track.

Measures. Information content, 2 − H(p), of the model's four-way distribution at each base, from the K = 7 iterative masked pass: positions are split into 7 disjoint strided sets, each masked in turn, and every position read back only from the pass that masked it. 2 bits means all the probability on one base; 0 means chance.

How to read it. High where the surrounding sequence determines what belongs at a position — coding sequence, splice sites, strong binding sites. The nuclear genome averages 0.199 bits, so most of the track is genuinely low and a tall column is the exception rather than the baseline.

What it does not mean. Confidence is not importance. A base the model predicts well is predictable FROM 165 Saccharomycetales genomes, which is not the same claim as functionally essential — a repetitive tract is highly predictable and carries little function. Read a peak against the conservation and GC lanes before reading it as meaning.

Shorkie_LM · unmasked bits · 0–2

Source. The same model and the same run as the masked track, with nothing masked.

Measures. Information content from a single forward pass in which the model can see the base it is scoring. This is the quantity the paper's Figure 2A logo is built on.

How to read it. Use it to see what masking costs. It runs about 3.4× higher (0.687 bits against 0.199) and picks the right base 97.5% of the time — but confident about WHICH base is not the same as sharply peaked, which is why 97.5% accuracy still averages under 0.7 of a possible 2 bits.

What it does not mean. NOT A PREDICTION. The model is largely reading its own input, so its certainty here mostly measures how well it copies. It is drawn because the paper uses it and because the gap between the two passes is informative — never as evidence of what the model knows.

phastCons · 7 yeasts posterior · 0–1

Source. UCSC phastCons7way for sacCer3 — a phylo-HMM over a 7-species alignment (S. cerevisiae, paradoxus, mikatae, kudriavzevii, bayanus, castellii, kluyveri), from hgdownload.soe.ucsc.edu/goldenPath/sacCer3/phastCons7way/.

Measures. The posterior probability that a base lies in a conserved element, 0 to 1. A statement about what evolution has held still across those seven yeasts.

How to read it. The independent check on the model: Shorkie_LM is alignment-free and this is alignment-based, so agreement is evidence and disagreement is a question. Genome-wide they correlate at r = 0.121, and both rank coding sequence above intergenic.

What it does not mean. It SATURATES inside genes — 40.1% of coding bases sit at 0.99 or above, median 0.974 — so the within-CDS correlation of 0.045 is partly range restriction, not purely disagreement. It also has no value for 0.65% of the genome, drawn as a gap and never as zero.

GC content · 50 bp fraction · 0–1

Source. Computed from the sacCer3 reference on this machine; no external data. The genome-wide figure comes out at 38.15% against the published 38.1%, which is the check that the computation is right.

Measures. The fraction of A/C/G/T in a centred 50 bp window that are G or C. 50 bp because a 5 bp window takes only six values and is not a composition, while the model's own 128 bp pooling grid would build the thing being controlled for into the control.

How to read it. A CONTROL, and a small result is the reassuring one. If the model's information content were mostly base composition it would show here: measured, r = −0.020 genome-wide, so composition explains about 0.04% of the variance in model certainty.

What it does not mean. Small overall is not zero everywhere. In intergenic sequence r = −0.221 — AT-rich sequence really is more predictable to the model, which is also why chrM at 17.1% GC is the most predictable chromosome in the genome (IC 0.457 against a nuclear 0.198). Read a peak in an AT-rich region with that in mind.

Shorkie · predicted RNA-seq a.u. · 0–1097.560669

Measures. Predicted RNA-seq coverage, averaged over the 384 `_T0_` tracks — the untreated, glucose, vegetative baseline the paper's own Figure 4 mutagenesis is scored on. One value per 16 bp, which is the model's own output bin; nothing finer is written because nothing finer exists.

How to read it. High over transcribed genes and low between them. Because the axis is logarithmic, a bar at half height is roughly 30x the value of one at a quarter height, not twice it — the median 16 bp bin reads 2.07 and the maximum 1,097.6, so a linear axis would draw the median at 0.2% of the lane and the track would be a flat line with spikes.

What it does not mean. Arbitrary units, and NOT comparable with a real coverage file: these are the model's predictions, on the scale its training data happened to have. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end and no window can start before position 0 — 17,408 bases genome-wide, left as no data rather than filled from a window that never scored them. A single forward pass, not reverse-complement averaged, so that this lane reports the same number as the coverage panel on /shorkie-lab/shorkie/ for the same locus.

Shorkie · predicted ChIP-exo a.u. · 0–15.910071

Measures. Predicted ChIP-exo coverage averaged over the 1,128 ChIP-exo tracks: where the model expects a protein to be cross-linked to DNA, across every factor in that corpus at once.

How to read it. A different assay from the RNA-seq lane and only r = 0.38 with it genome-wide. It is also ORF-enriched (1.20x, against RNA-seq's 17.94x), which is enough to look like an expression track at a glance and not enough to be one — the site has already shipped that confusion once, by reading the paper's channel order instead of the released targets sheet's.

Shorkie · predicted ChIP-MNase a.u. · 0–8.468615

Measures. Predicted ChIP-MNase coverage averaged over the 20 MNase tracks — nuclease accessibility, which is what a nucleosome map is built from.

How to read it. Genome-wide this correlates with the expression lane at only r = 0.08, so it is close to an independent statement about the same sequence rather than a restatement of it. It is worth having for a second reason: no measured nucleosome track ships here, because the canonical chemical map (Brogaard 2012, GSE36063) is published only as raw reads — the smallest supplementary file is 238 MB — and UCSC's sacCer3 carries no nucleosome, RNA-seq or TSS signal among its 49 leaf tracks.

Shorkie · predicted 1,000-strain RNA-seq a.u. · 0–271.634308

Measures. Predicted RNA-seq coverage averaged over the 1,014 natural-isolate tracks — the 1,000-genomes yeast panel, rather than the TF-induction timecourse the lane above uses.

How to read it. r = 0.49 with the T0 lane: related, and far from the same track. The two disagree wherever expression depends on strain background rather than on the induction state.

Shorkie · gradient x input d log2 cov · -1.342147–1.342147

Source. Shorkie (Chao et al. 2025, bioRxiv 2025.09.19.677475), the fold-f0 checkpoint, run over sacCer3 in 1,493 windows of 16,384 bp on the same 8,192 bp cores as the language-model tracks, so the two models' lanes are aligned base for base. Not a published track. The derivative is taken through the PyTorch port, rc-averaged, and multiplied by the one-hot input.

Measures. How much each individual base contributes to the model's predicted log2 RNA-seq coverage — d log2(sum of T0 coverage + 1) / d input, multiplied by the input. Because the input is one-hot, this keeps the reference base's own contribution and is exactly zero at the three bases that are not there; that is the correct rendering of the quantity, not a simplification of it. Averaged over both strands, which is a test-time augmentation the paper adopts and not a symmetry: the model is not reverse-complement equivariant.

How to read it. SIGNED, so the lane has a zero rule in its middle. A bar UP is a base whose presence raises the predicted expression of its window; a bar DOWN is one that lowers it. The axis is symmetric and logarithmic in both directions because the quantity is heavy-tailed: the median base reads |0.0008| against a maximum of 1.34, so a linear axis would draw a typical base at 2.5% of half-height.

What it does not mean. This is NOT the attribution shown on /shorkie-lab/shorkie/, and it will not reproduce those figures. There, a reader picks a gene and the gradient is taken of that gene's predicted coverage. Genome-wide there is no chosen gene, so the target is each window's whole cropped interior — the only definition that exists at every base. Gradients superpose, so this is the sum of the per-gene attributions of everything in view. Second: a gradient is a LOCAL linear sensitivity, not the effect of actually changing the base. Against the shipped mutagenesis planes it agrees about DIRECTION and much less about magnitude: at 22 of 23 loci the single strongest substitution points the same way, and base by base the median correlation is 0.369 (range 0.05-0.65). The one exception is instructive rather than troubling — at GAL3 the gradient reads +0.0013 at the base mutagenesis calls strongest, which is essentially zero against a genome-wide median |gradient| of 0.0008. That is gradient SATURATION, the documented failure mode of a local derivative, and the reason mutagenesis and integrated gradients are on this page beside it. Full in-silico mutagenesis is the honest answer to 'what if this base changed', and it is not affordable genome-wide: measured, 1,231 hours.

Shorkie · fold disagreement sd / mean|g| · 0–2

Source. All EIGHT released Shorkie checkpoints (Chao et al. 2025, bioRxiv 2025.09.19.677475), folds f0-f7, each fetched from the paper's own storage bucket, each verified at 14,253,567 parameters and each byte-distinct from the others. Every one was run over the whole of sacCer3 in the same 1,493 windows on the same 8,192 bp cores as every other lane here, producing eight complete genome-wide gradient x input tracks; this lane is their per-base standard deviation divided by their per-base mean magnitude. Every other attribution lane in this browser comes from f0 alone.

Measures. How much the eight training runs DISAGREE about a base, relative to how much they claim it matters. A coefficient of variation: 1.0 means the spread across the checkpoints is as large as the effect they average to, 0.3 means they broadly agree about its size. This is a statement about the LANES ABOVE IT, not about the sequence — it is the only track here that measures the method rather than the genome.

How to read it. Against sk-gradient directly, position by position. A tall bar under a tall attribution bar is a base one checkpoint found and the others did not, and the per-base claim there should not be leaned on; a low bar under a tall one is a base all eight agree about. The genome-wide median is 0.775, so a typical scored base carries a spread about three quarters the size of its own effect — which is the honest summary of how much a single-checkpoint attribution track can be trusted at base resolution, and it agrees with the locked-panel result on /shorkie-lab/shorkie/, where the sign of a strong effect is stable across folds while the RANKING overlaps only 41-55% between any two. Mildly ANTI-correlated with loudness (Spearman -0.26 within the scored half): the folds agree relatively better where the effect is large, which is the reassuring direction. The axis is fixed at 0-2 so that mid-lane is exactly the 1.0 reference; 0.05% of scored bases exceed 2.0 and draw at full height.

What it does not mean. OPTIMISATION VARIANCE, NOT UNCERTAINTY. The eight folds share an architecture, a training corpus and a recipe, differing only in the cross-validation split and the run of the optimiser. So this bounds how much of a single checkpoint's attribution is an accident of training — it says nothing about whether the architecture is right, and a base where all eight agree is not thereby correct, only reproducibly claimed. Second: it is BLANK on half the genome, and that is by construction. Below the genome-wide median effect size, a ratio of two quantities both at the noise floor is not a trust statement, and the first version of this lane guarded that with an epsilon large enough to invert the reading — measured, the epsilon form ran +0.56 with loudness while the underlying ratio runs -0.26. A gap here means 'the attribution lane is not claiming much here', never 'the folds agree'. Third: the folds are compared on gradient x input, the cheapest of the four attribution methods and the one most exposed to saturation; a disagreement here is not automatically a disagreement about the mutagenesis lane.

Shorkie · condition-dependence fraction · 0–2

Source. Derived, not predicted separately: the 3,053 TF-induction RNA-seq tracks resolve into 13 timepoints (0-180 min) x 337 regulators, and this is the spread across those 13 timepoint means, (max - min) / (mean + 1), per 16 bp bin. Shorkie (Chao et al. 2025), fold f0.

Measures. How much the model expects a position's expression to move across the induction timecourse — condition-DEPENDENCE, not condition. A gene pinned at its maximum in every condition scores near zero however loudly it is transcribed; a gene that is silent in one state and active in another scores high however quiet it is on average.

How to read it. Against the coverage lane above it, not on its own. High here and low there is a regulated gene caught in its off state; low here and high there is a constitutive one. Over their own gene bodies the ordering is HOP2 0.455 (meiosis-specific, silent in vegetative growth), MMS2 0.352, POP4 0.302 and GAL3 0.297 (glucose-repressed) at the top, and the glycolytic enzymes at the bottom — PDC1 0.066, ADH1 0.081, FBA1 0.082, TDH3 0.083. A tenfold separation from a quantity nothing was tuned on. The axis is fixed at 0-2 rather than at the data maximum, because 1.0 has a meaning here — the spread across timepoints equals the mean — and a round axis is what lets two places on the genome be compared. 99 bins of 758,738 (0.013%, in 12 runs, maximum 2.494) exceed 2.0 and draw at full height.

What it does not mean. This lane exists BECAUSE the obvious alternative failed a measurement. The plan was 13 genome-wide timepoint lanes so a reader could pick a condition; they are indistinguishable — the lowest pairwise correlation among all 13 is 0.9923 and T5 against T0 is 1.0000 — because each averages ~300 regulators and averaging 300 induction experiments washes out every individual induction. The variation is real and lives in the INDIVIDUAL tracks: over GAL1's gene body the 3,053 of them span 43.7x. Those cannot ship genome-wide (337 regulators would be 276 MB), so individual conditions are available inside the 23 analysed windows and nowhere else. Second caveat: a spread of thirteen means is a floor on the true condition-dependence, never a measurement of it.

Shorkie · mutagenesis (ISM) logSED · -0.647629–0.647629

Source. Shorkie (Chao et al. 2025, bioRxiv 2025.09.19.677475), fold f0. Every one of the three substitutions at all 16,384 positions of each analysed window, on both strands — 98,304 forward passes a window — then rc-averaged, mean-centred across the four bases and projected on the reference, which is the paper's own recipe in all three files that implement it. The browser reads the same packs the per-locus panels do, through their own decoder.

Measures. What actually happens to the model's predicted log2 RNA-seq coverage when the base that is there is replaced. Not a derivative of that, not an integral of one, and not a block ablation: the finite difference itself, which is why it is the standard this page measures the other three methods against. Positive means the base that is present is HOLDING THE PREDICTION DOWN, so changing it would raise expression; negative means the base is doing work the model relies on.

How to read it. At base zoom it becomes the paper's sequence logo — letters above the zero rule, mirrored below. Read it against gradient x input directly above: the two agree about direction at 22 of 23 loci and correlate at a median of only 0.369 base by base, and where they disagree it is usually the gradient going flat on a saturated promoter rather than the mutagenesis being wrong.

What it does not mean. BLANK ON 96.9% OF THE GENOME, and that is a fact about what was computed rather than about the sequence. Full mutagenesis is 2,950 s a window — 1,231 hours, 51 days of GPU, for all 1,493 windows — so it was run on the 23 windows this site analyses and nowhere else. A gap here means 'not measured', never 'no effect'; the lane prints the fraction of the view that is missing so the distinction is on the screen. It also scores each window's own gene body, not the whole cropped interior the gradient lanes use, so the two answer subtly different questions even where both are present.

Shorkie · integrated gradients d log2 cov · -0.172005–0.172005

Source. Shorkie (Chao et al. 2025, bioRxiv 2025.09.19.677475), the fold-f0 checkpoint, run over sacCer3 in 1,493 windows of 16,384 bp on the same 8,192 bp cores as the language-model tracks, so the two models' lanes are aligned base for base. Not a published track. Integrated gradients at 32 steps from an all-zero-DNA baseline, averaged over both strands — and the completeness target is rc-averaged too, since the average of two complete decompositions is a decomposition of the average.

Measures. The same quantity as the gradient lane — each base's signed contribution to predicted log2 RNA-seq coverage — obtained by integrating the gradient along the straight path from a sequence with no DNA to this one, instead of reading the slope at this one. Its attributions sum to f(sequence) − f(baseline), which is the property it exists for, and it is deliberately NOT mean-centred across the four bases: that identity is a telescoping integral of the raw gradient, and subtracting a per-position mean breaks it (measured, 8–650% completeness error centred against 0.4–13% un-centred).

How to read it. The same way as the gradient lane — up raises the prediction, down lowers it — but do NOT read it as a smoothed copy of it. Measured on chrI the two correlate at r = 0.60 per base, 0.44 at 64 bp bins, and their strongest 2,000 bases overlap by only 27%. That gap is the point: a gradient is a local slope, so where a promoter is saturated it reads near zero while the path integral still records the base as load-bearing.

What it does not mean. Like the gradient lane, this differentiates each window's whole cropped interior rather than a chosen gene — the only target defined at every base — so it does NOT reproduce the region-conditioned figures on /shorkie-lab/shorkie/, where a reader picks a gene first. Gradients superpose, so what is drawn is the sum over everything in view. Two things specific to this lane. First, IG has a REFERENCE POINT and the gradient lane does not, and the reference is not neutral: from an all-zero-DNA input the model predicts 12.43, which is above 62% of real 16 kb windows. So the path from 'no DNA' to a real sequence runs downhill over most of the genome, and this lane comes out 56.2% negative where gradient × input is 50.2% — a fact about the baseline, not about the sequence, so the two lanes' overall balance must not be read against each other. Second, 32 steps is a numerical approximation of an integral, not the integral: measured on real windows the attributions sum to within 0.5–2.3% of f(sequence) − f(baseline), which is the identity the method exists for.

Shorkie · occlusion d log2 cov · -1.671213–1.671213

Source. Shorkie (Chao et al. 2025, bioRxiv 2025.09.19.677475), the fold-f0 checkpoint, run over sacCer3 in 1,493 windows of 16,384 bp on the same 8,192 bp cores as the language-model tracks, so the two models' lanes are aligned base for base. Not a published track. 256 ablations a window at 64 bp, both strands, each one a forward pass — 6.4 hours of GPU for the genome.

Measures. What the model's predicted log2 RNA-seq coverage actually does when 64 bp of input is removed, by zeroing its four DNA channels and re-running. This is the only EXACT method on the page: every other lane reports a derivative or an integral of one, and this reports the measured difference. Zeroing is how the paper's language model masks a position and is indistinguishable to the model from a run of N — so it asks whether the stretch carries information at all, where the motif-knockout panel's shuffle asks whether its ARRANGEMENT matters. Those are different questions and give different answers.

How to read it. Down means removing that stretch LOWERS the prediction, so the model was relying on it; up means removing it raises the prediction, which is what a repressive element looks like. Because a whole 64 bp goes at once, a single decisive base and a diffuse stretch of sixty-four weak ones read the same here — the per-base lanes above separate them.

Shorkie - predicted H2B a.u. · 0–16.98889

Source. Shorkie (Chao et al. 2025, bioRxiv 2025.09.19.677475), fold f0, run over sacCer3. The ChIP-MNase block is 20 tracks over 9 histone marks; this lane is the mean of one mark's replicates rather than of all 20.

Measures. Predicted ChIP-MNase coverage for H2B - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

How to read it. Against the expression lane, not alone -- and switch marks in this one lane to see the canonical 5'/3' split, which the model reproduces from sequence with no chromatin input at all. Measured over the 23 analysed genes, normalised to each gene's own mean and oriented 5' to 3': H3K4me3 reads 1.50 over the first quarter against 0.85 over the last, H3K9ac 1.38/0.86 and H3K27ac 1.32/0.88 -- the promoter-proximal marks. H3K36me3 inverts it, 0.86/1.14, which is the co-transcriptional gene-body mark. Bulk H2B and H3 stay flat (0.97/1.08 and 0.93/1.09), as histones with no positional preference should.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. No MEASURED nucleosome or histone track ships here for comparison: the canonical chemical map (Brogaard 2012, GSE36063) is published only as raw reads and UCSC's sacCer3 carries no nucleosome signal, so there is nothing on this page to check these against.

Shorkie - predicted H2BK123ub a.u. · 0–15.305061

Measures. Predicted ChIP-MNase coverage for H2BK123ub - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

Shorkie - predicted H3 a.u. · 0–19.805643

Measures. Predicted ChIP-MNase coverage for H3 - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

Shorkie - predicted H3K27ac a.u. · 0–5.607394

Measures. Predicted ChIP-MNase coverage for H3K27ac - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

Shorkie - predicted H3K36me3 a.u. · 0–9.756899

Measures. Predicted ChIP-MNase coverage for H3K36me3 - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

Shorkie - predicted H3K4me3 a.u. · 0–8.234179

Measures. Predicted ChIP-MNase coverage for H3K4me3 - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

Shorkie - predicted H3K79me3 a.u. · 0–2.880923

Measures. Predicted ChIP-MNase coverage for H3K79me3 - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

Shorkie - predicted H3K9ac a.u. · 0–7.274961

Measures. Predicted ChIP-MNase coverage for H3K9ac - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

Shorkie - predicted H4K12ac a.u. · 0–6.985842

Measures. Predicted ChIP-MNase coverage for H4K12ac - nuclease accessibility under an antibody for that mark, which is what a chromatin map is built from.

Shorkie - predicted SUA7 ChIP-exo a.u. · 0–473.241211

Source. Shorkie (Chao et al. 2025, bioRxiv 2025.09.19.677475), fold f0, run over sacCer3. The ChIP-exo block is 1,128 tracks over 765 targets; this lane is the mean of one target's replicates rather than of all 1,128.

Measures. Predicted ChIP-exo coverage for SUA7 (TFIIB - pre-initiation complex) - where the model expects this protein to be crosslinked to DNA.

How to read it. A sharp peak is a predicted binding footprint; a broad one over a promoter is usually the pre-initiation complex rather than a sequence-specific site. The nine general-machinery targets in this picker exist to make that distinction visible - they are also the targets that differ most from the pooled ChIP-exo mean, which is why that mean is dominated by promoter shape.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted SPT15 ChIP-exo a.u. · 0–48.441666

Measures. Predicted ChIP-exo coverage for SPT15 (TBP - pre-initiation complex) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted TFA1 ChIP-exo a.u. · 0–122.933655

Measures. Predicted ChIP-exo coverage for TFA1 (TFIIE - pre-initiation complex) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted TFB4 ChIP-exo a.u. · 0–353.792877

Measures. Predicted ChIP-exo coverage for TFB4 (TFIIH - pre-initiation complex) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted RAD3 ChIP-exo a.u. · 0–157.646637

Measures. Predicted ChIP-exo coverage for RAD3 (TFIIH helicase) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted CET1 ChIP-exo a.u. · 0–132.492691

Measures. Predicted ChIP-exo coverage for CET1 (mRNA capping enzyme) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted HTZ1 ChIP-exo a.u. · 0–54.234638

Measures. Predicted ChIP-exo coverage for HTZ1 (H2A.Z - promoter-flanking variant) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted SPT7 ChIP-exo a.u. · 0–9.891804

Measures. Predicted ChIP-exo coverage for SPT7 (SAGA coactivator) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted HHF1 ChIP-exo a.u. · 0–36.146435

Measures. Predicted ChIP-exo coverage for HHF1 (histone H4) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted RAP1 ChIP-exo a.u. · 0–31.419741

Measures. Predicted ChIP-exo coverage for RAP1 (general activator - ribosomal and glycolytic promoters) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted ABF1 ChIP-exo a.u. · 0–17.251659

Measures. Predicted ChIP-exo coverage for ABF1 (general regulator - ARS-binding) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted REB1 ChIP-exo a.u. · 0–25.190792

Measures. Predicted ChIP-exo coverage for REB1 (general regulator - terminator) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted TBF1 ChIP-exo a.u. · 0–37.983974

Measures. Predicted ChIP-exo coverage for TBF1 (general regulator - subtelomeric) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted CBF1 ChIP-exo a.u. · 0–16.522972

Measures. Predicted ChIP-exo coverage for CBF1 (centromere and MET regulon) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted FHL1 ChIP-exo a.u. · 0–40.18956

Measures. Predicted ChIP-exo coverage for FHL1 (ribosomal protein regulon) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted SFP1 ChIP-exo a.u. · 0–10.059575

Measures. Predicted ChIP-exo coverage for SFP1 (ribosome biogenesis) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted GAL4 ChIP-exo a.u. · 0–8.878935

Measures. Predicted ChIP-exo coverage for GAL4 (galactose regulon activator) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted MSN2 ChIP-exo a.u. · 0–10.913806

Measures. Predicted ChIP-exo coverage for MSN2 (general stress response (STRE)) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted HSF1 ChIP-exo a.u. · 0–10.297993

Measures. Predicted ChIP-exo coverage for HSF1 (heat-shock factor) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted PHO4 ChIP-exo a.u. · 0–9.413221

Measures. Predicted ChIP-exo coverage for PHO4 (phosphate starvation) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted INO4 ChIP-exo a.u. · 0–4.88596

Measures. Predicted ChIP-exo coverage for INO4 (inositol / phospholipid) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted UME6 ChIP-exo a.u. · 0–9.924767

Measures. Predicted ChIP-exo coverage for UME6 (early meiotic repressor) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted GCN4 ChIP-exo a.u. · 0–5.031454

Measures. Predicted ChIP-exo coverage for GCN4 (amino-acid starvation) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted SWI4 ChIP-exo a.u. · 0–15.02518

Measures. Predicted ChIP-exo coverage for SWI4 (SBF - cell-cycle G1/S) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

Shorkie - predicted STE12 ChIP-exo a.u. · 0–10.389616

Measures. Predicted ChIP-exo coverage for STE12 (mating and filamentation) - where the model expects this protein to be crosslinked to DNA.

What it does not mean. Arbitrary units, and a PREDICTION rather than an experiment - this is where the model expects the assay to read, not where anyone measured it. The first 1,024 bases of every chromosome are blank because the head crops that much from each window end. And 25 targets of 765 (or 9 marks) is a selection: a factor absent from the picker is absent because it was not shipped, never because the model predicts nothing for it. ChIP-exo crosslinks whatever is in the complex, so a predicted peak for a coactivator is not evidence that it touches DNA - the browser's motif popup names seven factors here whose absent JASPAR matrix is explained by exactly that.

The annotation lanes carry the same four fields; open what this is beside any of them in the track panel.

What the constraint track is not

It is one number a base — how concentrated the model's four-way distribution is — and not which base the model expects. The full constraint logo, the composition-floor comparison and the region statistics need all four probabilities at every position, which is 49 M values genome-wide; those stay on the 23 primary regions, which ship complete packs. Use the jump menu above to move between the two.

High information content means the model is confident, not that a base is functionally important — the two correlate but are not the same claim, and the model was trained on 165 Saccharomycetales genomes, so a confidently predicted base is one that is predictable from that comparative context. Coding sequence comes out more constrained than the genome average in 14 of 14 tested windows; LTRs and transposons come out less. chrM is far more predictable than any nuclear chromosome because it is AT-rich and repetitive, which is a property of the sequence rather than of its function — and it is also where phastCons has the least to say, at 77% coverage against over 99% on every nuclear chromosome.