Genome browser

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:

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.

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.

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.

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.

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 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.

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.

1,783 tiles of 65,536 bins across 9 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.

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.

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%.

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.

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 (5 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.

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.

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

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.