Reading the analysis
Once an assembly is clean, OpenPretext computes a set of measurements from the contact map: contact decay, compartments, misassembly signatures, and model predictions. This page explains what each readout means, with interactive figures for the main ones. If you are new to curation, begin with the field guide.
What the numbers in the panel mean
Open the 3D Analysis panel and it reports a set of slopes, scores, and plots. This section explains how to read each one and what a healthy value looks like.
The exponent and the decay curve
Drag the slider. The contact map updates and the decay curve changes slope at the same time. The green band, on the slider and between the dashed lines, marks the healthy range.
Contact decay, P(s)
How fast contact frequency falls as two points move apart along a chromosome, measured as the slope of the decay on a log-log plot. A well-packed genome sits between −1.0 and −1.5, drawn as dotted reference guides on the chart. Below the curve, a companion panel plots the local slope against distance, so a short-range plateau or a long-range roll-off shows up as a departure from the healthy band even when the single fitted slope looks fine.
When to worry: a slope near 0 means contacts barely fade with distance, which points to poor compaction or a misassembly stitching distant regions together. A slope past −2 usually means the data has been over-corrected.Fit quality, R²
How closely the decay follows a single clean power law. Above 0.9 the curve is tight. A low value means the points scatter, often because contigs of different quality are mixed together.
Look closer: the per-chromosome list breaks this down, so you can see which chromosomes fit cleanly and which are worth inspecting. A whole-genome R² can look modest while individual chromosomes fit well.The composite health score
One number from 0 to 100 that folds five measurements together: contiguity (N50), decay quality (P(s)), integrity (fewer misassembly flags), compartment strength (A/B), and library quality (the cis to trans ratio, which needs scaffolds assigned). Above 70 is a well-assembled genome.
Read it as a trend, not a grade: a fresh assembly scores low on integrity and library before anything has been cut or scaffolded. It increases as curation progresses, and that trend is more informative than any single value.Insulation and boundaries
The insulation score dips wherever a region is walled off from its neighbors. The deepest dips mark boundaries between chromosome neighborhoods, drawn as red markers along the map. In a clean, well-ordered assembly the strongest boundaries line up with real structure, not with the joins between contigs.
When to worry: a boundary sitting exactly on a contig edge, with the map continuous across it, is a seam in the assembly rather than biology.Join support at every junction
For each contig junction, OpenPretext checks whether contact keeps flowing across the boundary. A correct join carries contact over the seam about as strongly as it holds inside each contig. A wrong join or a misorder shows the opposite: contact drops away across the boundary, a dark square opens up off the diagonal, and the junction is flagged as a weak join with a red marker on the map. The check is scaffold aware, so a junction between two contigs you have assigned to different scaffolds counts as an intended chromosome boundary and is left alone.
When to worry: a weak join inside a single scaffold, where two pieces meant to be one chromosome barely touch across the seam. This usually means a cut is needed, or that two contigs require reordering before they will join cleanly.Haplotigs: the alternate haplotype left behind
A retained haplotig is a piece of the second haplotype that was never merged into the primary assembly, so the same region ends up represented twice. It leaves two marks. In the contact map it shows a bright off-diagonal block against its one homologous primary, because Hi-C reads from the alternate haplotype map to both copies. In read depth it sits near half the assembly's coverage, because the reads split across the two haplotypes. OpenPretext uses the contact block to find candidates and coverage to confirm them, then lets you tag the confirmed ones so they drop out of the primary set.
Confirmed against unconfirmed: a bright block on its own is not proof. Segmental duplications and repeats make the same pattern. Only when the block and half coverage agree is a contig reported as a confirmed haplotig. Without a coverage track, candidates stay unconfirmed, so load coverage before you trust the call.A/B compartments and the saddle plot
Chromatin sorts into two compartments: A, gene-rich and active, and B, quiet and dense. On the map they show as a checkerboard, where A talks to A and B talks to B across long distances. The saddle plot distills that checkerboard into one square: bins sorted from most B to most A, then averaged, so strong corners with a weak middle mean the two compartments are cleanly separated. The strength number puts a figure on that contrast.
Which side is A: Hi-C alone splits the two compartments but cannot say which is the active one, since the sign of the split is arbitrary. OpenPretext settles it with GC content when you load a reference FASTA, labeling the gene-rich, GC-rich side as A. Without a FASTA the split is still real but the A and B labels are left unoriented.Virtual 4C
Alt-click any point on the map to pull out that single locus's contact profile across the whole genome, the way a 4C experiment would. It turns a two-dimensional map into one readable track anchored at one place.
When to worry: a sharp spike far from the diagonal but on the same chromosome is worth a look. It is either a wrong join pulling distant sequence together, or a real long-range contact.Observed over expected
Contact frequency drops steeply with distance, and that decay dominates the raw map so much that faint long-range structure is hard to see. The O/E view divides every cell by the genome-wide average at its separation, so the decay cancels out and what is left is contact that is stronger or weaker than expected. On a diverging blue-to-red scale, red is enrichment and blue is depletion, and the compartment checkerboard and off-diagonal blocks stand out. This is the same observed/expected map the compartment analysis is built on, shown directly.
Overview only: the transform needs the genome-wide average, which exists for the whole-genome overview but not for a single detail tile. Zoom in past the overview and the raw tiles take over, so read O/E as a low-zoom diagnostic, not a replacement for the detailed map.Contiguity and the reference-quality standard
Alongside the health score, the stats panel reports the numbers assembly projects track: contig and scaffold N50, auN (a length-weighted average that moves more smoothly than N50), the largest scaffold, and the fraction of the assembly assigned to chromosomes. Scaffold N50 is the one that climbs as you group contigs into chromosomes, so it is the number to watch during curation. These are shown against the Earth BioGenome Project standard written as 6.C.Q40: a contig N50 above 1 Mb, chromosome-scale scaffolds, and fewer than one base error in 10,000.
What it cannot tell you: the Q in that standard is base accuracy, and measuring it needs the sequencing reads, which a Hi-C contact map does not carry. OpenPretext reports the contiguity and chromosome-assignment parts against the targets and leaves base accuracy marked as not assessed, rather than implying an assembly meets a standard it has only half-checked.A and B, and the saddle plot
Chromatin sorts into two compartments: A, gene-rich and active, and B, quiet and dense. On the map they show as a checkerboard; the saddle plot condenses that checkerboard into a single square, which makes the degree of separation easier to read.
Adjusting compartment strength
Slide the strength. The checkerboard on the map sharpens and the saddle plot's corners pull apart. Strong AA and BB corners with a weak middle mean the two compartments are cleanly separated.
The four signatures
Most assembly errors leave a characteristic shape on the contact map. The four below cover the common cases, each corrected with a cut, a join, or a flip.
Select a signature
Each button draws the pattern that error makes on the map. The note describes what you are seeing and the operation that corrects it.
Inversion
What to do
Join support across a junction
A join is supported only if Hi-C contact continues across it. OpenPretext scores every contig junction by how well the signal carries over the seam. Where the signal holds, the join is supported. Where it drops away, the junction is flagged for inspection.
Weakening a join until it is flagged
Two contigs meet at the junction in the middle of the map. Slide the junction signal down and the contact across the boundary fades into a dark square. Below a threshold the junction is marked as a weak join, with a red marker where the two contigs meet.
What a model can add: Evo2HiC
OpenPretext can run predictions from a companion model alongside your real data. These do not measure your genome. They estimate what a model expects and lay it beside the map so you can compare. They need the local server running, and the full effect needs real model weights; the default mode does a light smoothing so you can see the wiring.
Sharpen a sparse map
Denoise and upscale a thin contact map so faint structure is easier to see, without touching the values you actually curate against.
Epigenomic tracks from sequence
Estimate signals such as accessibility or histone marks straight from the assembled sequence, shown as extra tracks along the map.
A map from sequence alone
Predict a contact map from sequence with no experiment, a reference for what the folding should look like when the real map is noisy or incomplete.
Treat every prediction as a hypothesis to check against the real map, never as a replacement for it. The measured contacts are the ground truth; the model is a second pair of eyes.
Reading these numbers in the app
Every readout described here appears in the app's 3D Analysis panel. Load an assembly, compute the analyses, and the panel links back to these explanations.