A trained ChromBPNet model that you cannot interpret is just an expensive curve fitter. The reason researchers invest in this framework is that its models can be opened up: you can ask which bases in a sequence drive the predictions, gather those driving patterns across the genome, and recover the transcription factor motifs the model has learned. Two tools do that work. Contribution scores assign credit to individual bases, and TF-MoDISco clusters the high-credit patterns into motifs. This article explains both, in the order you would actually use them, and how to read what they give you without inventing biology that is not there.
Step one: contribution scores, credit at single-base resolution
A contribution score answers a precise question for a given region and a given prediction: how much did each base in this sequence contribute to what the model predicted? ChromBPNet computes these scores with a DeepLIFT-style attribution approach, a family of methods that distributes the prediction back across the input bases. The output is a track over your region, one value per position, where high positive values mark the bases the model’s prediction leans on.
Read a contribution track over a strong peak and a pattern jumps out. Most positions sit near zero; the model is indifferent to them. Scattered short stretches light up, often a handful to a couple of dozen bases wide, and those lit stretches are the sequence the model believes creates accessibility there. When several lit stretches appear in one peak, you are often looking at the combined handiwork of multiple binding events. The track converts a black-box prediction into a map of candidate functional sequence, at the resolution the biology actually operates on.
Why per-base credit changes what you can ask
Summary statistics about a region, however accurate, cannot tell you mechanism. Contribution scores can, because they localize. Three uses follow directly:
- Hypothesis generation: the lit stretches in a peak are your candidate binding sites, nominated by the model rather than by scanning for known patterns.
- Variant interpretation: a variant sitting inside a high-contribution stretch has a mechanism-shaped story behind its predicted effect, while one in a dead zone starts with explaining to do.
- Model auditing: if contribution tracks across many peaks light up on the same suspicious pattern, the model may be keying on an artifact, and you have learned something about your training data before it embarrasses you in a figure.
That third use is underrated. Interpretation is also quality control, and a model whose important bases look biologically coherent is a model whose predictions deserve more trust everywhere else, including variant scoring.
Step two: TF-MoDISco, from scattered patterns to motifs
Contribution tracks are per region, and a genome has a lot of regions. You cannot eyeball thousands of lit stretches and assemble the cast of regulators by hand. TF-MoDISco is the tool that does the assembly. It takes the high-contribution patterns found across many regions, clusters the ones that resemble each other, and distills each cluster into a motif: the recurring sequence signature of a transcription factor’s binding preference.
The logic deserves a slow read, because it inverts the usual motif workflow. The traditional approach scans sequence with a library of known motifs and reports matches. TF-MoDISco works from the model outward: the patterns come from what the network actually uses, clustered by similarity, and only afterward do you compare them with known transcription factor signatures to name them. Motifs the model relies on but the libraries describe poorly can still surface, because nothing in the clustering step requires prior permission from a database.
Reading a motif result like a skeptic
A TF-MoDISco run hands you a set of motifs, and the working researcher’s job is appraisal, not applause. For each motif, ask the standard questions. Does it resemble the known binding signature of a transcription factor that makes sense in this cell type? A beautiful motif for a factor that is not expressed in your cells is a mystery to investigate, not a result to publish. How much data backs it: motifs assembled from many contributing regions across the genome carry more weight than a cluster built from a handful. And where does it appear: a motif concentrated in the peaks of the relevant cell type, positioned sensibly within them, behaves like biology, while one scattered without pattern behaves like a warning.
Expect a mixture. Real runs return the well-known regulators of the cell type, some motifs matching factors you did not expect, and a residue of patterns that resist naming. That mixture is the honest output of learning from data, and it is why interpretation is a research step rather than a report generator.
The bias connection, one more time
Interpretation is where the bias-factorized design proves itself in public. In a model trained without bias separation, contribution scores cheerfully assign high credit to sequences the assay enzyme prefers, and TF-MoDISco will dutifully cluster those into crisp motifs that are pure instrument artifact. The clusters look exactly like results. In a bias-factorized ChromBPNet model, the enzyme’s preferences were accounted for during training, so the contribution landscape the interpretation tools explore belongs to the accessibility model, and the motifs it yields are candidates for real regulatory biology. Same interpretation pipeline, profoundly different evidentiary value, decided back at training time.
A sensible interpretation workflow
Put together, the working order is steady:
- Validate the trained model first; interpretation of an undertrained or mismatched model produces confident nonsense.
- Compute contribution scores over your peak regions and inspect a handful of strong peaks by eye to calibrate your expectations.
- Run TF-MoDISco over the scored regions to cluster patterns into motifs.
- Appraise each motif against the biology of your cell type, its supporting count of regions, and its placement within peaks.
- Feed the survivors into the rest of your study: variant interpretation, comparisons across cell types or conditions, and the design of follow-up experiments.
From motifs back to variants
Interpretation also feeds the framework’s other headline use. Once TF-MoDISco has told you which motifs your model trusts in a cell type, variant scoring gains a mechanistic lens: a variant that lands inside a high-contribution instance of a strong motif, and changes the bases that motif depends on, is a far more compelling candidate than one with a similar score in anonymous sequence. Conversely, a variant predicted to change accessibility while touching nothing the interpretation pass found important deserves a second look before it earns a follow-up experiment. The two analyses cross-examine each other, and candidates that survive both are the ones worth carrying into the wet lab.
This is the deeper reason interpretation sits at the center of the ChromBPNet workflow rather than at its decorative edge. Predictions tell you what the model thinks. Contribution scores and motifs tell you why it thinks it, in the vocabulary of transcription factor binding. A framework that can explain itself in that vocabulary can be argued with, improved, and trusted selectively, which is all a research tool can honestly offer.
Conclusion
Interpreting a ChromBPNet model is a two-step craft. Contribution scores, computed with a DeepLIFT-style attribution, tell you which bases in each region the prediction actually rests on. TF-MoDISco gathers those decisive patterns across regions and clusters them into motifs, the sequence signatures of the transcription factors shaping accessibility in your cell type. Read the results with a skeptic’s checklist, cell-type sense, supporting depth, sensible placement, and they become one of the strongest outputs in functional genomics: a data-driven census of the regulatory grammar your model has learned, and a set of mechanistic hypotheses precise enough to test.
Read next:
