Insights and logos

logo = model.insights.logo(15)                       # all sequences of length 15
common = model.insights.logo(15, pseq_above=1e-8)    # only the common ones
rare = model.insights.logo(15, pseq_below=1e-8)      # only the rare ones

logo.consensus_sequence
logo.fraction             # share of the probability at length 15 this logo describes
pd.DataFrame(logo.as_dict())                         # {residue: heights}, ready for logomaker

For each position of a given length, a logo gives the distribution of residues and its information content (information_bits, information_nats), with the stack heights to draw it from in height_bits or height_nats (one row per position, one column per residue). It is computed over every sequence of that length, so no small-sample correction applies. Information is measured against a uniform background; for nucleotide models, which have five residues including N, pass reference_residues=4 for the usual two-bit axis.

Common and rare sequences

pseq_above and pseq_below restrict a logo to sequences above or below a probability; the two fractions, each a share of the probability at that length, add up to one. surprisal_above and surprisal_below take the threshold as a surprisal $-\ln P(s)$ in nats instead, so surprisal_above selects the rare sequences, like pseq_below. Give at most one condition.

These conditional logos are computed on a fine grid of probabilities: resolution_error reports the effect of the grid on a logo, max_resolution_error= refuses a logo whose grid error would exceed it, and bins= or spacing= set the grid.

model.insights.logos(15, edges=[...]) returns several logos at once, one for each band between consecutive surprisal edges (in nats), commonest first. model.insights.cost(15, ...) returns the states, edges and memory a logo needs, checked against the memory budget.

Member	Meaning
consensus_sequence, consensus_probability	The most probable residue at each position, and its probability there.
information_bits, information_nats	Information content per position.
height_bits, height_nats	Stack heights per position and residue.
fraction	Share of the probability at this length the logo describes.
n_support	Per position, how many residues carry any probability.
as_dict(unit="bits")	{residue: heights}.