pystreme

GPU-batched reimplementation of STREME for de novo motif discovery, callable in-process from Python and Jupyter: no subprocess, no FASTA round trip, results as Motif objects, DataFrames and figures.

from pystreme import MotifDiscovery, summary

disc = MotifDiscovery.from_bed("peaks.bed", "genome.fa", size=200, device="cuda")
motifs = disc.fit(n_motifs=3)
summary(motifs)

What it does

• Fixed-width extraction around peak centers (HOMER -size semantics), 2-bit encoded, GPU-resident, with an order-k Markov background fitted on the control set.
• STREME-style round loop (fit): dinucleotide-shuffled (or GC-matched, or explicit) control, hold-out split, k-mer seeds, batched-convolution PWM refinement, hold-out re-test of each motif's threshold, site erasing, patience-based stopping.
• Positional output for free: per-site offsets from the peak center and a CentriMo-style central-enrichment statistic, which separates the bound factor's own motif from co-occurring ones.
• Independently useful pieces: a batched PWM scanner (scan), a SEA-style enrichment test for a known motif (enrichment), tables (summary, sites_frame), database annotation (annotate), logos and positional histograms (plot), MEME-format export (to_meme).

Does it work?

Yes, measured against the real thing -- see Validation:

	pystreme	STREME 5.5.0
7232 SP1 ChEC-seq peaks, 10 motifs	same primary motif, 8 of 10 shared (Tomtom)
wall-clock, one A40 (reserved)	62 s	191 s
scanner vs FIMO (L1)	Pearson r = 1.00000, 99.95% identical best sites

Where to go next

• Usage guide: every argument of fit, how to read a Motif, controls, exports, pitfalls.
• Running on a cluster: LSF and SLURM recipes, and the pitfalls of shared clusters.
• Design document: why it is built this way, and what was deliberately left out.
• API reference.

Status

Phases 0-5 of the design document are done. Not done by choice: a yeast ChEC-seq regime for the STREME head-to-head, and repeat-aware GC matching.