Scientific question

Can we propose sequences whose motif content is tunable—here combining a PWM with a simple latent “strength” decoder—and track that change along an interpolation path?

What we would conclude with this

PWM sampling gives an interpretable generative baseline; latent mixing between PWM and background mimics how deep models might smoothly vary motif usage. This is a toy illustration, not a trained genome-scale generative model.

Synthetic data

A 12 bp DNA PWM (CSV) and a background base composition; sequences are 40 bp with a fixed motif window. Seed: 42 (data), 7 (sampling). See demos/scientific-generative-sequences/data/generate.py and src/run.py.

Approach

  1. Classical: sample motif positions from row-wise PWM; negatives use background only.
  2. Latent: map z ∈ ℝ² through a logistic decoder to strength s(z) ∈ (0,1); per-position probabilities are s·PWM + (1−s)·background (renormalized).
  3. Score sequences with PWM log-odds; plot latent path and mean score vs interpolation t.

Key outputs

Motif composition

Score distributions

Latent interpolation

Reproduce

cd demos/scientific-generative-sequences
python3 data/generate.py
python3 src/run.py

Dependencies: demos/requirements.txt.