Explore the balance between self-association and partner binding. Screen bZIP charge substitutions, inspect every contact, and compare the structural evidence.
Live browser calculationsNo sequence uploadResearch use only
What runs nowRegister detection → charge-contact scoring → constrained mutation screening. This is an explainable rule-based model, not trained ML. Structural predictions are precomputed; no AlphaFold job is launched.
Ready to analyze. Results are computed from the sequence and selected partners on this device.
02 / INSPECT THE MECHANISM
Selectivity readout
Computed locally
WT self score—
Variant self score—
Self score Δ—
Worst partner Δ—
Δ = variant − wild type. More negative self Δ is the screening objective; partner Δ near zero preserves the model score. These are contact counts, not binding energies or affinities.
Detected heptad register
Heptad
g
a · core
d · core
e
Partner score comparison
Pair
WT
Variant
Δ count
Δ sensitivity²
+1 attraction0 neutral−1 repulsion
03 / PRIORITIZE EXPERIMENTS
Constrained variant screen
Searches E/K/Q substitutions at e/g positions. Filters: self Δ < 0 and selected partner tolerance. Ties: worst partner Δ, fewer hydrophobic e/g edits, mean partner Δ, then fewer substitutions. This is not a stability model.
Top 20 qualifying variants
Variant · click to inspect
Edits
Self Δ
Worst partner Δ
Mean partner Δ
04 / REVIEW INDEPENDENT EVIDENCE
Structure, with its limitations visible.
Precomputed AlphaFold2-Multimer
● Chain A ● Chain B · Cα trace · drag / arrow keys
Changing the live sequence does not generate or update these structures. The WT homodimer file uses a longer construct than the zipper-only confidence table.
Do not rank binding strength by ipTM.
In the supplied zipper-only set, six of eight homodimer models are antiparallel. ATF3–JUN ipTM varies only from 0.72 to 0.73. Neither result establishes mutant affinity or therapeutic activity.
Inspect all 16 zipper-only prediction summaries
Supplied model confidence and available topology annotations
¹ Ten same-heptad g↔e′ contacts assume a parallel dimer and five complete heptads. Core packing, helix propensity, solvent, DNA cooperativity and cellular context are not modeled.
² The sensitivity readout assigns Q/N a hypothetical partial charge. It is a stress test, not a physical charge assignment or a calibrated energy calculation.
The supplied literature-peptide check concerns ATF3W variants, not experimental validation of m3 or m8. A shortlist still needs binding, selectivity and functional assays.
Today: transparent contact rules plus stored AF2 model outputs. Next: collect matched Kd/Tm and partner-panel measurements; compare protein embeddings and structure features against this baseline.
Use held-out sequence families, calibrated uncertainty and active learning to choose informative experiments. Combine scores only after independent validation; no learned ATF3 selectivity model is deployed here.
Uses the supplied ATF3_bZIP engine, sequence snapshots, reference tables and PDB files. Run records include input sequence, constraints and SHA-256 identifiers.