QASC — Quantum Allosteric Scanner

Continuous-time quantum-walk prediction of allosteric sites from a single, unbound (apo) protein structure. Given only per-residue Cβ coordinates and the active-site residues, QASC ranks every residue by how strongly a quantum walk seeded at the active site couples to it, and returns the top-5 candidate allosteric residues.

Dependency-light — numpy + scipy only. No molecular dynamics, no bound (holo) structure, no deep-learning stack, and no fitted parameters at inference.

---

Method

QASC models a protein as a residue contact graph and propagates a signal from the active site with a continuous-time quantum walk, U(t) = exp(-iHt). The per-residue score fuses two complementary quantum signals:

1. CTQW communicability — the infinite-time-averaged transfer probability of the quantum walk on the Kirchhoff (Laplacian) Hamiltonian. It measures how strongly each residue is dynamically connected to the active site, and recovers accessible and cryptic pockets.

2. IPR-resonant transfer — anchor-seeded transfer weighted by each eigenmode's inverse participation ratio. It surfaces spatially localized modes that trap coherence in collapsed or deeply buried packing defects — pockets that diffusion-like methods can miss.

The two are combined by a noisy-or fusion

score(i) = 1 - (1 - rank[CTQW(i)]) * (1 - 0.5 * rank[IPR(i)])

where rank[·] is the per-protein rank percentile. The score is spatially pocket-smoothed, and the five highest-scoring residues distal to the active site are returned. The fusion is a single fixed formula applied identically to every protein — nothing is fit to the scored targets.

Success is measured with a one-sided permutation test (the challenge's official criterion): the known allosteric residues must score significantly higher than random distal background residues.

---

Results

Predicted blind from the apo structure, QASC is permutation-significant on all three Cleveland Clinic Global Quantum + AI Challenge 2026 targets:

Target	Pocket regime	Permutation p	Significant
KRAS G12C	cryptic Switch-II	0.006	✓
BCR-ABL1	buried Myristoyl	0.001	✓
Cardiac myosin	apo-collapsed mechanical	0.035	✓

NPERM = 20000, reproducible end-to-end with python examples/run_targets.py. Out-of-fold generalization across hundreds of held-out proteins, the quantum-versus-classical ablation, and autonomous active-site seeding are reported objectively in RESULTS.md.

---

Install

python -m venv .venv && source .venv/bin/activate
pip install -e .

Usage

Reproduce the three-target result:

python examples/run_targets.py

Score your own protein:

import numpy as np
from qasc.model import allosteric_scores, predict_top5

cb     = ...          # (N, 3) per-residue Cbeta coordinates
anchor = ...          # indices of the active-site residues

scores = allosteric_scores(cb, anchor)   # (N,) per-residue allosteric score
top5   = predict_top5(cb, anchor)        # 5 candidate allosteric residue indices

---

Repository layout

qasc/
  graph.py       contact graph, Kirchhoff Hamiltonian, distance-to-anchor
  quantum.py     CTQW communicability, IPR-resonant transfer
  model.py       rank percentile, noisy-or fusion, pocket-smoothing, top-5
  evaluate.py    permutation significance test
data/targets/    the three challenge targets (cb, anchor, y, resnums)
examples/        run_targets.py — reproduces the headline results

Data

Each data/targets/*.npz holds only the apo inputs and evaluation labels: cb (Cβ coordinates), anchor (active-site residue indices), y (ground-truth allosteric residues, used only for the permutation test), and resnums (author residue numbers). Structures are derived from public RCSB PDB entries.

Context

Built for the 2026 Global Quantum + AI Challenge, Cleveland Clinic track — "Unlocking undruggable targets: quantum simulation of allosteric signal propagation." This repository contains the final model only; the exploratory work behind it is not included.

License

MIT — see LICENSE.