Adaptive acquisition functions (UCB-guided Bayesian optimization) will reduce the number of experimental trials required to identify coalition-stable drug synergy pairs by ≥40% compared to fixed EI acquisition, when applied to residual graph isomorphism networks trained on MSH3-KPC-3 interaction graphs.

When a residual graph isomorphism network (R-GIN) trained on MSH3–KPC-3 protein-protein/drug interaction graphs is used as a surrogate model for Bayesian optimization over a drug-pair combination space, an Upper Confidence Bound (UCB) acquisition function will identify all coalition-stable drug synergy pairs (defined as pairs whose Shapley-value-derived synergy score exceeds a fixed threshold τ and remains stable under leave-one-out perturbation of the coalition) using ≥40% fewer labeled experimental trials (oracle queries) than a fixed-parameter Expected Improvement (EI) acquisition function, measured as the number of queries needed to reach 95% recall of the true stable-pair set, averaged over ≥30 independent runs with different random seeds and initial designs, at statistical significance p<0.05 (paired Wilcoxon signed-rank test) and effect size Cohen's d≥0.5.

Minimum viable test (MVT): simulate a closed-loop BO campaign against a frozen "oracle" — the fully trained R-GIN surrogate evaluated on held-out ground-truth synergy labels — with two acquisition functions (UCB vs EI) competing to find all coalition-stable pairs from a fixed candidate pool of ~1,000 drug pairs derived from MSH3-KPC-3 interaction data. Each acquisition function runs 30 independent trials (different random seeds, different initial 10-point Latin hypercube designs). Track cumulative queries to reach 95% recall of ground-truth stable set. Compare query-count distributions.

# Phase 1: Surrogate training
graph = build_MSH3_KPC3_interaction_graph(sources=[STRING, DrugBank])
model = ResidualGIN(layers=5, hidden=256, dropout=0.2)
train(model, labeled_pairs_70pct, epochs=200, early_stop_patience=15)
assert calibration_ECE(model, held_out_30pct) < 0.10
assert AUROC(model, held_out_30pct) >= 0.75

# Phase 2: Ground truth coalition-stability labeling
for pair in candidate_pool:
    shapley_score = monte_carlo_shapley(pair, coalition_graph, n_perm=1000)
    stability = leave_one_out_variance(pair, coalition_graph) < 0.10 * shapley_score
    ground_truth[pair] = (shapley_score > tau) and stability

# Phase 3: BO loop (per acquisition function, per seed)
def bo_loop(acq_fn, seed, pool, ground_truth, budget=300):
    rng = seed_rng(seed)
    observed = latin_hypercube_init(pool, n=10, rng=rng)
    gp = fit_GP_on_embeddings(model.embed(observed))
    queries = 10
    while recall(observed, ground_truth) < 0.95 and queries < budget:
        next_pair = acq_fn.select(gp, pool - observed)   # UCB: mu + beta*sigma
        label = ground_truth[next_pair]                   # oracle reveal
        observed.add((next_pair, label))
        gp = fit_GP_on_embeddings(model.embed(observed))   # or update posterior
        queries += 1
    return queries

results_UCB = [bo_loop(UCB(beta=2), s, pool, gt) for s in range(30)]
results_EI  = [bo_loop(EI(), s, pool, gt) for s in range(30)]

reduction_pct = 100 * (mean(results_EI) - mean(results_UCB)) / mean(results_EI)
wilcoxon_test(results_UCB, results_EI)