Abstract
<jats:p>De novo molecular generation often collapses multiple design objectives into a single scalar reward, obscuring trade-offs between potency, synthetic accessibility and drug-likeness. We present a Paretoguided Monte Carlo tree search (MCTS) framework that maintains a non-dominated front across multiparameter optimisation (MPO), synthetic Bayesian accessibility (SYBA), resistance-resilience (RRS) and polypharmacology-network (PNS) scores, guided by a scaffold-aware fragment policy derived from ChEMBL27 frequencies and Tanimoto compatibility. A four-method benchmark across 20 independent seeds (fixed search budget) shows that random search achieves the highest mean reward (0.733 5), followed by MCTS (0.727 6), deterministic greedy search (0.721 1) and a genetic algorithm (0.702 7); the MCTS–random gap is small (mean ∆ = 0.005 9, 95% CI [0.000 8, 0.011 0], paired t19 = 2.41, p = 0.026). MCTS is therefore competitive on scalar reward and, within our fragment-vocabulary oracle design, is the method that returns a multi-objective Pareto front of non-dominated candidates (four solutions, hypervolume 1.236 6) spanning the potency–accessibility trade-off, including the high-potency / low-accessibility extreme that fixed-weight scalar aggregation discards. To our knowledge this is the first Pareto-guided MCTS to fold resistance-resilience and polypharmacology-network scores into the generation objective, in contrast to generic affinity/drug-likeness metrics in prior Pareto-MCTS generators, with the honest caveat that in the present front these domain scores remain comparatively homogeneous. Component ablations identify the scaffold-aware policy (∆ = +0.148) and the Pareto front itself (∆ = +0.108) as the dominant drivers of performance. All code, molecule sets and score tables are deposited under an open licence. These results position MCTS as a complementary multi-objective explorer rather than a universal single-objective optimiser. Scientific Contribution: We introduce a Pareto-guided MCTS framework for de novo antimalarial design that maintains a non-dominated front across potency, synthetic accessibility, resistance-resilience and polypharmacology-network scores, guided by a scaffold-aware fragment policy. Against a four-method benchmark across 20 seeds, MCTS is competitive with random search on scalar reward while uniquely recovering a multi-objective Pareto front spanning the potency–accessibility trade-off. Component ablations identify the scaffold-aware policy and the Pareto front itself as the dominant drivers of performance.</jats:p>