Short MD simulations coupled with a data-driven rank-aggregation score for virtual screening: ranking accuracy, activity-relevant features, and domain of applicability

Make-on-Demand (MoD) virtual screening provides access to billions of synthesizable compounds, but each candidate must be synthesized before it can be tested, changing the objective from the raw hit rate to ranking accuracy of top candidates. We show that 15 ns of molecular dynamics (3$\times$5 ns replicas) is sufficient to meaningfully improve this ranking: an interpretable, data-driven rank-aggregation model built on dynamics-derived contact features consistently outranks docking on two structurally distinct targets ($\rho$=0.65 vs. 0.49), a result that holds across hundreds of alternative train/test splits. On a target where docking is uninformative or even anti-correlated with potency, MD alone recovers a clear ranking signal ($\rho$=0.46 vs. $-$0.15), showing the orthogonality of the approach. With a couple of fitted parameters, i.e., feature signs, the model resists the overfitting that degrades parametric ML models on the same features (random forests and gradient-boosted trees reach only $\rho$=0.39-0.44), and remains usable with training sets as small as 30-50 compounds. MD adds the most value precisely where similarity-based ranking is weakest: in the moderate-novelty regime and when training data are scarce or dominated by weak or inactive binders. A label-free variant requiring no activity labels already ranks compounds well from inactive-only training, enabling potency prediction before any actives are confirmed. Neither MD nor the docking baseline resolves congeneric pairs (Tc$\geq$0.7), marking the boundary where higher-fidelity methods such as FEP are likely needed. The short simulation time makes the method practical for scoring hundreds to thousands of MoD candidates on a single GPU node within days.