DESPOT

A carbonyl oxygen approached head-on along a lone pair and the same oxygen approached edge-on are, to a classical knowledge-based potential, the same event: both are recorded as an oxygen–donor pair at 2.9 Å. Conventional statistical scoring functions tabulate the probability of an interaction as a function of distance alone, which treats the space around every protein atom as isotropic and discards precisely the angular information that makes hydrogen bonds, halogen bonds and aromatic stacking directional. DESPOT (Direction-Enhanced Scoring POTentials) is an all-atom anisotropic potential built to keep that information.

It does so by inverting the usual formulation. Instead of asking how interactions are distributed through space, DESPOT asks, for every discretized point in a symmetry-aware frame around a protein atom, which ligand atom type occupies that point — with a dedicated "void" type standing for space that is preferentially left empty. Because the reference state is positionally averaged rather than interaction-conditioned, the resulting potential is ligand-independent and encodes steric exclusion directly, something a distance-only reference state cannot express. The paper pairs this atom-level model with a geometry-conditioned residue-level formulation, DESPOT-screen, and combines the two inverse-Boltzmann scores into a consensus, DESPOT-combo.

DESPOT was developed in the Laboratory for Biomolecular Modelling and Design at KU Leuven, across its Chemistry and Electrical Engineering departments, and released as open-source software with its derived potentials archived separately. It sits beside catalog scoring functions such as EquiScore, BioScore and DOScore, but takes the opposite methodological route: interpretable tabulated statistics rather than a learned neural representation. The name is unrelated to DESPOT1/DESPOT2, the long-established quantitative-MRI relaxometry technique.

#Key Features

• Symmetry-aware discretization: Atoms are classified from hybridization and heavy-atom coordination into isotropic, axially symmetric and fully anisotropic classes, and the surrounding space is binned in radial shells, spherical sectors or full voxels accordingly.
• A void atom type: Treating emptiness as an outcome of the occupancy distribution lets the potential penalize steric clash natively, without a bolted-on repulsion term.
• Three related scoring modes: The atom-level score, the residue-level DESPOT-screen, and their DESPOT-combo consensus are derived from the same statistics and the same atom typing, so their differences isolate the formulation rather than the data.
• Interaction fields from the scoring model: Evaluating the same potentials over an empty grid produces multi-channel molecular interaction fields, one channel per ligand atom type, for binding-site characterization and fragment placement.
• Per-atom decomposition: Because the score is fully additive, each ligand atom carries its own contribution, exposing which substituents are well placed and which are not.

#Technical Details

Atom typing follows the fconv scheme used by DrugScoreX, refined by first- and second-degree heavy-atom bonding environments and co-defined with symmetry class, yielding 67 protein and 145 ligand atom types. Interactions are recorded from 1–6 Å in 0.1 Å radial bins, with angular bins at 3.0° laid out on a Driscoll–Healy equiangular grid so each shell can be smoothed analytically as a spherical-harmonic filter rather than by real-space convolution; pseudo-energies from the inverse-Boltzmann step are capped to [-5, +5] so no single sparse bin dominates. The potentials are derived from 110,943 drug-like complexes drawn from CROWN, a curated set of quality-filtered, protonated, energy-minimized PDB complexes, after removing any entry sharing both >70% sequence identity and >70% ECFP4 Tanimoto similarity with CASF-2016 and capping representation at 500 complexes per ligand code. On CASF-2016 the atom-level score reaches Pearson r = 0.61 scoring power and Spearman ρ = 0.639 ranking power, while DESPOT-combo recovers a near-native pose at rank one for 255 of 285 targets (89.5%) and improves enrichment over every isotropic potential tested.

#Applications

The potentials are applied unchanged to new complexes, so the practical uses are rescoring docked poses, ranking compounds in a virtual screen, and characterizing a pocket before any ligand exists. Medicinal chemists get a per-atom readout that points at which functional group is paying a geometric penalty, which is more actionable during lead optimization than a single aggregate score; structural biologists get interaction-field channels derived from observed complexes rather than from force-field probe calculations.

#Impact

DESPOT's substantive claim is that anisotropy matters most where distance-only scoring is weakest — rejecting geometrically implausible poses and separating binders from decoys — while gains in affinity correlation are modest, an asymmetry the paper's controlled ablations against reimplemented isotropic and residue-level baselines make explicit. Its leakage analysis is a second contribution: the authors argue that overfitting is underexamined in knowledge-based scoring, and that scoring functions whose training data overlap CASF-2016 post inflated numbers. The work is a preprint and has not been peer reviewed, and the reported evaluation is confined to CASF-2016 rather than prospective screening, so the screening advantage remains a retrospective one.