Documentation

Docs / Known Limitations

Known Limitations

No scoring function is perfect. This page documents the current limitations of LatticeZero's docking and scoring tools. We believe honest disclosure builds more trust than marketing claims.

Rigid Receptor

LatticeZero treats the receptor as a rigid body. The scoring grid is pre-computed from a single protein conformation and does not change during docking or scoring.

Impact: Targets with significant induced-fit effects (loop rearrangements, side-chain rotations upon ligand binding) may show reduced scoring accuracy. Flexible residues near the binding site are not modeled.

Workaround: For targets where flexibility is critical, consider preparing multiple receptor conformations (ensemble docking) and scoring against each grid separately. Compare rankings across conformations to identify robust hits.

No Quantum Mechanical Treatment

The scoring function uses classical physics approximations - pairwise potentials, Coulombic electrostatics with a distance-dependent dielectric, and empirical hydrogen bond terms. No quantum mechanical (QM) calculations are performed.

Impact: Charge transfer, polarization, and orbital-mediated interactions (e.g., sigma-holes in halogen bonds) are approximated rather than computed from first principles. The halogen bond term (E_halogen) uses a geometric model rather than QM electrostatics.

Where this matters most: Metalloprotein active sites where metal coordination involves d-orbital effects, and halogen-bonded complexes where sigma-hole directionality is critical.

No Explicit Solvent

Water molecules are not explicitly simulated. Desolvation effects are approximated through the E_desolv term (based on atomic solvation parameters and buried surface area) rather than through molecular dynamics or integral equation methods.

Impact: Water-mediated interactions - where a water molecule bridges the ligand and protein - are not directly modeled. The waterAnchorFrac geometric feature provides a proxy for conserved water contacts, but does not compute actual water positions or energies.

Where this matters most: Targets with structurally conserved water molecules in the binding site (e.g., HIV protease catalytic waters, metalloenzyme coordination waters).

No Entropy Estimation

The scoring function does not estimate the entropy change upon binding. The strain term captures internal ligand strain (conformational cost) but not translational, rotational, or vibrational entropy losses, nor the entropy gain from displacing ordered water.

Impact: The scores correlate with binding favorability but should not be interpreted as free energy estimates. LatticeZero predicts relative rankings (which ligand binds better), not absolute binding free energies (ΔG).

Ligand Flexibility

IsoPose searches ligand conformational space using a genetic algorithm with torsional sampling. While effective for drug-like molecules with moderate flexibility (up to ~10 rotatable bonds), the search may be incomplete for:

Highly flexible ligands (>12 rotatable bonds) - GA may not adequately sample all relevant conformations
Macrocycles - Ring conformations require specialized sampling not currently implemented
Peptides - Backbone flexibility and secondary structure preferences are not modeled

Practical guidance: For highly flexible compounds, consider generating conformer ensembles externally (e.g., with RDKit ETKDG) and using IsoScore to rescore them.

No Covalent Docking

LatticeZero models non-covalent binding only. Covalent inhibitors - compounds that form a chemical bond with a protein residue (typically Cys, Ser, or Lys) - are scored based on their non-covalent interactions only.

Impact: For known covalent inhibitors, the scoring will capture the non-covalent pose recognition but miss the covalent bond contribution. This typically results in underscoring of covalent binders relative to non-covalent binders.

Metalloprotein Limitations

While the E_metal term models metal ion coordination, the current implementation has limitations:

Supported metals: Zn, Mg, Mn, Fe, Ca, Cu. Other metals are not specifically parameterized.
Coordination geometry: Idealized geometric preferences (tetrahedral for Zn, octahedral for Mg) are used. Unusual coordination geometries may be misjudged.
Metal-binding pharmacophores: Only common metal-binding groups (carboxylates, hydroxamates, thiols, phosphonates) are recognized. Novel metal-chelating motifs may not score correctly.

Absolute Binding Affinity

LatticeZero scores are designed for ranking (which compound in a library is most likely to bind), not for predicting absolute binding affinities (ΔG or Ki/IC50 values). The total score is a weighted sum of heterogeneous physics terms and does not have direct thermodynamic units.

Practical implication: A score of -50 does not mean ΔG = -50 kcal/mol. Scores are meaningful for relative comparison within a single target and profile, not across different targets or profiles.

Benchmark Scope

Current validation is primarily on the DEKOIS2 benchmark suite, which has its own limitations:

Chemical diversity: DEKOIS2 targets represent common drug target classes but do not cover all protein families.
Decoy quality: Matched decoys test discrimination against similar-property non-binders. Real-world screening libraries have different property distributions.
Pre-docked poses: IsoScore benchmarks use pre-positioned conformers, not de novo docked poses.

See Benchmark Results for our full validation data including confidence intervals and stability assessments.

Reporting Issues

If you encounter unexpected scoring behavior or results that seem inconsistent with the known physics of your target, we encourage you to report it. Understanding failure modes helps us improve the scoring function.