Glossary

Key terms, metrics, and definitions used throughout LatticeZero.

Active (compound)

A molecule experimentally confirmed to bind to the target protein. In benchmarking, actives are the "positives" that a scoring function should rank highly.

AUC (Area Under the ROC Curve)

A metric ranging from 0 to 1 that measures how well a scoring function separates actives from decoys. AUC = 1.0 means perfect separation; AUC = 0.5 means random (no better than chance). LatticeZero uses tie-corrected Mann-Whitney U AUC to prevent inflation from tied scores.

Autotune

LatticeZero's unsupervised profile generation tool. Given a receptor PDB and compound library SDF (no active/decoy labels), Autotune analyzes the binding pocket and interpolates between validated reference profiles to generate a custom scoring profile. See Autotune.

BEDROC (Boltzmann-Enhanced Discrimination of ROC)

An enrichment metric that weights early recognition more heavily than AUC. Useful when you care most about finding actives in the top 1-5% of a ranked list. Uses an exponential decay parameter alpha (default: 20).

Binding Pocket

The 3D cavity on a protein surface where ligands bind. In LatticeZero, the pocket is defined by a center point and radius during Target Prep. The scoring grid covers this region.

Bootstrap Confidence Interval

A statistical method for estimating uncertainty. LatticeZero resamples actives and decoys with replacement (typically 1,000 iterations) to compute a 95% confidence interval around AUC and other metrics.

Clash

A steric overlap between atoms. The E_clash scoring term penalizes ligand atoms that are too close to protein atoms. Moderate clashes may indicate pose prediction issues; severe clashes suggest the pose is physically unrealistic.

Cross-Validation (Holdout)

A validation technique where data is split into training and test sets. The Optimizer uses multi-seed 5-fold holdout: it trains on 80% of ligands and evaluates AUC on the held-out 20%, repeating 5 times. The mean holdout AUC is the primary quality metric.

Decoy

A molecule designed to resemble actives in physicochemical properties (molecular weight, logP, hydrogen bond count) but not bind to the target. Decoys are the "negatives" in benchmarking. DEKOIS2 provides carefully matched decoy sets.

DEKOIS2

Demanding Evaluation Kits for Objective In Silico Screening, version 2. A widely-used benchmark suite providing curated active/decoy sets for dozens of pharmaceutical targets. LatticeZero validates scoring profiles against DEKOIS2 targets. See Benchmark Results.

Desolvation

The energetic cost of removing water molecules from both the ligand and the protein surface upon binding. Represented by the E_desolv scoring term.

Differential Evolution (DE)

The global optimization algorithm used by the LatticeZero Optimizer. DE maintains a population of candidate solutions and evolves them through mutation, crossover, and selection. It's robust against local minima and doesn't require gradient information.

Dispersion

Attractive van der Waals interactions between atoms. The E_disp term captures London dispersion forces, which are distance-dependent and contribute favorably to binding. Dominant for hydrophobic contacts.

Docking

The computational process of predicting how a small molecule (ligand) binds to a protein target. IsoPose performs docking using a genetic algorithm to search for optimal poses.

Enrichment Factor (EF)

A metric measuring how many more actives are found in the top X% of a ranked list compared to random selection. EF1% = 20 means the top 1% contains 20x more actives than expected by chance.

Feature Matrix

In the Optimizer, a pre-computed N x F matrix where N = number of ligands and F = number of scoring features. Used for fast vectorized scoring: scores = feature_matrix @ weight_vector.

Genetic Algorithm (GA)

The pose search method used by IsoPose. A population of candidate poses evolves through selection, crossover, and mutation until convergence. Parameters include population size, mutation rate, and maximum generations.

LZ Pack

A pre-validated target bundle containing a compiled scoring grid, optimized profile, reference compounds, and benchmark results. LZ Packs provide production-ready scoring for well-studied targets. See LZ Packs.

Grid (Scoring Grid)

A 3D lattice of pre-computed interaction energies covering the binding pocket. During scoring, each ligand atom's contribution is looked up from the grid, enabling ultra-fast evaluation. Compiled during Target Prep.

Hydrogen Bond

A directional electrostatic interaction between a hydrogen donor (N-H, O-H) and an acceptor (lone pair on N, O, S). The E_hbond term scores these interactions. The hbondGeo term evaluates the geometric quality (angle and distance) of hydrogen bonds.

IsoPose

LatticeZero's GPU-accelerated molecular docking tool. Performs conformational search using a genetic algorithm followed by scoring against the pre-compiled grid. Processes approximately 1 ligand every 3 seconds. See IsoPose.

IsoGrid

The compiled scoring grid data structure generated during Target Prep. Contains pre-computed interaction energies for each grid point across all 14 scoring terms. Stored as a binary file and loaded into GPU memory for scoring.

IsoScore

LatticeZero's ultra-fast rescoring engine. Takes pre-docked 3D coordinates as input (no pose search) and evaluates them against the scoring grid at ~4,000 ligands per second using WebGPU. See IsoScore.

L2 Regularization

A penalty added to the Optimizer's objective function that discourages extreme weights: lambda * ||w||^2. Prevents overfitting and keeps profiles physically interpretable.

Ligand

A small molecule that binds to a protein target. In molecular docking, ligands are the molecules being scored against the receptor.

Metalloprotease

A protease that uses a metal ion (typically zinc) in its catalytic mechanism. ACE is the canonical example in LatticeZero. Metal coordination scoring (E_metal) is important for these targets.

Pose

A specific 3D orientation and conformation of a ligand within the binding pocket. Docking generates poses; scoring evaluates them.

Pose Quality Score (PQS)

A classification system that assesses whether a docked pose is physically plausible. Each pose receives a verdict: GOOD (plausible), FIXABLE (minor issues), or TRASH (physically unrealistic). PQS is computed from E_clash, E_rep, insideFrac, and strain. See IsoScore Methodology.

Profile (Scoring Profile)

A set of weights (one per scoring term) that control how each physics term contributes to the total score. Different profiles are optimized for different target classes. See Scoring Profiles.

Profile Pack

A standardized export format (v2.0) that bundles a scoring profile with its validation metadata: weights, AUC, tier, validation method, dataset, engine version, and provenance information. Used for sharing profiles between sessions and with the CLI tools.

Receptor

The protein target that ligands bind to. Used interchangeably with "target" in the context of molecular docking.

Repulsion

Short-range steric clash energy. The E_rep term penalizes atoms that are closer than the sum of their van der Waals radii. High repulsion indicates steric clashes in the pose.

Rescoring

Evaluating pre-existing ligand poses against a scoring function without performing new pose search. IsoScore is a rescoring engine - it takes 3D poses as input and returns scores.

ROC Curve (Receiver Operating Characteristic)

A plot of true positive rate (sensitivity) vs. false positive rate (1 - specificity) at varying score thresholds. The area under this curve (AUC) quantifies overall discrimination ability.

Scoring Function

The mathematical function that evaluates how well a ligand pose fits the target binding site. LatticeZero's scoring function has 14 physics-based terms. The total score is a weighted sum of these terms. See Physics Reference.

SDF (Structure-Data File)

A file format for storing molecular structures with optional data fields. LatticeZero accepts SDF files as input for scoring and can export results in SDF format with scores in molecule properties.

Strain

Internal ligand strain energy - the energetic cost of the ligand adopting its bound conformation relative to a relaxed state. High strain suggests the ligand must distort to fit the pocket, which is thermodynamically unfavorable.

Target Class

A classification of protein targets by structural and functional similarity (e.g., kinase, protease, nuclear receptor, GPCR, metalloenzyme, PPI, enzyme). Target class guides the selection of scoring profiles, Optimizer priors, and Autotune interpolation.

Target Prep

LatticeZero's 4-step receptor preparation pipeline: upload PDB, clean structure (remove waters, add hydrogens), define binding pocket (center + radius), and compile the scoring grid (IsoGrid). See Target Prep.

Tier

A quality classification for scoring profiles based on validated AUC: Platinum (>= 0.90), Gold (>= 0.80), Silver (>= 0.60), Internal (< 0.60). Higher tiers indicate stronger validated discrimination between actives and decoys.

WebGPU

A modern W3C web standard for GPU computation in browsers. Provides low-level access to the GPU for both graphics and compute workloads. LatticeZero uses WebGPU compute shaders (written in WGSL) for all scoring computations, enabling browser-based molecular docking without software installation.

WGSL (WebGPU Shading Language)

The shader programming language for WebGPU, defined by the W3C. LatticeZero's scoring kernels - the inner loops that evaluate each ligand atom against the scoring grid - are written in WGSL and execute directly on the user's GPU.

Weight (Profile Weight)

A numeric multiplier applied to a scoring term within a profile. Higher absolute weight = greater influence on the total score. Weights can be negative (indicating the feature separates actives and decoys in the opposite direction from the default assumption). Weights are optimized by the Optimizer or generated by Autotune.

Y-Scramble

A validation control where active/decoy labels are randomly permuted before profile optimization. If the resulting AUC is near 0.5 (random), the model is confirmed to be learning real signal. If Y-scramble AUC is high, the model may be overfitting to noise rather than capturing genuine binding physics.

Adaptive Budget Escalation

A server-side docking strategy that starts with a fast "scout" search and only escalates to deeper searches when needed. Easy ligands finish quickly (~0.3s), while difficult cases automatically receive more compute time (up to ~90s). Reduces total screening time without sacrificing quality on hard cases.

Auto-Policy Routing

An IsoPose feature that automatically selects the best GA search strategy based on pocket properties. Analyzes volume, charge distribution, buriedness, and metal presence to choose between baseline, aggressive, conservative, and exploratory policies.

Chemical State Enumeration

Generating multiple plausible protonation states and tautomers of a ligand before docking. Different charge states can dramatically change binding, so docking all plausible forms and keeping the best result improves accuracy for ionizable molecules.

Conformer Ensemble

Generating multiple diverse 3D conformations of a ligand (using RDKit ETKDGv3) and docking each independently. Returns the best pose across all conformers. Breaks the "steric veto ceiling" where the input conformation prevents correct pose prediction.

Early Stopping

A GA optimization technique that terminates the search when improvement stalls. If the best score hasn't improved for 15 consecutive generations (after 20 minimum), the search ends early, saving 20-40% of compute time.

PoseRepair

A post-GA clash resolution step that applies small random perturbations to top poses to reduce steric overlap with the receptor. Improves >95% of docked poses. Enabled by default.

TopK Poses

Returning the K best poses per ligand instead of just one. Poses are diversity-filtered by RMSD clustering to ensure distinct binding modes. Higher presets return more poses (Balanced: 3, High: 5).

Pocket Distance The Euclidean distance from the defined pocket center to the centroid (geometric center) of a docked pose. Values under the pocket radius (typically 8–15 Å) indicate the pose is correctly placed in the binding site. Available as pocket_distance in docking results.

Input Displacement The distance a ligand moved from its input 3D coordinates to the final docked position. When input comes from SMILES (coordinates near [0,0,0]), this value is typically 30–80 Å, reflecting the distance from the origin to the binding pocket - not an error. Also known as pose0_centroid_distance. See Pocket Distance for the correct metric to verify pose placement.