Chapter 9: Structure-based Drug Design with Active Learning
Introduces structure-based drug design, explains why docking at massive scale motivates progressive docking, and illustrates the approach with a validated hit-finding example.
After this chapter you will be able to
- Structure-Based Design and Docking. Explain how knowledge of a target’s 3D structure guides molecule selection and how docking helps rank potential binders.
- Progressive Docking Cost Reduction. Explain why exhaustive docking is costly at billion-molecule scale and describe how progressive docking substitutes predictions for many docking runs.
- SBDD and SBVS relationship. Distinguish the broad scope of structure-based drug design from structure-based virtual screening as one early-stage application.
- Protein Conformational Flexibility. Explain why a protein should not be treated as a single rigid shape when considering ligand binding.
- Binding Sites and Docking. Explain what protein-ligand docking seeks to predict and why a protein binding site can favor particular ligands.
- Docking Scores and Checks. Distinguish a docking score from measured binding strength and describe how a promising pose is checked before advancing a compound.
- Conformational sampling of poses. Explain that conformational sampling searches possible protein–ligand geometries, or poses, within the protein’s binding site, and describe why changing conformations make the search difficult.
- Pose scoring and ranking. Explain that docking scores poses to identify a compound’s most likely binding mode and prioritize compounds, with the lowest, most favorable score as its target.
- Selective Protein–Ligand Isolation. Trace how the complex is separated into protein and ligand PDB files, including how the ligand is identified. Explain why selecting every non-protein, non-water atom can give the wrong ligand file.
- Preparation Choices and Docking Reliability. Explain why separate PDB files are not automatically ready for reliable docking. Describe representative protein and ligand preparation decisions and how those decisions can affect results.
- Binding Box Definition. Explain how the known ligand position is used to define a binding box and why its size matters.
- Vina Pose Scoring. Describe what Vina's scoring function evaluates and its stated limitation.
- Pose Search and Selection. Trace Vina's repeated pose search from random starts through refinement and reported poses.
- Pose Ranking by Affinity. Explain how the docking poses are ordered and which result is used to rank them.
- Docking Pose Validation. Explain why a favorable docking score needs further checking and describe the validation options given here.
- On-demand libraries and docking cost. Explain what synthesis-on-demand libraries offer and why their scale makes docking every compound impractical.
- Deep docking candidate prioritization. Explain how deep docking reduces the need to dock every compound in a large library.
- Active learning sample selection. Distinguish the active-learning approach described here from training once on a predefined dataset.
- Limited Testing Budget. Explain why active learning is useful when there are far more candidates than can be tested or labeled, and why random selection can waste a limited budget.
- Surrogate And Oracle Loop. Trace how a small labeled set, a surrogate model, and an oracle work together in the repeated active learning process.
- Deep docking surrogate network. Describe how the PyTorch surrogate uses molecular fingerprints to predict docking scores, and explain the roles of its main components.
- Gaussian process surrogate tradeoffs. Explain why a Gaussian process can be useful early in active learning and why a project might later switch to a deep-learning surrogate.
- Surrogate Training Objective. Explain that the surrogate is trained with supervised learning to predict binding affinities, minimizing mean squared error between predictions and actual values with the Adam optimizer.
- Training and Experiment Monitoring. Explain why the section uses TensorBoard logging during training and the active learning experiment, including the kinds of changes and comparisons it is intended to track.
- Initial sampling cold start. Explain why an initial sample is needed before a surrogate model can guide compound selection, and why those first evaluations matter.
- Random and diverse selection. Distinguish random sampling from diversity-based sampling, including why random draws can be redundant in an imbalanced chemical library.
- Uncertainty-Based Molecule Selection. Explain why active learning selects molecules the model is uncertain about and how that uncertainty can be identified.
- Representative Chemical Space Coverage. Explain why selected molecules should cover different parts of chemical space, and describe ways to encourage that coverage.
- Exploration and Exploitation Trade-Off. Distinguish exploring new molecular possibilities from exploiting promising predicted regions, and explain why acquisition-function choice affects the balance.
- Four Acquisition Function Priorities. Distinguish greedy acquisition, uncertainty sampling, probability of improvement, and expected improvement by what each prioritizes.
- Docking Oracle Labels. Explain how an oracle turns an unlabeled molecule into a labeled one in this active learning application, and why evaluating every molecule is impractical.
- Precomputed Proxy Oracle. Describe why deepdock_oracle is used during development, what data it requires, and what changes when moving to real docking.
- Iterative Deep Docking Loop. Trace how an initial docked sample becomes training data for a surrogate model, then how repeated selection, docking, retraining, and evaluation continue until the budget is exhausted.
- Systematic Strategy Comparisons. Explain how organized experiments and parameter sweeps help compare active learning configurations, and describe what the section's example comparison shows.
- Active Learning Suitability. Explain the conditions under which active learning is worth its overhead and how the choice of oracle limits its value.
- Failure Modes and Validation. Describe plausible ways the loop can be misled and the checks needed before trusting or acting on its rankings.
- Use-Dependent Oracle Choice. Explain why docking is not necessarily the best or only oracle for active learning in drug discovery.
- Lead Optimization Aims. Describe when lead optimization begins and what it aims to improve about promising hits.
- Free Energy for Potency Optimization. Explain how free energy calculations estimate ligand binding affinity and why their predictions can guide potency optimization.
- FEP Ligand Transformation. Describe how FEP uses a transformation between ligands and molecular dynamics simulations to estimate binding free energy.
- Active Learning Protocol Selection. Trace how active learning selects and validates an FEP protocol using experimental binding data, FEP results, and a surrogate model.
What it builds on
- What is Virtual Screening? (chapter 2) · Introduces virtual screening of compound libraries.
- Understanding the Dataset as a Structure Data File (chapter 2) · Prior work with molecular structure files provides context for extracting structures from a PDB file.
- Implementing a Neural Network (chapter 8) · The earlier introduction to implementing a neural network provides the basis for this PyTorch model.
- Training & Applying our Neural Network (chapter 8) · Supervised neural-network training techniques.
- Fingerprints: Representing Molecules as Numbers (chapter 2) · Molecular fingerprints used by RDKit's MaxMinPicker.
- Lead Identification & Lead Optimization (chapter 1) · The distinction between identifying hits and optimizing leads.
Section by section
Introduction
Introduces structure-based drug design, explains why docking at massive scale motivates progressive docking, and illustrates the approach with a validated hit-finding example.
- Structure-Based Design and Docking: Explain how knowledge of a target’s 3D structure guides molecule selection and how docking helps rank potential binders.
Why it matters: This explains both what docking contributes to hit identification and why its rankings need further evaluation. - Progressive Docking Cost Reduction: Explain why exhaustive docking is costly at billion-molecule scale and describe how progressive docking substitutes predictions for many docking runs.
Why it matters: It shows how a screening campaign can prioritize expensive calculations when docking every candidate is impractical. - CACHE Hit-Finding Example: Describe how the CACHE-1 example connects computational screening to experimental testing, and identify Deep Docking’s role in narrowing the candidate pool.
Why it matters: The example distinguishes computationally predicted hits from binders confirmed by experiments.
Key terms: structure-based drug design (SBDD), molecular docking, scoring-function values, progressive docking, docking-score proxies, deep learning (DL), CACHE-1, Deep Docking, predicted hits, experimentally confirmed binders
9.1 Docking: A Core SBDD Technique
Defines structure-based drug design and places structure-based virtual screening within its early stages.
- SBDD and SBVS relationship: Distinguish the broad scope of structure-based drug design from structure-based virtual screening as one early-stage application.
Why it matters: This distinction places computational screening within the broader effort to find and improve molecules using target structures.
Key terms: structure-based drug design (SBDD), structure-based virtual screening (SBVS), 3D structure
9.1.1 Protein-Ligand Docking
This section explains what protein-ligand docking predicts, why protein structures complicate it, and how to interpret and check docking results.
- Protein Conformational Flexibility: Explain why a protein should not be treated as a single rigid shape when considering ligand binding.
Why it matters: Protein flexibility is a reason a docking result may not capture every relevant binding arrangement. - Experimental Structure Uncertainty: Describe why an experimentally determined protein structure is a useful but uncertain starting point for docking.
Why it matters: Uncertainty in the starting structure can affect conclusions drawn from docking. - Binding Sites and Docking: Explain what protein-ligand docking seeks to predict and why a protein binding site can favor particular ligands.
Why it matters: It connects a docking pose to the physical and chemical features of the target site. - Docking Scores and Checks: Distinguish a docking score from measured binding strength and describe how a promising pose is checked before advancing a compound.
Why it matters: A favorable score can help prioritize compounds, but it cannot establish binding on its own.
Key terms: protein conformations, energy landscape, rotatable bonds, X-ray crystallography, electron-density map, atomic model, protein binding sites, protein-ligand docking, binding mode, active site
9.1.2 Minimal Protein-Ligand Docking Workflow
This section outlines the stages of a minimal docking workflow and explains how docking searches for and ranks possible protein–ligand binding poses.
- Three docking workflow stages: State that the molecular docking workflow is divided into input preparation, structure preparation, and docking.
Why it matters: These stages give you a simple way to organize the docking process. - Conformational sampling of poses: Explain that conformational sampling searches possible protein–ligand geometries, or poses, within the protein’s binding site, and describe why changing conformations make the search difficult.
Why it matters: Sampling explains what docking must search before it can choose a binding pose. - Pose scoring and ranking: Explain that docking scores poses to identify a compound’s most likely binding mode and prioritize compounds, with the lowest, most favorable score as its target.
Why it matters: Scoring turns candidate poses into a predicted binding mode and a way to prioritize compounds.
Key terms: input preparation, structure preparation, docking, conformational sampling, poses, binding site, pose evaluation, scoring functions, binding mode, lowest (most favorable) score
9.1.3 Prepare the Protein & Ligand Structures
This section uses a crystallized EGFR–LN2057 complex to show how to isolate docking inputs and why careful protein, ligand, and binding-site preparation matters.
- Docking Inputs and Binding Site: Describe the protein structure, ligand structure, and binding-site definition needed for docking, using the EGFR–LN2057 complex as the example. Explain how the known site differs from the search area in blind docking.
Why it matters: The structures and search region determine what the docking experiment evaluates. - Selective Protein–Ligand Isolation: Trace how the complex is separated into protein and ligand PDB files, including how the ligand is identified. Explain why selecting every non-protein, non-water atom can give the wrong ligand file.
Why it matters: A selection can run without error yet put the wrong atoms in the ligand file and misplace the docking box. - Preparation Choices and Docking Reliability: Explain why separate PDB files are not automatically ready for reliable docking. Describe representative protein and ligand preparation decisions and how those decisions can affect results.
Why it matters: A docking score is only as trustworthy as the structural and chemical choices behind its inputs.
Key terms: protein, ligand, binding site, blind docking, PDB, MDTraj, Trajectory, atom indices, resname QQJ, receptor
9.1.4 Run A Docking Experiment
This section shows how to define a binding box and run AutoDock Vina, then explains how Vina scores, searches for, and reports ligand poses.
- Binding Box Definition: Explain how the known ligand position is used to define a binding box and why its size matters.
Why it matters: The box determines where docking searches and can affect both the results and computation time. - Charge Preparation and Vina Scores: Distinguish the effect of computed Gasteiger charges on Vina scores from the continuing importance of structure preparation.
Why it matters: It prevents the mistaken conclusion that charge-related structure preparation can be skipped just because Vina ignores these particular charges. - Docking Inputs and Interaction Maps: Trace how the docking call uses prepared structures, settings, and precalculated interaction maps.
Why it matters: The maps explain how Vina can look up interaction energies during a search and why reusing them can avoid redundant work. - Vina Pose Scoring: Describe what Vina's scoring function evaluates and its stated limitation.
Why it matters: A docking score helps rank predicted poses, but its empirical basis is a reason to treat predictions cautiously. - Pose Search and Selection: Trace Vina's repeated pose search from random starts through refinement and reported poses.
Why it matters: It explains why docking produces a set of predicted poses rather than a single placement.
Key terms: binding box, known binding site, MDTraj, Gasteiger partial charges, protonation, atom typing, hydrogen placement, PDBQT, VinaDocking, compute_vina_maps
9.1.5 Docking Validation & Interaction Fingerprints
This section explains how to interpret and validate docking poses and scores without treating a favorable score as proof of binding.
- Pose Ranking by Affinity: Explain how the docking poses are ordered and which result is used to rank them.
Why it matters: It helps you identify the top-ranked pose without confusing the docking score with its diagnostic components. - Diagnostic Energy Components: Describe what the additional energy outputs indicate and how they help flag questionable poses.
Why it matters: The extra columns can reveal strain that is worth investigating in an otherwise promising pose. - Docking Pose Validation: Explain why a favorable docking score needs further checking and describe the validation options given here.
Why it matters: A strong docking score is a starting point for judging a pose, not a substitute for checking whether it is plausible or binds in reality.
Key terms: docking score, affinity column, vina scoring function, pose, intermolecular energy, internal energy, torsion energy, conformational strain, affinity thresholds, control
9.2 Active Learning for Hit Identification: Deep Docking
This section explains why enormous synthesis-on-demand libraries make exhaustive docking impractical and introduces deep docking as a machine-learning approach to prioritizing compounds through active learning.
- On-demand libraries and docking cost: Explain what synthesis-on-demand libraries offer and why their scale makes docking every compound impractical.
Why it matters: The size of these libraries creates the need for a more selective screening strategy. - Deep docking candidate prioritization: Explain how deep docking reduces the need to dock every compound in a large library.
Why it matters: Predictions help reserve costly computational work for candidates worth pursuing. - Active learning sample selection: Distinguish the active-learning approach described here from training once on a predefined dataset.
Why it matters: It explains why selective sampling is part of the deep-docking approach, not just a one-time prediction step.
Key terms: Synthesis on Demand (SoD), molecular docking, virtual catalogs, deep docking, binding affinities, active learning, informative samples, predefined dataset
9.2.1 Active Learning: Smart Choices with Limited Resources
This section explains how active learning uses a limited testing budget to select informative candidates and improve a model through repeated testing.
- Limited Testing Budget: Explain why active learning is useful when there are far more candidates than can be tested or labeled, and why random selection can waste a limited budget.
Why it matters: It explains why choosing what to test is itself an important part of drug discovery. - Informative Candidate Selection: Distinguish uncertainty-based from representativeness-based selection using the apple examples, and describe how diverse initial choices help begin the search.
Why it matters: It gives two concrete reasons to choose one untested candidate over another. - Surrogate And Oracle Loop: Trace how a small labeled set, a surrogate model, and an oracle work together in the repeated active learning process.
Why it matters: It connects strategic choices to a repeatable process that improves as new results arrive.
Key terms: active learning, unlabeled data, initial sampling, uncertainty-based selection, representativeness-based selection, labeled data, surrogate model, oracle, active learning loop
9.2.2 The Deep Learning Surrogate Model
This section defines the neural-network surrogate used in deep docking and explains when a Gaussian process may be a better surrogate.
- Deep docking surrogate network: Describe how the PyTorch surrogate uses molecular fingerprints to predict docking scores, and explain the roles of its main components.
Why it matters: The surrogate provides a faster prediction in place of obtaining every score from the expensive oracle. - Gaussian process surrogate tradeoffs: Explain why a Gaussian process can be useful early in active learning and why a project might later switch to a deep-learning surrogate.
Why it matters: The amount of labeled data affects which surrogate is practical at different stages of active learning.
Key terms: surrogate model, Morgan fingerprints, docking score predictions, ReLU, dropout, nn.Sequential, Gaussian processes, GPs, oracle, mean
9.2.3 Training the Surrogate Model
This section explains how the deep learning surrogate is trained and how its progress is monitored during active learning.
- Surrogate Training Objective: Explain that the surrogate is trained with supervised learning to predict binding affinities, minimizing mean squared error between predictions and actual values with the Adam optimizer.
Why it matters: This objective defines what the surrogate is being trained to predict. - Mini-Batches and Training Settings: Describe why training uses mini-batches and how the stated batch size and learning rate may be adjusted.
Why it matters: These settings make training practical without treating the defaults as universally optimal. - Training and Experiment Monitoring: Explain why the section uses TensorBoard logging during training and the active learning experiment, including the kinds of changes and comparisons it is intended to track.
Why it matters: Logging helps reveal whether the surrogate remains a useful proxy as the experiment progresses.
Key terms: surrogate model, supervised learning, mean squared error, binding affinities, Adam optimizer, mini-batches, batch size, learning rate, convergence, TensorBoard
9.2.4 Initial Sampling
The section explains how to choose molecules for the first docking evaluations when an active-learning surrogate model has no labeled data.
- Initial sampling cold start: Explain why an initial sample is needed before a surrogate model can guide compound selection, and why those first evaluations matter.
Why it matters: The first docking evaluations set the starting point for the active-learning process. - Random and diverse selection: Distinguish random sampling from diversity-based sampling, including why random draws can be redundant in an imbalanced chemical library.
Why it matters: The sampling choice affects how much useful variety each expensive docking evaluation contributes. - MaxMin and alternative seeds: Describe how MaxMin builds a diverse initial set and state the other initial-selection options named in the section.
Why it matters: These are concrete ways to choose a useful starting set before model-guided selection is possible.
Key terms: initial sampling, surrogate model, labeled data, structure-activity landscape, random sampling, diversity-based sampling, chemical space, scaffolds, MaxMin, MaxMinPicker
9.2.5 Acquisition Functions for Active Learning
This section explains how acquisition functions choose molecules for docking by weighing model uncertainty, chemical-space coverage, and predicted binding affinity.
- Uncertainty-Based Molecule Selection: Explain why active learning selects molecules the model is uncertain about and how that uncertainty can be identified.
Why it matters: Choosing difficult cases can make each docking result more useful for improving the model. - Representative Chemical Space Coverage: Explain why selected molecules should cover different parts of chemical space, and describe ways to encourage that coverage.
Why it matters: Broader sampling reduces the chance of overlooking novel active compounds. - Exploration and Exploitation Trade-Off: Distinguish exploring new molecular possibilities from exploiting promising predicted regions, and explain why acquisition-function choice affects the balance.
Why it matters: The selection strategy affects how efficiently docking finds promising molecules without missing new ones. - Four Acquisition Function Priorities: Distinguish greedy acquisition, uncertainty sampling, probability of improvement, and expected improvement by what each prioritizes.
Why it matters: These priorities explain why different acquisition functions can propose different docking batches. - Monte Carlo Dropout Uncertainty: Describe how Monte Carlo dropout supplies an uncertainty estimate for neural-network predictions in this section.
Why it matters: It gives acquisition functions an uncertainty estimate even when the neural network does not provide one directly.
Key terms: uncertainty sampling, decision boundary, ensemble disagreement, representativeness, diversity sampling, density-weighted methods, clustering-based approaches, chemical scaffolds, acquisition functions, exploration
9.2.6 The Oracle
This section explains how docking supplies labels to an active learning system, why a precomputed proxy is useful during development, and how oracles can vary in cost and accuracy.
- Docking Oracle Labels: Explain how an oracle turns an unlabeled molecule into a labeled one in this active learning application, and why evaluating every molecule is impractical.
Why it matters: The cost of obtaining docking labels is why the system must choose which molecules to evaluate. - Precomputed Proxy Oracle: Describe why deepdock_oracle is used during development, what data it requires, and what changes when moving to real docking.
Why it matters: The proxy makes development practical without confusing fast benchmark lookups with real docking calls. - Multi-Fidelity Oracle Choices: Distinguish low-, medium-, and high-fidelity oracles by their approximate accuracy and computational cost, and describe why an active learning system might choose among them.
Why it matters: Choosing evaluation rigor as well as molecules offers another way to manage limited computational resources.
Key terms: oracle, AutoDock Vina, docking score, benchmark label, deepdock_oracle, proxy oracle, vina_oracle, pre-computed results, deterministic, multi-fidelity oracle
9.2.7 The Active Learning Loop
This section connects deep docking components into an active learning loop, shows how to monitor and compare runs, and explains when its results are useful or unreliable.
- Iterative Deep Docking Loop: Trace how an initial docked sample becomes training data for a surrogate model, then how repeated selection, docking, retraining, and evaluation continue until the budget is exhausted.
Why it matters: This is how the pipeline directs a limited docking budget toward promising molecules. - Progress and Chemical Space Monitoring: Describe the metrics and visualizations used to follow an active learning run, including what the PCA chemical-space plot cannot reliably show here.
Why it matters: Progress plots can reveal problems during a run, but a reductive chemical-space plot should not be overinterpreted. - Systematic Strategy Comparisons: Explain how organized experiments and parameter sweeps help compare active learning configurations, and describe what the section's example comparison shows.
Why it matters: Comparable runs make conclusions about strategy choices more reliable. - Active Learning Suitability: Explain the conditions under which active learning is worth its overhead and how the choice of oracle limits its value.
Why it matters: These conditions help determine whether the extra selection and retraining work is justified. - Failure Modes and Validation: Describe plausible ways the loop can be misled and the checks needed before trusting or acting on its rankings.
Why it matters: Finding high-scoring candidates efficiently is not the same as establishing that they will bind or can be made.
Key terms: active learning loop, surrogate model, acquisition function, oracle, budget, TensorBoard, top 100 molecules, training loss, PCA, chemical space coverage
9.3 Active Learning for Lead Optimization: Free Energy Perturbation Experiments
This section explains that active learning is not limited to docking and places free energy calculations in the context of lead optimization.
- Use-Dependent Oracle Choice: Explain why docking is not necessarily the best or only oracle for active learning in drug discovery.
Why it matters: It helps you see docking as one possible application of active learning, not a requirement. - Lead Optimization Aims: Describe when lead optimization begins and what it aims to improve about promising hits.
Why it matters: It explains why the section shifts from finding hits to improving them.
Key terms: oracle, active learning, docking, hit compounds, lead optimization, binding affinity, off-target activity
9.3.1 The Role of Free Energy Calculations
This section explains how free energy calculations support lead optimization and how active learning can identify accurate, target-specific FEP protocols.
- Free Energy for Potency Optimization: Explain how free energy calculations estimate ligand binding affinity and why their predictions can guide potency optimization.
Why it matters: Binding-affinity estimates help teams prioritize structural changes before committing to synthesis and testing. - FEP Ligand Transformation: Describe how FEP uses a transformation between ligands and molecular dynamics simulations to estimate binding free energy.
Why it matters: This explains what the computationally intensive FEP experiments in the optimization workflow actually evaluate. - Target-Specific FEP Protocols: Describe what an FEP protocol specifies and why its settings may need optimization for a particular target.
Why it matters: The dependence on target-specific settings is the reason to optimize the protocol, not just choose which ligands to simulate. - Active Learning Protocol Selection: Trace how active learning selects and validates an FEP protocol using experimental binding data, FEP results, and a surrogate model.
Why it matters: The surrogate helps focus costly FEP calculations on promising settings while a separate test set checks the final choice.
Key terms: free energy calculations, binding affinity, potency optimization, Free Energy Perturbation (FEP), molecular dynamics (MD), FEP protocol, receptor structure, ligand structure and binding pose, force fields, solvent model
9.4 Summary
Where this is used later
- The Evolution of AI-Driven Drug Discovery (chapter 13) · Earlier docking work provides another example of a task-specific tool.
- Challenges of Drug-Target Affinity Prediction (chapter 11) · Introduces protein-ligand docking as a structure-based approach.
- Platform Overview (chapter 13) · Introduces protein-ligand docking.
- Defining Multimodal AI Systems (chapter 13) · The active learning loop previously introduced for drug discovery.
- Integration Patterns Across Platforms (chapter 13) · Develops the active learning loop referenced by this section.
- Future Directions (2023-2030) (chapter 13) · The active learning loop as an earlier example of iterative model-guided selection
Concepts explained in this chapter
- active learningActive learning is a machine-learning approach that iteratively selects the most informative unlabeled examples for labeling to improve a model with limited labeling resources.
- surrogate modelA surrogate model is an inexpensive predictive model that approximates the performance or outputs of a more computationally expensive process.
- docking scoreA docking score is a numerical estimate produced by molecular docking to rank the predicted favorability of a ligand's binding pose or interaction with a target.
- uncertainty samplingUncertainty sampling is an active-learning acquisition strategy that selects examples for which the current model is most uncertain.
- labelerAn oracle is an authoritative labeling mechanism that converts an unlabeled data point into an accurate label at a specified computational, human, or financial cost.
- acquisition functionAn acquisition function is a rule that scores candidate data points according to their expected usefulness for selecting the next samples in sequential learning.
- molecular dockingMolecular docking is computational simulation of how a molecule fits and orients within a target binding site to estimate its binding behavior.
- progressive dockingDeep docking is an active-learning approach that uses machine-learning models to predict binding affinities across large compound libraries and prioritize compounds for docking or synthesis.
- TensorBoardTensorBoard is a visualization and experiment-tracking tool for inspecting machine-learning metrics, model behavior, and generated data.
- AutoDock VinaAutoDock Vina is an open-source molecular-docking program that predicts ligand poses and assigns docking scores for target binding sites.
- ablation studyAn ablation study evaluates the contribution of a system component by removing or modifying it and measuring the resulting change in performance.
- free energy perturbationFree energy perturbation is a computational method that estimates free-energy differences between related molecular systems from a statistical-mechanical transformation.
- uncertaintyUncertainty is a quantitative or qualitative representation of limited confidence in a model prediction.
- expected improvementExpected improvement is an acquisition criterion that estimates the expected amount by which a candidate will improve on the best observed result.
- probability of improvementProbability of improvement is an acquisition criterion that estimates the probability that a candidate will outperform the best observed result.
- structure-based drug designStructure-based drug design is a drug-design approach that uses three-dimensional molecular structures and interactions to guide compound discovery and optimization.
- free energy calculationA free energy calculation is a computational method for estimating the change in thermodynamic free energy associated with a molecular process such as ligand binding.
- bounding boxA binding box is a three-dimensional spatial region that constrains where a docking algorithm searches for ligand binding poses.
- vina scoring functionThe Vina scoring function is AutoDock Vina's empirical function for estimating ligand binding favorability from pairwise atomic interaction terms and penalties.
- binding modeA binding mode is a specific predicted spatial arrangement and interaction pattern between a ligand and its target binding site.
- FEP protocolAn FEP protocol is a defined collection of molecular structures, force fields, solvent and sampling choices, simulation settings, and transformation parameters for an FEP calculation.
- structure preparationStructure preparation is the process of configuring molecular or protein structures for computational analysis by correcting, completing, or assigning required structural information.
- structure-based virtual screeningStructure-based virtual screening prioritizes compounds using structural information about a biological target or binding site.
- Monte Carlo dropoutMonte Carlo dropout estimates predictive uncertainty by retaining stochastic dropout during inference and aggregating multiple forward-pass predictions.
- data representativenessRepresentativeness is the degree to which selected samples reflect the distribution and diversity of the data domain.
- exploration-exploitation trade-offThe exploration-exploitation trade-off is the choice between gathering information from unfamiliar regions and favoring options already predicted to perform well.
- binding siteA binding site is the region of a protein or other biomolecule where a ligand or compound interacts.
- crystallographyCrystallography is a technique for determining molecular or material structures from the diffraction of radiation by crystals.
- ligandA ligand is a molecule that binds to a biomolecular target and can alter its activity or signaling.
- molecular conformationA conformation, or conformer, is a particular three-dimensional arrangement of a molecule accessible through internal structural flexibility without changing its connectivity.
- molecular targetA biological target is a biomolecule whose activity can be modulated to produce a desired medicinal effect.
- protein structureProtein structure is the three-dimensional arrangement of a protein's atoms and constituent chains.
- interaction fingerprintAn interaction fingerprint is a structured encoding of the types and locations of interactions between a ligand and a target.
- binding poseA docking pose is a predicted three-dimensional position, orientation, and conformation of a ligand relative to a target binding site.
- qualityA scoring function is a function that assigns a numerical value to a candidate according to specified criteria.
- dropoutDropout is a regularization technique that randomly disables units during training to reduce reliance on particular representations.
- ReLUReLU is an activation function that outputs zero for negative inputs and otherwise returns the input value.
- feedforward networkA position-wise feed-forward network is a neural network applied independently to each sequence position, commonly using two linear transformations and a nonlinear activation.
- MaxMinThe MaxMin algorithm iteratively selects points that are maximally distant from points already selected.
- samplingRandom sampling is the selection of items uniformly at random from an available population or pool.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.
