What you will learn in Build AI Drug Discovery Pipelines
Use AlphaFold-inspired models to screen, generate, and optimize drug candidates
Chapter by chapter
Chapter 1: The Drug Discovery Process
The chapter introduction explains why drug discovery is difficult and why computational screening and machine learning are useful.
What you learn in chapter 1 →Chapter 2: Ligand-based Screening: Filtering & Similarity Searching
Introduces virtual screening as a computational way to prioritize promising compounds for focused experimental testing.
What you learn in chapter 2 →Chapter 3: Ligand-based Screening: Machine Learning
Introduces the stages and vocabulary of a machine-learning workflow for cardiotoxicity prediction and explains why the workflow is iterative.
What you learn in chapter 3 →Chapter 4: Solubility Deep Dive with Linear Models
The chapter introduces why drug solubility matters in drug discovery and how linear models can help predict it from compound properties.
What you learn in chapter 4 →Chapter 5: Classification: Cytochrome P450 Inhibition
Introduces drug metabolism as a drug-discovery concern and explains why changes in cytochrome P450 activity can affect drug exposure and interactions.
What you learn in chapter 5 →Chapter 6: Case Study: Small Molecule Binding to an RNA Target
The introduction explains why small-molecule binding to RNA is therapeutically promising but difficult to predict, and frames QSAR as a way to relate molecular descriptors to measured binding.
What you learn in chapter 6 →Chapter 7: Unsupervised Learning: Repurposing Drugs, Curating Compounds, & Screening Fragments
The chapter introduces unsupervised learning as a way to find structure in unlabeled drug-discovery data and explains the basic purposes of dimensionality reduction, clustering, and density estimation.
What you learn in chapter 7 →Chapter 8: Introduction to Deep Learning
This introduction previews the chapter’s move from earlier ligand-based machine learning work to applied deep learning with PyTorch and outlines the sections ahead.
What you learn in chapter 8 →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.
What you learn in chapter 9 →Chapter 10: Generative Models for De Novo Design
The introduction frames drug discovery as a molecular design challenge and explains how a learned latent space could focus the search for new molecules.
What you learn in chapter 10 →Chapter 11: Graph Neural Networks for Predicting Drug-Target Affinity
The chapter introduction explains why graph representations may improve drug-target affinity prediction and motivates extending GraphDTA from a drug-graph model to a dual-graph design.
What you learn in chapter 11 →Chapter 12: Transformers for Protein Structure Prediction
This introduction explains why protein structures matter to drug discovery and why the gap between known sequences and experimentally determined structures motivates computational prediction.
What you learn in chapter 12 →Chapter 13: Multimodal AI Systems for End-to-End Drug Discovery
The introduction frames end-to-end drug discovery as prioritizing target-and-molecule hypotheses for experiments and clarifies the chapter’s three uses of “multimodal.”
What you learn in chapter 13 →Concepts explained in this book
- latent spaceA latent representation is a learned internal encoding of input data in a continuous or structured representation space.
- embeddingAn embedding is a learned numerical vector representation of an object that captures properties relevant to a task.
- ESM-2ESM-2 is a protein language model trained on large collections of amino acid sequences to learn sequence representations and predict mutation effects.
- 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.
- attentionAn attention mechanism computes input-dependent weights that determine how strongly elements of a sequence or representation influence one another.
- encoderAn encoder is a model component that transforms an input into a lower-dimensional representation while retaining selected information.
- transformer architectureA transformer architecture is a neural-network architecture that models relationships among sequence elements using attention and position-wise transformations, potentially with encoder and decoder stacks.
- masked language modelingMasked language modeling is a self-supervised technique that trains a model to predict tokens hidden or replaced within an input sequence.
- learning rateA learning rate is an optimization step-size configuration that controls the magnitude of parameter updates during iterative learning.
- ESMFoldESMFold is a protein structure prediction model that uses protein-language-model embeddings to predict three-dimensional structure without generating a multiple sequence alignment.
- surrogate modelA surrogate model is an inexpensive predictive model that approximates the performance or outputs of a more computationally expensive process.
- mean squared errorMSE is mean squared error, the average of the squared differences between predicted and observed numerical values.
- decoderA decoder is the component of a sequence model that generates an output sequence from an internal representation and previously processed tokens.
- linear layerA linear layer is a neural-network layer that applies an affine transformation to its input using a weight matrix and bias vector.
- linear regressionLinear regression is a regression method that models a linear relationship between input features and a target variable.
- machine learningMachine learning is a computational approach that learns patterns or predictive functions from data rather than relying solely on explicitly programmed rules.
- DGraphDTADGraphDTA, also called DualGraphDTA, is a dual-graph neural model for drug-target affinity prediction that represents both drug molecules and protein targets as graphs.
- neural networkAn artificial neural network is a parameterized computational model composed of interconnected units arranged in layers.
- 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.
- vanishing gradient problemA vanishing gradient is a training failure mode in which gradients become extremely small during backpropagation, preventing earlier layers from learning effectively.
- classifierClassification is a supervised learning task that predicts a discrete class label for each input.
- clusteringClustering is an unsupervised learning technique that partitions data instances into groups of similar instances.
- self-supervised learningSelf-supervised learning trains a model using supervisory signals or prediction targets derived from the input data itself rather than external labels.
- MegaMolBARTMegaMolBART is a transformer-based generative model that learns representations and generates molecular structures from molecular string data.
- contact mapA contact map is a representation indicating which pairs of residues or atoms in a biomolecule are spatially close or interacting.
- generalizationGeneralization is a model's ability to maintain predictive performance on previously unseen data drawn from the relevant distribution.
- variational autoencoderA variational autoencoder is an autoencoder that represents inputs with parameterized probability distributions in latent space and is trained using reconstruction and distribution-regularization objectives.
- autoencoderAn autoencoder is a neural model that encodes inputs into a latent representation and decodes that representation to reconstruct the inputs.
- DiffDockDiffDock is a diffusion-based molecular docking model that predicts ligand binding poses and associated confidence scores from protein-ligand inputs.
- amino acid sequenceProtein primary structure is the linear order of amino acid residues in a protein chain.
- gradientA gradient is a vector containing the partial derivatives of a scalar function with respect to its variables and pointing toward steepest increase.
- molecular conformationA conformation, or conformer, is a particular three-dimensional arrangement of a molecule accessible through internal structural flexibility without changing its connectivity.
- closed-loop learningA closed-loop learning system iteratively feeds experimental results back into predictive models or design decisions to improve subsequent outputs.
- dimensionality reductionDimensionality reduction is the transformation of data into fewer dimensions while retaining important information.
- feature importance scoreFeature importance is a quantitative estimate of how much an input feature contributes to a predictive model's decisions.
- sigmoidThe logistic function, or sigmoid, is an S-shaped function that maps real-valued inputs to values between zero and one.
- AlphaFoldAlphaFold2 is a deep learning model for predicting protein three-dimensional structures using sequence-derived evolutionary information and structural reasoning.
- feature vectorA feature vector is an ordered collection of numerical or categorical features representing one input instance.
- one-hot encodingOne-hot encoding represents each categorical token as a sparse vector with one active coordinate and all other coordinates inactive.
- BioNeMoBioNeMo is a platform for developing, optimizing, and serving generative and predictive AI models for biology and drug discovery.
- hidden layerA hidden layer is an intermediate neural-network layer whose artificial neurons transform representations between the input and output layers.
- DataLoaderPyTorch DataLoader is a utility for batching, optionally shuffling, and iterating over samples from a dataset.
- GRUA gated recurrent unit is a recurrent neural-network unit that uses learned gates to control information retained and propagated across sequence steps.
- regularizationRegularization is a modeling technique that constrains learning to reduce overreliance on particular features and improve performance on unseen data.
- multi-head attentionMulti-head attention is an attention mechanism that combines multiple independently parameterized attention operations to capture different relationships in a sequence.
- ReLUReLU is an activation function that outputs zero for negative inputs and otherwise returns the input value.
- evaluation datasetA validation set is a held-out collection of examples used during development to assess model performance and guide configuration choices.
- cross entropy lossCross-entropy is a loss function that measures the discrepancy between a target probability distribution and a predicted probability distribution.
- activation functionAn activation function is a function applied to a neuron's weighted input to produce its output, typically introducing nonlinearity.
- area under the ROC curveAUC is the area under the receiver operating characteristic curve and measures ranking-based discrimination between classes.
- RFdiffusionRFdiffusion is a diffusion model that generates protein backbone structures, optionally conditioned on structural constraints or motifs.
- reparameterization trickThe reparameterization trick expresses stochastic sampling from a parameterized distribution as a differentiable transformation of distribution parameters and independent noise.
- logitA logit is an unnormalized real-valued score for a possible output token before conversion to probabilities.
- QSAR modelQSAR modeling predicts molecular properties or biological activities from quantitative relationships between chemical structure and measured outcomes.
- PyTorchPyTorch is an open-source machine-learning framework for building and training neural networks with tensor computation and automatic differentiation.
- generative modelingA generative model learns patterns or a probability distribution from data and can produce new samples with similar characteristics.
- cross-attentionCross-attention is an attention mechanism in which queries from one sequence or representation attend to keys and values derived from another.
- encoder-only transformerAn encoder-only transformer is a transformer consisting of encoder layers that produces contextual representations without an autoregressive decoder.
- positional embeddingA positional encoding is a vector representation of a token's location that is added to its content embedding to provide sequence-order information.
- graph neural networkA graph neural network is a neural-network model that computes representations or predictions from graph-structured data by propagating information among connected nodes.
- trainingModel training is the process of fitting a model to learn a mapping from input features to target labels or values.
- MolMIMMolMIM is a conditional molecular generative model that produces compounds subject to specified property ranges and structural requirements.
- random forestA random forest is an ensemble model that combines predictions from multiple decision trees trained with randomized data samples or feature selections.
- molecular dockingMolecular docking is computational simulation of how a molecule fits and orients within a target binding site to estimate its binding behavior.
- protein folding problemProtein folding is the process by which an amino acid sequence adopts a three-dimensional structure.
- class imbalanceClass imbalance is a dataset condition in which some outcome values or classes occur much more frequently than others.
- overfittingOverfitting is the condition in which a model fits training examples too closely and performs poorly on new data.
- protein structureProtein structure is the three-dimensional arrangement of a protein's atoms and constituent chains.
- weight initializationWeight initialization is the procedure for assigning initial numerical values to a model's learnable weights before training.
- similarity searchSimilarity search retrieves or ranks molecules according to their structural similarity to one or more query molecules.
- exploding gradientThe exploding gradient problem is a training failure in which gradients grow excessively during backpropagation, causing unstable parameter updates.
- molecular descriptorA molecular descriptor is a numerical or symbolic representation of a molecule's physicochemical, structural, topological, or quantum-mechanical properties.
- teacher forcingTeacher forcing is a sequence-model training method that supplies ground-truth previous tokens as decoder inputs instead of the model's own predictions.
- backpropagationA backward pass is the reverse computation that propagates derivatives from a loss through a model to obtain parameter gradients.
- forward passA forward pass is the computation that transforms model inputs into predictions by propagating them through the model.
- generationInference is the process of applying a trained model to input data to produce predictions.
- model calibrationCalibration is the agreement between a model's predicted probabilities and the observed frequencies of outcomes among examples receiving those probabilities.
- protein structure predictionProtein structure prediction is the computational inference of a protein's three-dimensional structure from its amino acid sequence or related data.
- principal component analysisPCA is principal component analysis, a linear dimensionality-reduction method that projects data onto orthogonal directions of greatest variance.
- deep learningA deep neural network is a neural network composed of multiple computational layers that learn hierarchical representations from data.
- dropoutDropout is a regularization technique that randomly disables units during training to reduce reliance on particular representations.
- SimpleFoldSimpleFold is a protein structure prediction model that generates structured protein conformations from noisy representations using flow matching.
- hyperparameter tuningHyperparameter tuning is the systematic selection of model configuration values that are not learned directly from the training examples.
- tensorA tensor is a multidimensional numerical array that stores data and supports operations for machine-learning computation.
- log-ratioA log-likelihood ratio is the logarithm of the ratio between two likelihoods, used here to compare the plausibility of mutated and reference protein sequences.
- pseudo-perplexityPseudo-perplexity is an evaluation metric for masked language models that aggregates the model's conditional uncertainty over tokens masked in a sequence.
- zero-shot learningZero-shot transfer is the application of a model to a task or modality without task-specific labeled examples for that application.
- 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.
- molecular graphA molecular graph represents a molecule as a graph whose nodes are atoms and whose edges are chemical bonds.
- applicability domainAn applicability domain is the region of input chemical or feature space in which a predictive model is expected to make reliable predictions.
- chemical spaceChemical space is the set of possible chemical structures and compounds considered for discovery or analysis.
- language modelA language model assigns probabilities to sequences or predicts the next token from preceding tokens.
- training datasetA training set is the collection of examples provided to a learning algorithm for fitting a model.
- sidechainA side chain is the variable chemical group attached to an amino acid's common molecular backbone.
- residual connectionA residual connection adds a layer's input representation to that layer's transformed output.
- f-scoreThe F1 score is the harmonic mean of precision and recall for a classification system.
- virtual screeningVirtual screening is computational prioritization of chemical compounds as drug candidates by predicting target interactions or drug-relevant properties.
- pkdpKd is the negative base-ten logarithm of a ligand's equilibrium dissociation constant, with larger values indicating stronger binding affinity.
- mean absolute errorMean absolute error is the average of the absolute differences between observed target values and model predictions.
- hidden stateA hidden state is a learned intermediate vector that summarizes sequence information available to a recurrent or sequential model at a particular position.
- compound filterCompound filtering is the computational removal of molecules with undesirable properties or substructures from a screening collection.
- model-agnostic methodModel-agnostic interpretability comprises methods that explain or analyze a model's predictions without relying on its internal implementation or architecture.
- ligand-based virtual screeningLigand-based virtual screening prioritizes compounds by using information from known ligands rather than an explicit target structure.
- transfer learningTransfer learning is a technique that adapts knowledge learned from one dataset or task to improve performance on another task.
- reconstruction lossA reconstruction loss quantifies the discrepancy between an input and the output reconstructed from its encoded representation.
- posterior collapsePosterior collapse is a variational autoencoder failure mode in which the decoder ignores the latent representation and the learned posterior approaches the prior.
- blueprintA blueprint layer is an architectural layer containing documented, customizable reference workflows that compose multiple models and services.
- conditional generationConditional generation is the generation of data or structures subject to specified input constraints or conditioning information.
- DNABERTDNABERT is a transformer model trained on DNA sequences to learn patterns associated with gene regulation and sequence function.
- foundational model layerA foundation model layer is an architectural layer containing pretrained models that provide reusable representations or generation capabilities across downstream tasks.
- GeneformerGeneformer is a transcriptome-level model that learns gene co-expression patterns from single-cell RNA sequencing data.
- NIMNVIDIA NIM is a deployable inference microservice for serving machine-learning models through standardized interfaces.
- Phenom modelPhenom-Beta is a vision-transformer biological foundation model that learns biologically meaningful representations from cellular microscopy images.
- ProteinMPNNProteinMPNN is an inverse-folding model that designs amino-acid sequences predicted to adopt a supplied protein backbone structure.
- ProtGPT2ProtGPT2 is an autoregressive generative protein model that produces amino-acid sequences intended to form viable protein structures.
- scBERTscBERT is a transformer model optimized for analyzing sparse, noisy single-cell RNA sequencing data.
- pharmacophore modelingLigand-based pharmacophore modeling derives common interaction features and geometry by comparing conformations of compounds with related activity.
- epistasisEpistasis is the dependence of a mutation's effect on the presence of other mutations in the same sequence.
- ESMEvolutionary Scale Modeling is a family of protein language models trained on large collections of amino acid sequences.
- low-rank adaptationLoRA is a parameter-efficient fine-tuning method that learns low-rank updates while keeping most pretrained weights fixed.
- Rotary Position EmbeddingA rotary position embedding encodes relative positional information by rotating query and key representations within attention computations.
- miscalibrationMiscalibration is the discrepancy between a model's predicted probabilities and the observed frequencies of the corresponding outcomes.
- matrix-matrix multiplicationMatrix multiplication is the arithmetic operation that combines matrices to produce a matrix of weighted linear transformations.
- interceptAn intercept is the constant term in a linear model that represents the predicted output when all input features are zero.
- layer normalizationLayer normalization normalizes features within each individual sample and applies learned scale and shift parameters.
- scaled-dot product attentionScaled dot-product attention computes attention weights from query-key dot products divided by the square root of the key dimension and applies them to value vectors.
- sinusoidal positional encodingA sinusoidal position encoding represents sequence positions with fixed sine and cosine functions added to model inputs.
- computation graphA computational graph is a directed graph whose nodes represent computational operations or values and whose edges represent their dependencies.
- embedding layerA token embedding is a learned vector representation assigned to a discrete input token.
- novelty scoringNovelty scoring estimates how underexplored or differentiated a therapeutic target or discovery opportunity is relative to existing research and investment.
- DAVIS datasetDAVIS is a benchmark dataset of drug-target pairs with measured binding-affinity values, commonly used to evaluate affinity-prediction models.
- KIBA datasetA KIBA score is a unified bioactivity metric derived by integrating heterogeneous measurements of drug-target binding strength.
- 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.
- binding affinityBinding affinity is the strength of association between a ligand and its biological target, commonly quantified by an equilibrium constant or related parameter.
- varianceDistribution variance is the expected squared deviation from a distribution's mean and measures its spread.
- position-specific scoring matrixA position-specific scoring matrix is a matrix assigning residue or symbol scores that vary by sequence position, typically based on observed frequencies or substitution preferences.
- lipophilicityLogP is the logarithm of a compound's octanol-water partition coefficient and is a measure of lipophilicity.
- SMILESSMILES is a line notation that encodes molecular atoms, bonds, and connectivity as a text string using defined symbols and grammar rules.
- reinforcement learningReinforcement learning is a machine-learning approach in which an agent learns sequential actions through feedback from rewards or penalties.
- encoder-decoder architectureA sequence-to-sequence architecture maps an input sequence to an output sequence, typically through an encoder and decoder.
- pipelineA sequential pipeline is an architecture in which distinct processing stages execute in order and pass each stage's outputs to the next stage.
- molecular fingerprintA molecular fingerprint is a fixed-length, often binary, representation encoding structural features or substructures present in a molecule.
- automatic differentiationAutomatic differentiation is the systematic computation of exact derivatives of programs by applying the chain rule to their elementary operations.
- model architectureA neural network architecture is the arrangement and connectivity of layers and neurons that defines a neural network's computational structure.
- neuronAn artificial neuron is a computational unit that transforms weighted inputs, usually with a bias and activation function, into an output.
- AUC-PRThe area under the precision-recall curve is a summary metric that aggregates precision and recall across classification decision thresholds.
- interpretabilityModel interpretability is the degree to which a model's predictions, representations, or decision process can be understood by humans.
- cross-validationCross-validation estimates model performance or selects configurations by repeatedly training and evaluating on different partitions of the training data.
- Brier scoreBrier score is the mean squared error between predicted probabilities and binary observed outcomes.
- isotonic regressionIsotonic regression is a non-parametric regression method that fits a monotonic function to map scores to calibrated outcomes.
- Platt scalingPlatt scaling is a parametric probability-calibration method that fits a logistic regression mapping from model scores to outcome probabilities.
- precision-recall curveA precision-recall curve plots classifier precision against recall as the classification threshold varies.
- ROC curveA receiver operating characteristic curve plots true positive rate against false positive rate as the classification threshold varies.
- reliability diagramA reliability diagram, or reliability plot, compares predicted probabilities with observed outcome frequencies across probability bins to assess calibration.
- root mean squared calibration errorRoot mean squared calibration error is the square root of the mean squared difference between predicted probabilities and observed outcome frequencies.
- Youden IndexYouden’s Index is the difference between a classifier's true positive rate and false positive rate, equivalently sensitivity plus specificity minus one.
- protein language modelA protein language model is a model trained on protein sequences to learn sequence representations or predict biologically meaningful sequence patterns.
- nonlinear mappingNonlinearity is the property of a function or model in which outputs cannot be expressed as a linear transformation of inputs.
- principal componentA principal component is a transformed feature formed as a linear combination of original features that captures a successive amount of data variance.
- recurrent neural networkA recurrent neural network cell is the repeated computational unit that combines the current input with the previous hidden state to produce an output and updated hidden state.
- ATP-binding pocketAn ATP-binding pocket is the kinase binding-site cavity where adenosine triphosphate binds during catalysis.
- high-throughput screeningHigh-throughput screening is the automated testing of many compounds in parallel for activity against a biological or chemical assay.
- VAE-CYCVAE-CYC is a variational molecular autoencoder architecture that combines sequence encoding, latent-variable sampling, autoregressive decoding, and cyclical KL regularization.
- agglomerative clusteringAgglomerative clustering is a bottom-up hierarchical clustering method that starts with singleton clusters and successively merges the closest clusters.
- centroidA centroid is the arithmetic mean position of the points assigned to a cluster.
- dendrogramA dendrogram is a tree diagram that represents the hierarchy and merge or split levels produced by hierarchical clustering.
- exclusion sphereAn exclusion sphere is the neighborhood around a selected cluster centroid whose flagged members cannot serve as members of another cluster or as another centroid.
- hierarchical clusteringHierarchical clustering is a clustering technique that builds a hierarchy of nested groups by iteratively merging or splitting clusters.
- inter-cluster distanceIntercluster distance is a rule for measuring separation between two clusters from the distances among their members.
- K-means clusteringK-means clustering partitions data into a specified number of groups by iteratively assigning points to nearby centroids and recomputing the centroids.
- prototype-based clusteringPrototype-based clustering represents each cluster by a prototype and assigns instances according to their proximity to those prototypes.
- redundancy ratioA redundancy ratio is the fraction of unique compound pairs whose similarity exceeds a specified threshold.
- performanceBig O notation describes an algorithm's asymptotic growth in time or memory as input size or dimensionality increases.
- space complexityMemory complexity is an asymptotic measure of the memory required by an algorithm as input size grows.
- Taylor-Butina clusteringTaylor-Butina clustering is a fingerprint-based clustering method that selects high-neighbor compounds as centroids and assigns nearby compounds using an exclusion threshold.
- Ward's methodWard's method is an agglomerative linkage criterion that merges the pair of clusters producing the smallest increase in within-cluster variance.
- hydrogen bondA hydrogen bond, or hydrogen bonding interaction, is a directional noncovalent attraction involving a hydrogen atom bonded to an electronegative atom and another electronegative atom.
- multiple objectivesMulti-objective optimization is the optimization of multiple potentially competing objectives simultaneously to identify balanced solutions.
- ligandA ligand is a molecule that binds to a biomolecular target and can alter its activity or signaling.
- parameterA model parameter is a numerical value that governs a model's mapping from inputs to outputs and is typically learned from data.
- foundation modelA foundation model is a broadly pretrained model that can support multiple downstream tasks through adaptation or reuse of its learned representations.
- supervised learningSupervised learning trains a model from examples paired with target labels or values.
- torch.nn.Modulenn.Module is the PyTorch base class for defining neural-network components with registered parameters, submodules, and computation.
- cold-start evaluationCold-start evaluation tests a model on entities or entity types that are absent from its training data.
- density estimatorDensity estimation is an unsupervised technique that estimates the probability distribution underlying sampled data.
- 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.
- pharmacophoric featureA pharmacophoric or pharmacophore feature is a molecular interaction capability, such as hydrogen-bond donation, hydrogen-bond acceptance, hydrophobicity, or aromaticity, represented in a pharmacophore.
- weight matrixA weight matrix is a rectangular array of model parameters whose entries associate input features with class-specific contributions.
- 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.
- unsupervised learningUnsupervised learning trains a model from data without provided target labels.
- mean poolingMean pooling forms one representation by averaging the representations of sequence positions.
- concordance indexThe concordance index is a metric that measures how often a model correctly orders pairs of examples according to their true values, with tied predictions receiving half credit.
- protonationProtonation is the addition of a proton to a molecule or ionizable chemical group.
- molecular fragmentA molecular fragment is a small chemical structure that can bind a biological target and serve as a starting point or building block for a larger compound.
- biasStatistical bias is systematic error caused by a model's simplifying assumptions that prevent it from capturing the true structure of the data.
- Kennard-Stone algorithmThe Kennard-Stone algorithm selects a representative subset of samples by iteratively choosing points that maximize distance from already selected points.
- conformational spaceConformational space is the collection of accessible three-dimensional molecular arrangements arising from structural flexibility.
- protomer spaceProtomer space is the set of chemically relevant protonation-state variants considered for a ligand.
- drug repurposingDrug repurposing is the identification of new therapeutic uses for an existing drug outside its original medical indication.
- self-organizing mapA self-organizing map is an unsupervised neural-network technique that maps high-dimensional data onto a lower-dimensional grid while preserving neighborhood relationships.
- oddsOdds are the ratio of the probability of an event to the probability of its complement.
- odds ratioAn odds ratio is the ratio of two odds, often comparing the odds under a changed feature value with the odds under a reference value.
- decision treeA decision tree is a hierarchical predictive model that applies feature-based decision rules to reach an output.
- autoregressive modelAutoregressive generation produces a sequence one element at a time, conditioning each prediction on previously generated elements.
- latent distributionA latent distribution is a probability distribution over unobserved variables that encode underlying factors of observed data.
- word dropoutWord dropout is a regularization technique that randomly replaces input tokens with an unknown-token marker during training.
- qualityA scoring function is a function that assigns a numerical value to a candidate according to specified criteria.
- logistic regressionLogistic regression is a statistical classification method that models class probabilities using a logistic function of input features.
- cyclical annealingCyclical KL annealing repeatedly increases a variational autoencoder's KL weight from zero to a maximum and resets it across multiple cycles.
- data leakageData leakage is the inadvertent use of information outside a training partition during model fitting, producing biased performance estimates.
- softmaxSoftmax is a function that converts a vector of scores into nonnegative values summing to one.
- tokenizationTokenization is the conversion of structured input into a sequence of discrete tokens for computational processing.
- curse of dimensionalityThe curse of dimensionality is the phenomenon in which statistical estimation, geometric coverage, and related computational tasks become increasingly difficult as the number of dimensions grows.
- enrichment factorEnrichment factor is a screening metric that compares the proportion of active compounds found in an early-ranked subset with the proportion expected by random selection.
- Kullback–Leibler divergenceKullback-Leibler divergence is a measure of how one probability distribution differs from a reference probability distribution.
- recallTrue positive rate is the proportion of actual positive instances correctly predicted as positive.
- baggingBagging is an ensemble technique that trains multiple instances of a learning algorithm on bootstrap-sampled subsets of data and aggregates their predictions.
- multimodal foundation modelA multimodal biological foundation model jointly represents biological information from multiple modalities such as sequences, structures, expression measurements, and phenotypes.
- ensembleEnsemble learning combines multiple individual predictive models to form a composite model with improved robustness or accuracy.
- homology modelingHomology modeling predicts a protein structure by using experimentally determined structures of evolutionarily related proteins as templates.
- dissociation constantA dissociation constant is an equilibrium measure of the tendency of a molecular complex to dissociate, with lower values indicating stronger binding.
- smiles pair encodingSMILES Pair Encoding (SmilesPE) is a subword tokenization method that learns frequent token pairs or molecular substrings in SMILES strings and encodes them as merged tokens.
- feature engineeringFeature engineering is the manual design or selection of input representations intended to expose task-relevant patterns to a predictive model.
- Chemistry42Chemistry42 is a drug-design platform that generates and evaluates candidate molecular structures with multiple computational models.
- constraint-satisfaction problemConstraint satisfaction is the process of finding outputs that obey specified hard requirements while possibly optimizing additional soft objectives.
- differential expression analysisDifferential expression analysis identifies genes or transcripts whose expression levels differ systematically between biological conditions or groups.
- druggabilityA druggability assessment evaluates whether a biological target has features that make it amenable to modulation by a drug-like molecule.
- TNIKTNIK is a kinase involved in inflammatory and fibrotic signaling pathways and investigated as a therapeutic target in pulmonary fibrosis.
- forced vital capacityForced vital capacity is the volume of air that a person can forcibly exhale after taking a maximal breath.
- GENTRLGENTRL is a generative molecular-design algorithm that combines latent-space generation with reinforcement learning and property-based rewards.
- idiopathic pulmonary fibrosisIdiopathic pulmonary fibrosis is a chronic progressive lung disease characterized by unexplained scarring and fibrosis of lung tissue.
- IND-enabling studiesAn IND-enabling study is a preclinical study package generated to support an investigational new drug application and first-in-human testing.
- PandaOmicsPandaOmics is a computational drug-discovery platform that analyzes omics data and scientific literature to support biological target identification.
- pathway enrichment analysisPathway enrichment analysis tests whether a set of genes is disproportionately represented in predefined biological pathways.
- threshold-based rewardA threshold-based reward assigns optimization feedback only when candidate properties meet specified minimum or maximum thresholds.
- precisionPrecision is the proportion of predicted positive instances that are truly positive.
- inverted dropoutInverted dropout scales retained activations during training so their expected magnitude matches the full activation signal at inference.
- unstable gradient problemThe unstable gradient problem is the failure of gradients in deep networks to maintain a useful scale as they propagate across layers.
- hyperparameterA hyperparameter is a configuration value selected before or outside model fitting that controls an algorithm, representation, preprocessing, or model complexity.
- molecular targetA biological target is a biomolecule whose activity can be modulated to produce a desired medicinal effect.
- uncertainty samplingUncertainty sampling is an active-learning acquisition strategy that selects examples for which the current model is most uncertain.
- adamAdam is an adaptive stochastic-gradient optimization algorithm that maintains running estimates of gradient moments to update model parameters.
- graph convolutional networkA graph convolutional network is a graph neural network that updates node representations by aggregating information from neighboring nodes with graph convolution operations.
- numerical differentiationThe finite difference method numerically approximates derivatives from function values at nearby input points.
- partial derivative analysisPartial derivative analysis estimates how infinitesimal changes in individual input descriptors affect a model's prediction.
- sensitivity analysisSensitivity analysis evaluates how changes in inputs or assumptions affect a model's outputs.
- SHAPSHAP is a model-interpretation method that attributes a prediction to input features using Shapley-value-based contributions.
- Shapley valueA Shapley value is a cooperative-game-theoretic attribution that assigns an average marginal contribution to an individual feature across coalitions.
- long-range dependencyA long-range interaction is a biologically relevant relationship between sequence positions that are far apart in the primary sequence but may be spatially close in the folded structure.
- bidirectional rnnA bidirectional recurrent neural network contains forward and backward recurrent networks whose outputs jointly represent each sequence position.
- exposure biasExposure bias is the mismatch caused when a sequence model is trained on ground-truth histories but must generate from its own possibly erroneous histories at inference.
- latent concatenationLatent concatenation is the direct joining of a latent vector with token embeddings at each sequence position before recurrent decoding.
- greedy algorithmGreedy decoding generates a sequence by selecting the highest-probability token at every step.
- temperature scalingTemperature sampling generates tokens from probabilities obtained by scaling logits with a temperature parameter that controls randomness and diversity.
- kinaseA protein kinase is an enzyme that transfers phosphate groups to proteins or other biological molecules.
- coarse-grainingCoarse-graining is the approximation of a detailed system by a model with fewer variables or reduced spatial and temporal resolution.
- gradient boostingGradient boosting is an ensemble learning technique that sequentially adds weak models trained to reduce the current ensemble's loss or errors.
- 3D molecular descriptorA 3D molecular descriptor is a numerical representation of properties derived from a molecule's three-dimensional geometry.
- Boltzmann-weighted averagingBoltzmann weighting is the averaging of state-dependent values with weights determined by their relative energies and temperature.
- Dimorphite-DLDimorphite-DL is a software package that predicts chemically plausible protonation states of molecules over a specified pH range.
- distance bounds matrixA distance bounds matrix stores lower and upper allowable distance bounds for pairs of atoms in a molecular structure.
- distance geometryDistance geometry is a method for generating molecular three-dimensional coordinates from constraints on distances between atoms.
- ETKDGETKDG is a distance-geometry conformer-generation method that incorporates experimental torsion preferences and chemical knowledge constraints.
- intermolecular force fieldA molecular force field is a parameterized mathematical model of atomic interactions and their contribution to molecular potential energy.
- Molecular Operating EnvironmentMOE is commercial molecular-modeling software for visualizing, analyzing, and manipulating molecular structures.
- protomerA protomer is a molecular form distinguished by its protonation state or by protonation-linked tautomeric structure.
- tautomerismTautomerization is the process by which tautomers interconvert through proton transfer and bond rearrangement.
- Universal Force FieldThe Universal Force Field is a general-purpose molecular mechanics force field parameterized to model structures and energies across broad classes of elements and compounds.
- early recognition problemEarly recognition is the ability of a ranking method to retrieve a large proportion of active compounds within a small initial fraction of a screened collection.
- molecule objectAn RDKit Mol is an in-memory molecular graph object representing atoms, bonds, and associated molecular information.
- probability densityA probability density function describes the relative likelihood density of a continuous random variable across possible values.
- grid searchGrid search is a hyperparameter optimization method that evaluates models across a specified Cartesian product of candidate parameter values.
- RandomizedSearchCVRandomized search is a hyperparameter optimization method that evaluates configurations sampled from specified parameter distributions or candidate ranges.
- kinase inhibitorA kinase inhibitor is a compound that reduces the catalytic or signaling activity of a kinase.
- 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.
- drug-likenessDrug-likeness is the degree to which a compound has physicochemical and structural characteristics associated with suitability as a drug.
- Lipinski's rule of fiveThe Lipinski rule of five is a heuristic set of molecular-property guidelines for estimating a compound's oral drug-likeness.
- random feature selectionFeature subsampling is the random selection of a subset of input features for constructing an individual model component or tree split.
- hydrophobic effectThe hydrophobic effect is the tendency of nonpolar molecular groups to associate away from water, often causing them to become buried within biological structures.
- root mean squared errorRoot mean squared error (RMSE) is the square root of the mean squared differences between predicted and observed values, expressed in the target's original units.
- support vector regressionSVR is support vector regression, a supervised regression method that fits a function using support vectors and an epsilon-insensitive loss.
- selectivitySelectivity is the extent to which a compound binds to or affects its intended target relative to other proteins or biological targets.
- hydrophobicityHydrophobicity is the tendency of a chemical group or molecule to partition into a nonpolar environment rather than an aqueous one.
- position probability matrixA position probability matrix records the probability of each amino acid at each sequence position.
- pseudocountPseudocount smoothing adds prior counts or probabilities to observed counts to avoid zero estimates and reduce overconfidence.
- ADMETAn ADMET property describes a drug compound's absorption, distribution, metabolism, excretion, or toxicity behavior in a biological system.
- antibodyAn antibody is an immune-system protein that specifically recognizes and can help neutralize a pathogen or other antigen.
- antibody heavy chainA heavy chain is one of the larger polypeptide chains forming an antibody and contributing to its antigen-binding and effector properties.
- antigenAn antigen is a molecule or molecular region specifically recognized by an antibody or other immune receptor.
- b-cell receptor sequence datasetA B-cell receptor sequence dataset is a collection of antibody receptor sequences annotated with biological targets or other immune-related properties.
- [CLS] tokenA CLS token is a special Transformer input token whose contextual representation can summarize an entire sequence for classification.
- fine-tuningFine-tuning is the continued training of a pretrained model on task-specific data, usually with its parameters initialized from pretraining.
- attention_maskAn attention mask is a tensor or rule that marks which sequence positions should contribute to attention or aggregation computations.
- classification headA classification head is a neural-network component that maps an input representation to scores or probabilities for discrete classes.
- variable regionA variable region is the sequence-diverse portion of an antibody that determines much of its antigen-binding specificity.
- edge indexedge_index is a two-row sparse tensor that records source and destination node indices for graph edges.
- multiple sequence alignmentA multiple sequence alignment is an arrangement of related biological sequences in aligned positions to reveal conserved and correlated patterns.
- optimizationAn optimizer is an algorithm that updates learned model parameters using loss gradients and an optimization rule.
- multilabel classificationMultilabel classification is a supervised learning task in which each instance can be assigned multiple class labels simultaneously.
- TensorBoardTensorBoard is a visualization and experiment-tracking tool for inspecting machine-learning metrics, model behavior, and generated data.
- batch vectorA batch vector assigns each flattened graph node to the original graph it came from during batched graph processing.
- AutoDock VinaAutoDock Vina is an open-source molecular-docking program that predicts ligand poses and assigns docking scores for target binding sites.
- early stoppingEarly stopping is a training-control technique that terminates iterative model fitting when additional updates no longer improve a selected evaluation criterion.
- generative chemistryGenerative molecular design is the computational creation of molecular structures intended to satisfy specified biological, chemical, or medicinal requirements.
- learning rate scheduleLearning rate scheduling is the adjustment of an optimizer's learning rate during model training according to a predefined or adaptive schedule.
- vocabularyA vocabulary is a mapping between the permitted text tokens and their integer identifiers.
- correlationThe Pearson correlation coefficient is a statistic ranging from -1 to 1 that measures the strength and direction of a linear association between two continuous variables.
- bandwidth parameterBandwidth is the scale parameter controlling the spatial range and smoothing of a kernel density estimate.
- kernel functionA kernel function is a smooth, symmetric, non-negative weighting function that determines how an observation contributes to an estimated density at a location.
- kernel density estimateKernel density estimation is a nonparametric method that estimates a continuous probability density by summing kernel functions centered on observed data points.
- target identificationTarget discovery is the identification of biological entities and disease relationships that may support new therapeutic hypotheses.
- IC50IC50, or the half-maximal inhibitory concentration, is the concentration of a compound required to reduce a specified biological activity by 50 percent under defined experimental conditions.
- canonical SMILESA unique SMILES is a canonicalized SMILES representation selected so that equivalent molecular structures receive the same textual representation.
- PyTorch GeometricPyTorch Geometric is an open-source extension library for PyTorch that provides data structures and operators for geometric deep learning on graphs and related domains.
- global structureGlobal structure is the overall arrangement and interrelationships among points or clusters across an entire data representation.
- jaccard distanceJaccard distance is a dissimilarity measure equal to one minus the ratio of set intersection size to set union size.
- local structureLocal structure is the pattern of relationships among data points that are near one another in a representation.
- manifoldA manifold is a lower-dimensional mathematical space embedded in a higher-dimensional space that locally resembles ordinary Euclidean space.
- manifold learningManifold learning is a family of dimensionality-reduction methods that models high-dimensional data as lying on or near a lower-dimensional manifold.
- UMAPUniform manifold approximation and projection is a nonlinear dimensionality-reduction technique for embedding high-dimensional data in a lower-dimensional space.
- knowledge graphA knowledge graph is a graph-based representation of entities and the typed relationships between them, supporting structured querying and inference.
- protein designProtein design is the creation or selection of amino acid sequences intended to adopt a specified three-dimensional structure and satisfy desired stability or functional constraints.
- compound libraryA compound library is a curated collection of chemical compounds available for screening or analysis.
- learning curveA learning curve is a diagnostic plot showing model performance or error as a function of training-set size.
- multimodal AI platformA multimodal AI platform is a computational platform that integrates models, data, and processing components across multiple biological or chemical modalities.
- artificial intelligenceArtificial intelligence is the field of computational methods for performing tasks that ordinarily require aspects of human intelligence, such as learning and reasoning.
- feature scalingFeature scaling transforms feature values to comparable numerical ranges before model fitting.
- representation learningRepresentation learning is a technique in which a model learns useful encoded features from raw data.
- binding poseA docking pose is a predicted three-dimensional position, orientation, and conformation of a ligand relative to a target binding site.
- adversarial validationAdversarial validation trains a classifier to distinguish examples from two data partitions in order to detect distribution differences between them.
- bootstrap samplingBootstrap sampling creates a sample by repeatedly drawing observations from a dataset with replacement.
- drug-target affinity predictionDrug-target affinity prediction is the computational estimation of the strength of binding between a drug molecule and a biological target.
- ensemble consensusEnsemble consensus is the degree to which the individual models in an ensemble produce mutually agreeing predictions for an input.
- gini importanceGini importance is a tree-ensemble feature-importance measure based on the weighted reduction in node impurity attributable to each feature.
- out-of-bag errorOut-of-bag evaluation estimates an ensemble's performance using the bootstrap samples excluded from training each individual model.
- similarity mapA similarity map is a molecular visualization that assigns atom-level weights according to how much each atom affects a model prediction when its contribution is ablated.
- alpha-helixAn alpha-helix is a right-handed coiled protein backbone structure stabilized by a repeating hydrogen-bond pattern.
- beta-sheetA beta-sheet is a protein secondary structure formed by hydrogen-bonded, laterally aligned beta-strands.
- beta-strandA beta-strand is an extended protein backbone segment that associates with other strands to form beta-sheets.
- hydrophobic coreA hydrophobic core is the water-excluding interior region formed by nonpolar residues packed within a folded protein.
- protein backboneA protein backbone is the repeating chain of main-chain atoms linking amino acid residues, excluding their side chains.
- secondary structureSecondary structure is a local, regular geometric arrangement of a protein backbone stabilized mainly by hydrogen bonds.
- tertiary structureTertiary structure is the complete three-dimensional arrangement of a single protein chain.
- drug candidateA clinical candidate is a drug-development compound selected for advancement into clinical testing based on its efficacy, safety, and development properties.
- model complexityModel complexity is the degree of flexibility or representational richness of a model's function class.
- extended-connectivity fingerprintAn extended-connectivity fingerprint is a molecular fingerprint that encodes atom-centered neighborhoods expanded through successive bond-connectivity diameters.
- epochAn epoch is one complete pass through all examples in a machine-learning training dataset.
- polynomial feature interactionA polynomial feature is a feature formed by applying powers or products of original input features to represent nonlinear interactions.
- 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.
- coverage errorCoverage error is a multilabel ranking metric measuring how many top-ranked labels must be included to cover all true labels for an instance.
- micro averagingMicro-averaging pools classification outcomes across classes before calculating an aggregate metric, giving more influence to frequent classes or labels.
- pairwise ranking lossRanking loss is a multilabel metric measuring the fraction of incorrectly ordered positive-negative label pairs.
- adenosine triphosphateAdenosine triphosphate is a nucleotide that supplies chemical energy for cellular processes and phosphate-transfer reactions.
- EGFREGFR is the epidermal growth factor receptor, a cell-surface protein that mediates signaling in response to epidermal growth factor.
- Enamine Hinge Binder libraryThe Enamine Hinge Binder library is a collection of small molecules designed to bind kinase hinge regions.
- KinodataKinodata is a curated dataset of kinase bioactivity measurements assembled for computational kinase research.
- G protein-coupled receptorA G protein-coupled receptor (GPCR) is a cell-surface receptor that transmits extracellular signals through intracellular heterotrimeric G proteins.
- hinge regionA hinge region is a kinase structural region within the ATP-binding pocket that commonly forms interactions with inhibitors.
- ligand-gated ion channelA ligand-gated ion channel is a membrane-spanning receptor that opens or closes an ion-conducting pore when a ligand binds.
- nuclear receptorA nuclear receptor is an intracellular ligand-activated receptor that regulates DNA transcription and gene expression.
- promiscuous compoundOff-target promiscuity is the tendency of a compound to inhibit or modulate multiple unintended biological targets.
- tyrosine kinase inhibitorA tyrosine kinase inhibitor is a compound that blocks the catalytic or signaling activity of a tyrosine kinase.
- tyrosine kinase-linked receptorA tyrosine kinase-linked receptor is a membrane receptor associated with tyrosine-kinase signaling that responds to extracellular ligands.
- best matching unitA best matching unit is the node in a self-organizing map whose weight vector is closest to a given input vector.
- Gaussian neighborhood functionA Gaussian neighborhood function assigns larger update weights to nodes nearer the best matching unit and smaller weights to nodes farther away according to a Gaussian profile.
- neighborhood functionA neighborhood function assigns the magnitude of a node update according to the node's distance from the best matching unit.
- quantization errorQuantization error is the average distance between input data points and their best matching units in a self-organizing map.
- topographic errorTopographic error is the proportion of data points whose first and second best matching units are not adjacent in a self-organizing map.
- unified distance matrixA unified distance matrix is a self-organizing-map visualization that represents the average distances between each node and its neighboring nodes.
- accuracyAccuracy is the proportion of all classified instances that receive the correct predicted class.
- molecular dynamicsMolecular dynamics is simulation of the time-dependent physical movement of atoms and molecules under modeled forces.
- moleculeA molecule is an electrically neutral group of atoms held together by chemical bonds.
- signalA feature is an individual measurable input variable or encoded attribute used to represent an example for machine-learning prediction.
- data explorationExploratory data analysis is the examination and summarization of dataset attributes, distributions, and relationships to identify patterns and inconsistencies.
- RDKitRDKit is an open-source cheminformatics toolkit for representing, manipulating, visualizing, and analyzing chemical structures.
- hERG inhibitionhERG blockage is inhibition of the human ether-à-go-go-related gene potassium channel by a chemical compound.
- Levinthal's paradoxLevinthal's paradox is the observation that exhaustive conformational searching cannot explain the rapid folding of real proteins.
- molecular scaffoldA chemical scaffold is the core structural framework shared by a group of related molecules.
- boostingBoosting is an ensemble technique that sequentially trains models so later models correct errors made by earlier models.
- weak learnerA weak learner is a predictive model with limited individual accuracy that can contribute to a stronger ensemble when combined with other learners.
- structure-property relationshipA structure-property relationship describes how molecular structure corresponds to a chemical, biological, or functional property.
- preclinical developmentA preclinical study evaluates a drug candidate's safety and efficacy before testing in human clinical trials, typically using animals or model organisms.
- pharmacophore-based screeningPharmacophore-based sampling selects compounds that match hypothesized spatial and chemical features required for target binding.
- hitA hit is a compound that shows measurable binding or activity against a selected biological target during discovery screening.
- gradient descentGradient descent is an iterative optimization method that updates parameters in the direction that decreases a differentiable objective function.
- hERG potassium channelThe hERG potassium channel is a cardiac ion channel whose drug blockade can disrupt electrical repolarization and cause cardiotoxicity.
- drug discoveryThe process of identifying and characterizing chemical or biological entities that could become medicines.
- design-make-test-analyze cycleA design-make-test-learn cycle is an iterative drug-discovery workflow that designs compounds, synthesizes them, tests them experimentally, and learns from the results.
- integration patternAn integration pattern is a reusable architectural arrangement for connecting computational components and coordinating their data flows.
- parallel ensemble methodA parallel ensemble is an architecture in which multiple algorithms perform the same task and their outputs are combined.
- stochastic gradient descentMini-batch gradient descent is an optimization method that computes each parameter update from a small subset of the training data.
- inductive biasInductive bias is the set of assumptions a learning method uses to prefer some hypotheses or representations over others when generalizing from data.
- normalizationNormalization is the rescaling of input values or representations to a standardized magnitude or range.
- Morgan fingerprintA Morgan fingerprint is a circular molecular fingerprint that encodes atom-centered structural neighborhoods within a specified radius.
- objective functionA cost function measures aggregate prediction error across multiple data instances, with lower values representing better performance.
- atomAn atom is the basic unit of a chemical element that participates in chemical bonding.
- chiralityChirality is the property of an entity that cannot be superimposed on its mirror image.
- covalent bondA covalent bond is a chemical bond formed when atoms share electrons.
- enantiomerAn enantiomer is one of a pair of stereoisomers that are non-superimposable mirror images of each other.
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.
