Rules {Active Analog Approach} can align molecule activities by analogous structures.
Rules can align molecule activities by structural group {active pharmaceutical ingredient} (API).
Non-parametric methods {alternating conditional expectations} (ACE) can analyze activity.
Input "neuron" layer can hold physico-chemical properties and feed to middle layer using sigmoidal function {transfer function} with weights for outputs. Middle-layer "neurons" feed to one output {artificial neural network} (ANN).
Mathematical tools {chemometrics} applied to structure-activity relationships can find correlations and regression, recognize patterns, classify compounds and properties, design experiments for random screening and measuring, and validate results.
Quantum mechanics can pair with empirical approaches {computer-assisted metabolism prediction} (CAMP).
Cell arrays can pool more than one sample in cells, which allows fewer cells. Methods {deconvolution} can track sample pooling.
convolution
Convolution puts each sample into several cells, in regular pattern. Testing looks for one effect. Some cells show effect, but most do not. If sample causes effect, all cells with that sample show effect. Cells that contain that sample form pattern, so pattern indicates sample name.
deconvolution
Deconvolution uses convolution method and resulting cell pattern to find sample name. For example, for 100-cell array, 10 samples can feed into 90 cells, each cell receiving two samples. Ten cells have control samples. See Figure 1. Samples are in 18 cells. If testing shows that all 18 have activity over threshold, then that sample is effective.
If sample interactions cause effect, deconvolution can find interactions. If testing shows that only one cell has activity over threshold, those two samples must interact to be effective.
Combining quantum mechanics and physico-chemical properties {empirical-quantum chemical} {combined empirical/quantum chemical approach} can predict chemical behavior.
Models {Korzekwa-Jones model} can be for P-450 hydrogen abstraction and depend on difference between radical free energy and hydrogenated-atom free energy, as well as radical ionization potential and constant additive term.
Steric effects and van der Waals forces can cause fields {Lennard-Jones potential}.
Plots {loading plot} can use variable weights.
Semiempiric methods {modified neglect of differential overlap} (MNDO) can ignore overlap.
Molecule-modeling programs {molecular modeling}, such as Alchemy III and SYBYL from Tripos, can use electrostatics or quantum mechanics.
Non-parametric methods {non-linear partial least-squares, drug} (NPLS) can find least squares.
Response-surface methods {non-parametric method}, such as ACE, NPLS, and MARS, can be non-parametric.
IF/THEN statement sets {rule induction system, drug} can make output from input.
Graphs {score plot} can plot compound activities.
In multidimensional property space, compound clusters make classes separated by distance {cluster analysis} (CA). CA reduces unimportant variables. Substructure, topological index, physico-chemical property, calculated physico-chemical property, or hydrophobicity can determine classes.
Using discrete or continuous data and embedded data can put compounds into groups by activity level {cluster significance analysis} (CSA). CSA locates small clusters in large spaces.
Methods {Cone and Hodgkin similarity index} can measure molecular similarity.
Models {discriminant-regression model} (DIREM) can locate small clusters in large spaces.
Methods {distance-b program} (EVE) can locate small clusters in large spaces.
Unsupervised methods {hierarchical cluster analysis} (HCA) can measure distances between all points and make point vs. distance dendograms.
Structures can cluster in large databases by rating different compounds by similarity {Jarvis-Patrick method}.
Supervised methods {k-nearest neighbor} (k-NN) can calculate new-object distances from all other objects, to locate small clusters in large spaces.
Processes {partitioning} can merge individuals into groups or split whole into clusters.
Values {similarity measure} can compare distances.
Methods {single class discrimination} (SCD) can locate small clusters in large spaces.
Classifications {supervised method} can use already known patterns and clusters.
Activity and descriptor correlation vectors {trend vector analysis} can rank compound similarity.
Hierarchical methods {Ward's clustering method} {Ward clustering method} can agglomerate compounds to find clustering.
Supervised methods {Soft Independent Modeling of Class Analogies} (SIMCA) can use region-boundary or envelope models, to locate small clusters in large spaces.
Clustering methods {class analogy} can be SIMCA methods.
Distance measures {city-block distance} between structure-space points can be the same as Manhattan distance.
Distance measures {Manhattan distance} between structure-space points can be the same as city-block distance.
Distance measures {Minkowski distance} between structure-space points can be the same as Lp-metric.
Distance measures {Lp-metric} between structure-space points can be the same as Minkowski distance.
Structure-space points have distances {Mahalanobis distance}.
Hierarchical methods {centroid linkage} that agglomerate compounds can find clustering.
Hierarchical methods {complete linkage} that agglomerate compounds can find clustering.
Hierarchical methods {single linkage} that agglomerate compounds can find clustering.
Processes have factors {factor analysis}. Physico-chemical or structural properties describe compounds and have components {descriptor, factor} {X-variable, factor} {X descriptor, factor}. Chemical activities relate to variables {response variable}.
Methods {canonical factor analysis} can be for factor analysis.
Methods {centroid method} can be for factor analysis.
QSAR {combinatoric QSAR} can find similarities using different descriptor combinations.
Moments of inertia, and dipole and quadrupole moments, can be descriptors to calculate molecular moments {Comparative Molecular Moment Analysis} (CoMMA). CoMMA depends on shapes and charges.
Properties and structures have relations {Correlation Analysis}.
Factor-analysis methods {correspondence analysis} {correspondence factor analysis} (CFA) can use variable frequencies relative to activities, finds chi-square values, and finds principal components.
Principal components {disjoint principal component} (DPP) can be independent.
Thresholds {eigenvalue-one criterion} can be how many components have eigenvalues greater than one.
Unsupervised linear methods {eigenvector projection} can find factors.
Models {Evolutionary Programming} (EP) can add and subtract randomly selected variables, with crossing-over, and evaluate for "fitness" or best fit.
Methods {evolving factor analysis} (EVA) can analyze ordered data.
Methods {percentage of explained variance} {explained variance percentage} can indicate number of components required to reach 90% of total variance.
Parameters and descriptors can linearly relate to free energy {extrathermodynamic approach}.
Factor-analysis methods {free energy perturbation} (FEP) can use free-energy changes.
Binary descriptors can note molecule-substructure presence or absence {Free-Wilson approach}.
Linear property sets can have different values, change values by crossing-over between related such genes, and have random change {Genetic Function Algorithm} (GFA), to select best fit.
Values {Hammett sigma value} can relate to electronic and electrostatic properties.
Activity, partition coefficients for hydrophobicity, ionization degree, and molecular size relate {Hansch equation}.
Variables {latent variable} can be linear-descriptor combination.
Supervised methods {linear discriminant analysis} (LDA), in which boundary surface minimizes region variance and maximizes variance between regions, can put compounds into groups by activity level.
log K = k1 * sigma + k2 {linear free energy equation, drug} (LFE).
Supervised methods {linear learning machine} (LLM) can divide n-dimensional space into regions, using discriminant function.
Factor-analysis methods {maximum-likelihood method} can find factors.
Metric or non-metric methods {multidimensional scaling} (MDS) can analyze similarity or dissimilarity matrices to find dimension number and place objects in proper relative positions.
Non-parametric methods {multivariate adaptive regression spline} (MARS) can find factors.
Models {Mutation and Selection Uncover Models} (MUSEUM) can add and subtract randomly selected variables, with no crossing-over, and evaluate for "fitness" or best fit.
Unsupervised linear methods {non-linear iterative partial least-squares} (NIPALS) can represent data as product of score matrix, for original observations, and loading-matrix transform, for original factors.
Topological mappings {non-linear mapping} (NLM) can be factor-analysis methods in which linear-variable combinations make two or three new variables.
Information about compound physico-chemical properties can predict compound chemical or physiological behavior in vitro and in vivo {predictive computational model}.
Variables {principal component} (PC) can be linear-descriptor combinations. Unsupervised linear method {principal component analysis, factor} (PCA) represents data as product of score matrix, for original observations, and loading-matrix transform, for original factors. PCA is factor-analysis method in which linear variable combinations make two or three new variables. PCA reduces unimportant variables.
Singular-value decomposition (SVD) can find best singular values for predicting {principal component regression} (PCR). SVD projects regression to latent structures.
Modified PCA {principal factor analysis} can find principal factors.
Methods {Procrustes analysis} can identify descriptor sets for describing similarity.
Methods {QR algorithm} can diagonalize matrices.
Unsupervised linear methods {rank annihilation} can find factors.
Residual variance approaches constancy {Scree-test, drug}, and plotted slope levels off {Scree-plot}, depending on component number.
In unsupervised linear methods {singular value decomposition, drug} (SVD), correlation matrix is product of score, eigenvalue, and loading matrices, with diagonalization using QR algorithm.
Factor-analysis methods {spectral mapping analysis} (SMA) can first take data logarithm to eliminate outliers and then subtract means from rows and columns, to leave only variation, showing which variables are important and how much.
Spaces {structure space} can have two or three principal components.
Methods {target-transformation factor analysis} can rotate features to match known pattern, such as hypothesis or signature.
Factors and response variable have relations {Unsupervised Method}, without using factor information or predetermined models.
Designs {factorial design} can try to ensure design-space sampling, if position varies.
Designs {fractional factorial design} can try to ensure design-space sampling, if position varies.
Three-level designs {response surface method} (RSM) can have three factors that quantify relationships among responses and factors. RSM includes MLR, OLS, PCR, and PLS linear designs; non-linear regression analysis (NLR); and non-parametric methods, such as ACE, NPLS, and MARS.
isomer-enumeration method {Cayley tree structure}.
Isomer-enumeration methods {CONGEN program} can be successors to DENDRAL.
Isomer-enumeration methods {DENDRAL program} can be forerunners of CONGEN.
isomer-enumeration method {Henze and Blair recursion formulas}.
Isomer-enumeration methods {Polya's enumeration theorem} {Polya enumeration theorem} can use group theory.
Electron orbitals {molecular orbital} can be for whole molecule.
Analyses {ab initio analysis} can use all electrons.
Adding atomic orbitals can approximate molecular orbitals {linear combinations of atomic orbitals} (LCAO).
Semiempiric methods {perturbative configuration interaction using localized orbitals} (PCILO) can use perturbations.
Analyses {semiempiric} can use valence electrons and parameterize core electrons.
Sigma electrons can contribute {simple delta index, drug}.
Factors, properties, or structures {regressor} can contribute to response values {regression, regressor} {Regression Analysis}.
Regression can project to latent structures {canonical correlation} (CC), to put compounds in classes.
Regression {continuum regression} (CR) can project to latent structures, to put compounds in classes.
Variance-covariance matrix {correlation matrix, drug} can scale to normalize data.
Regression can project to latent structures {kernel algorithm}, to put compounds in classes.
Methods {matrix diagonalization, drug} can simplify data variance-covariance matrix.
Parametric methods {non-linear regression} (NLR) can find descriptor coefficients by non-linear regression.
Regression can project to latent structures {ridge regression} (RR), to put compounds in classes.
Methods {Spearman rank correlation coefficient} can measure molecular similarity.
Complete, symmetric, square matrix {variance-covariance matrix} uses property values and structure values.
Regression can project to latent structures {adaptive least-squares} {ALS algorithm}, to put compounds in classes.
Methods {classical least-squares, drug} (CLS) can be the same as ordinary least-squares analysis.
Partial least-squares {Comparative Molecular Field Analysis} (CoMFA) can analyze grid around site atom and find grid-point electrostatic and steric interactions, to make sampled-point descriptors.
Compounds have different classes with different weights {fuzzy adaptive least-squares} (FALS).
Methods {Generating Optimal Linear PLS Estimations} (GOLPE) can use PLS and D-optimal design to select variables, and cross-validates.
Fitting methods {inverse least-squares} (ILS) can find regression line.
Methods {least-squares regression, drug} can be the same as ordinary least-squares analysis.
Methods {linear least-squares regression, drug} can be the same as ordinary least-squares analysis.
Partial least-squares methods {matrix bidiagonalization method, drug} can simplify data variance-covariance matrix.
Regression can project to latent structures {multi-block PLS}, to put compounds in classes.
Methods {multiple least-squares regression, drug} can be the same as ordinary least-squares analysis.
Methods {multiple linear regression} (MLR) can measure linear component dependence on physico-chemical or structural properties and finds descriptor coefficients.
Methods {multivariate least-squares regression, drug} can be the same as ordinary least-squares analysis.
Methods {non-least-squares} (NLS) can detect non-linear relationships.
Fitting methods {ordinary least-squares} (OLS) can find descriptor coefficients.
Methods {partial least-squares} (PLS) can use least-squares to find independent variables and dependencies among variables. It projects regression to latent structures. It maximizes latent-variable and observable covariation. It diagonalizes the matrix.
Methods {SAMPLS algorithm} can apply PLS to trend vector analysis.
Estimates {best linear unbiased estimator} (BLUE) can give smallest variance among estimators.
Error measures {standard error} can be square root of MSE.
SSE, SSR, or SST {sum of squares of differences} {squares of differences sum}.
SSE / (observation number + factor number - 1) {mean square error} (MSE).
Errors or residuals can cause sum {SSE} of squares of differences between observed and predicted responses.
Regression can cause sum {SSR} of squares of differences between observed and mean.
Sum {SST} of squares of differences between predicted and mean makes total: SST = SSE + SSR.
5-Chemistry-Biochemistry-Drug-Activity
Outline of Knowledge Database Home Page
Description of Outline of Knowledge Database
Date Modified: 2022.0225