Sci Comp

Pure TypeScript library of numerical methods for the Datagrok platform.

Modules

• Optimization:

  • single-objective (detailed docs):
    • Nelder-Mead - derivative-free simplex method
    • PSO (Particle Swarm Optimization) - stochastic population-based method
    • Gradient Descent - first-order method with numerical gradients, momentum, and learning rate decay
    • Adam - adaptive moment estimation with per-parameter learning rates
    • L-BFGS - limited-memory quasi-Newton method with two-loop recursion and Armijo line search
    • L-BFGS-B - limited-memory BFGS with native box constraints, Moré–Thuente strong-Wolfe line search, compact representation, and Morales–Nocedal 2011 subspace refinement
  • multi-objective:
    • MOEA/D

• Time Series:

  • extraction of 45 features per value column (statistics, trend, complexity, threshold-based metrics)

• Statistics (detailed docs):

  • two-group comparison: Welch's t-test, Mann-Whitney U (exact + asymptotic), Hedges' g effect size
  • rank correlation: Spearman ρ, severity-trend wrapper
  • multi-group vs control: Welch pairwise + Bonferroni, Dunnett's many-to-one
  • trend tests: Jonckheere-Terpstra (continuous), Cochran-Armitage (incidence, with Buonaccorsi-modified variant)
  • sequential threshold test for proportions (Young 1985)
  • categorical: Fisher 2×2 exact (two-sided minlike + one-sided + odds ratio), Boschloo's unconditional exact (uniformly more powerful)
  • dose-response: PAVA isotonic regression, Williams step-down with 1971/1972 critical-value tables
  • covariate-adjusted analysis: ANCOVA (LS means, slope homogeneity, effect decomposition)

• NCA (Non-Compartmental Analysis):

  • 8 PK parameters: Cmax, Tmax, AUClast, AUCinf, %AUCextrap, λz, t½, CL, Vz
  • 3 AUC methods (linear, log-linear, linear-up/log-down) × naive Float64 + Neumaier-compensated summation
  • 4 BLQ-handling rules × 4 phases
  • λz auto best-fit (subset search by adjusted R²) + manual
  • IV bolus c0 back-extrapolation (logslope / c1 / cmin / set0 chain), extravascular pre-dose insertion
  • validated against PKNCA on 26 reference profiles (theophylline, indomethacin, synthetic rat)

Installation

npm install @datagrok-libraries/sci-comp

Quick start

Optimization

Minimize the Rosenbrock function with Nelder-Mead:

import {singleObjective} from '@datagrok-libraries/sci-comp';

const rosenbrock = (x: Float64Array): number => {
  let sum = 0;
  for (let i = 0; i < x.length - 1; i++)
    sum += 100 * (x[i + 1] - x[i] ** 2) ** 2 + (1 - x[i]) ** 2;
  return sum;
};

const nm = new singleObjective.NelderMead();
const result = nm.minimize(rosenbrock, new Float64Array([-1.2, 1.0]), {
  maxIterations: 5_000,
  tolerance: 1e-12,
});
// result.point ≈ [1, 1], result.value ≈ 0

See single-objective optimization docs for

• constrained optimization
• implemented methods
• sync and async objectives
• callbacks
• registry
• benchmarks
• examples

Time-series feature extraction

Extract tsfresh-compatible features from time-series data:

import {timeSeries} from '@datagrok-libraries/sci-comp';

const df: timeSeries.featureExtraction.TimeSeriesDataFrame = {
  ids: new Uint32Array([1, 1, 1, 1, 1]),
  time: new Float64Array([0, 1, 2, 3, 4]),
  rowCount: 5,
  columns: [{
    name: 'temperature',
    data: new Float64Array([20.1, 20.5, 21.0, 20.8, 21.3]),
  }],
};

const result = timeSeries.featureExtraction.extractFeatures(df);
// result.nSamples = 1, result.columns.length = 45
// e.g. temperature__mean, temperature__standard_deviation, temperature__linear_trend__attr_slope, ...

See feature extraction docs for

• full feature list (45 features across 11 categories)
• input/output format
• multiple samples and multiple columns
• validation
• naming convention
• examples

Statistics

Run Welch's t-test on two samples with NaN handling:

import {stats} from '@datagrok-libraries/sci-comp';

const a = [69, 70, 66, 63, 68, 70, 69, 67, 62, 63];
const b = [68, 62, 67, 68, 69, 67, 61, 59, 62, 61];

const r = stats.welchTTest(a, b);
// r.statistic ≈ 1.511, r.pValue ≈ 0.149

Or compare each treated group to a control with Dunnett's family-wise-error-controlled test:

const control = [7.40, 8.50, 7.20, 8.24, 9.84, 8.32];
const treated = [
  {doseLevel: 1, values: [9.76, 8.80, 7.68, 9.36]},
  {doseLevel: 2, values: [12.80, 9.68, 12.16, 9.20, 10.55]},
];

const dun = stats.dunnettPairwise(control, treated);
// dun[1].pValueAdj ≈ 0.0058  → drug B differs from control at α = 0.05

See statistics docs for

• full method list (14 tests across 8 domains)
• input types (number[], Float32Array, Float64Array, Int32Array, …)
• NaN handling (NaN as missing-value sentinel, stripped per-method)
• worked examples reproducing published references (Dunnett 1955, NIST Iris, Williams 1971/1972, Young 1985, Montgomery 15.10 vs SAS PROC GLM)
• validation against scipy via JSON fixtures (179 cases, 14 fixture files)

NCA

Run the full Non-Compartmental Analysis pipeline on one PK profile:

import {nca} from '@datagrok-libraries/sci-comp';

const inputs: nca.ProfileInputs = {
  time:    new Float64Array([0, 0.25, 0.5, 1, 2, 4, 8, 12]),
  conc:    new Float64Array([0, 1.5,  2.4, 3.0, 1.8, 0.9, 0.3, 0.1]),
  blqMask: new Uint8Array(8),
  lloq:    0.05,
  dose:               2.5,
  doseUnits:          'mg',
  concentrationUnits: 'mg/L',
  timeUnits:          'h',
  route:              nca.ROUTE_PO,
  infusionDuration:   null,
  bodyWeight:         null,
};

const rules: nca.NcaRules = {
  aucMethod: 'linear-up-log-down',
  blq: {
    preFirstMeasurable: 'set-zero', embedded: 'set-zero',
    afterLast: 'set-zero', consecutiveAfterLast: 'set-zero',
  },
  lambdaZ: {
    mode: 'auto-best-fit',
    minPoints: 3, minRSquared: 0.85, excludeCmax: true,
    adjRSquaredFactor: 1e-4,
  },
  extrapWarnPct: 20, extrapErrorPct: 50,
  compensatedSummation: false,
};

const result = nca.computeNca(inputs, rules);
// result.values: Cmax, Tmax, AUClast, AUCinf, lambdaZ, halfLife, cl, vz, pctExtrap
// result.provenance: lambda_z fit details, BLQ trace, AUC method, warnings
// result.status: 'ok' | 'partial' | 'failed'

See NCA docs for

• algorithm details (BLQ phasing, lambda_z best-fit search, IV bolus c0 back-extrapolation)
• parameter contract (ProfileInputs, NcaRules, ComputeResult)
• validation against fixtures (26 profiles across 3 datasets, all within §9.2 tolerances)