QSAR AI & MLR Analyzer

Interpretable MLR, small-data machine learning, repeated external validation, and uncertainty quantification

Step 1: Upload Data

Full Compound Set (CSV or SD/SDF)

Or Upload Existing Training Set

Training Data (CSV or SD/SDF)

Select Target Variable

Data Clean-up / Compound Screening

Pre-model data clean-up

Preview questionable compounds before splitting or modelling. Invalid target and low descriptor-completeness rows can be removed; activity outliers and duplicate IDs are flagged unless you choose to remove them.

Minimum descriptor completeness (%)

Activity outlier IQR multiplier

Remove missing/non-numeric target

Remove low descriptor-completeness rows

Remove duplicate compound IDs/names except first occurrence

Remove activity outliers by IQR rule

Split Full Set

Training / Test Split

Select Descriptors (Features)

Exhaustive search options

For very large descriptor tables, Auto mode first ranks descriptors by target correlation, removes near-duplicate descriptors, then exhaustively tests the remaining candidate pool. Strict all-eligible mode is only practical for small descriptor sets. Long exhaustive searches now run in small background chunks so the page remains responsive.

Validation settings

Use a fixed seed for reproducible research. Stratified folds distribute low, medium and high activity values across CV folds. Descriptor standardization uses training-set means and SDs only; during CV, each training fold has its own scaling parameters to avoid data leakage. Repeated Split Validation repeatedly creates stratified external holdout sets from the full compound set using the checked descriptors, which helps answer reviewer concerns about reliance on one fixed 37/9 split.

Machine-learning benchmark

Models are tuned using only the training data. Reported cross-validation performance is nested, while external test performance uses a final model refitted on the full training set. Gaussian process regression provides a modern probabilistic small-data model; CV-conformal intervals add model-agnostic uncertainty estimates. Deep neural networks are intentionally excluded because very small QSAR datasets do not support defensible end-to-end training.

Result safety: every AI/ML run is tied to its exact target and descriptor signature. Changing the target, checked descriptors, or AI feature context archives and clears incompatible AI results. PDF export is disabled while an AI/ML run is active.

Recommended manuscript experiment: use the same descriptors and the same repeated 37/9 partitions for MLR, Ridge, PLS, SVR, Gaussian process, kNN and Random Forest. Compare typical external accuracy, variance and how often each method ranks first.

Fully nested QSAR validation

MLR mode: descriptor eligibility, target correlation ranking, near-duplicate removal, subset selection, VIF calculation, scaling and MLR fitting are repeated using only each outer training fold. PLS mode: candidates are ranked by correlation only (no decorrelation or VIF screening, since PLS tolerates correlated predictors directly), and the number of latent variables is chosen by inner cross-validation (smallest count within 0.03 of the maximum inner Q²). One out-of-fold R² is calculated from all compounds per outer repeat; R² values from individual small folds are not averaged. Load the complete compound set for the intended publication analysis. A candidate-pool size above 14 is only usable with Algorithm = PLS (MLR silently caps back to 14, since it exhaustively enumerates every subset and larger pools are computationally infeasible in the browser); a very wide candidate pool relative to an outer training fold's size can occasionally make one inner cross-validation fold numerically unstable for PLS too, in which case that outer fold is skipped and reported as "no stable inner model" rather than guessed. Leave-one-series-out outer folds reuse the series/group column selected below in the Leave-One-Series-Out panel. Because that outer partition is fixed by series membership rather than reshuffled, "repeated outer cross-validations" instead becomes a seed-sensitivity sweep: the same series are held out every repeat, but the inner cross-validation used to pick descriptors (MLR) or the component count (PLS) uses a different seed each repeat.

Important: the existing AI/ML Nested CV tunes algorithms for a fixed descriptor set. This module additionally nests descriptor (or PLS component-count) selection and therefore evaluates the complete Robust-QSAR-to-MLR (or Robust-QSAR-to-PLS) pipeline, optionally with series-based outer folds instead of k-fold.

Phase 4 locked-protocol search & Leave-One-Series-Out (manuscript-lock compliance)

Locked protocol settings (not editable): pairwise correlation prefilter |r| ≥ 0.90 (Section 1), near-zero-variance exclusion at ≥90% identical values (Section 1, applied automatically to the descriptor checklist above), subset sizes 2-6 with hard VIF ≥ 5 exclusion before ranking (Section 2), stratified 5-fold CV with the fixed seed above (Section 3), mechanical selection by max Q² minus a 0.03 margin, then smallest size, then lowest VIF, then lowest RMSE (Section 5).

Leave-One-Series-Out validation (Section 4)

Trains on all other series and predicts the held-out series, one series at a time, using the currently checked descriptors. Each series is reported individually, as required by the locked protocol; it is not averaged into a single pooled figure. If your data has no dedicated series/scaffold column, keep the default "Auto" option: it recovers series membership from the text before a trailing "-number" in the compound ID (e.g. "EJMC2018-18" and "EJMC2018-31" both become series "EJMC2018"). Do not point this at a numeric descriptor column — grouping by a continuous value produces meaningless one-off "series" of coincidentally similar compounds. The PLS model option refits PLS (instead of MLR) on the remaining series using the same checked descriptors as latent-variable inputs; the number of components is chosen by 5-fold cross-validation on the remaining series only (smallest count within 0.03 of the maximum CV Q²), never using the held-out series.

QSAR method guide: Use Robust QSAR for interpretable descriptor selection, then compare the final descriptor set with the Small-Data AI / ML Benchmark and Repeated AI / ML Robustness. Nested CV, repeated external partitions and conformal uncertainty should be interpreted together.

Glossary and method notes

Q²ext / Q²F1: external predictive squared correlation using the training-set mean as denominator.
Q²F2: external predictive squared correlation using the external test-set mean as denominator.
Q²F3: normalized external prediction error compared with training-set target variance.
VIF: variance inflation factor; values above 5 suggest multicollinearity, above 10 are severe. Near-perfect descriptor collinearity can make VIF approach infinity.
Condition #: infinity-norm condition estimate for X'X; very large values suggest unstable coefficients.
Descriptor standardization: z-score scaling of each descriptor using the training-set mean and standard deviation; this improves coefficient comparability and often reduces numerical ill-conditioning caused by different descriptor scales.
Standardized β: raw coefficient × SD(descriptor) / SD(activity); use |β| to compare relative descriptor influence.
h*: leverage warning threshold used in Williams applicability-domain diagnostics. Test compounds with h > h* are extrapolations.
Y-randomization: checks whether good performance may arise by chance after target permutation.
Recommended workflow: set a fixed CV seed, use Robust QSAR or Elastic Net for high-dimensional data, use Exhaustive R² only for small filtered candidate pools, then build the final MLR model and export the model/predictions.

Upload training data first

Step 2: Upload Test Set

Test Data (CSV or SD/SDF, optional after full-set split)

Build & Validate Model

Run History

Completed model and validation runs are saved automatically in this browser. The history stores settings and summary metrics, not uploaded compound rows or structures. It remains available after Reset All or reopening the page from the same browser profile and URL/location.

Automatically save completed runs 0 saved runs Storage: this browser/device only

Import History JSON

No run has been saved yet.

Date / time	Run	Target and data	Features	Key result	Seed	Actions
No saved runs.

Training Results

R^2 (Train)

RMSE (Train)

MAE (Train)

5-Fold Cross-Validation

R^2 (5-CV)

RMSE (5-CV)

MAE (5-CV)

Leave-One-Out CV

R^2 (LOO)

RMSE (LOO)

MAE (LOO)

Test Set Validation

R^2 (Test)

RMSE (Test)

MAE (Test)

QSAR Diagnostics

Q²ext

RMSE Test/Train

Max VIF

Q²F2

Q²F3

Condition #

Build a model to see diagnostics and warnings.

Build a model to see standardized descriptor coefficients.

Fully Nested QSAR Validation

Load the complete compound set, select the target, then run fully nested validation. Descriptor selection will be repeated independently inside every outer training fold.

Complete repeats	OOF R², mean ± SD	Median R²	R² range	Mean RMSE	Mean MAE	Successful outer models
No fully nested validation has been run.

Complete out-of-fold performance by repeat

Repeat	Status	Predicted n	OOF R²	OOF RMSE	OOF MAE	Successful folds	Failed folds
No fully nested validation has been run.

Descriptor-selection stability across outer models

Rank	Descriptor	Selected outer models	Selection frequency
No descriptor-stability result is available.

Small-Data AI / Machine-Learning Benchmark

Run the Small-Data AI / ML Benchmark after selecting descriptors or use top correlated descriptors.

Algorithm	Features	Setting	Train R²	Nested-CV Q²	Nested-CV RMSE	Test R²	Test RMSE	Conformal coverage	Mean interval width	Top descriptor importance
No AI/ML benchmark has been run.

Repeated AI / ML Robustness and Ranking Stability

Run Repeated AI / ML Robustness to compare algorithms on the same stratified external partitions. Hyperparameters are selected inside each training split.

Algorithm	Successful splits	Mean Test R² ± SD	Median	Min–Max	Mean RMSE	R² ≥ 0.70	R² ≥ 0.60	Ranked first
No repeated AI/ML analysis has been run.

External Test R² across repeated splits

Every algorithm is evaluated on the same external partition in each repeat. The frequency of ranking first directly tests whether apparent algorithm superiority survives resampling.

Repeated Split Robustness

Run Repeated Split Validation after selecting descriptors. This tests how stable external prediction is across many stratified train/test partitions.

Repeated split external Test R² distribution

Fixed descriptor set evaluated over repeated stratified external splits. Use this as robustness evidence beside the single fixed external test result.

Model Performance Comparison

Training: Actual vs Predicted

Test: Actual vs Predicted

Residuals Distribution (Training)

Williams Applicability-Domain Plot

Leverage h versus standardized residual. The vertical line marks h*, and horizontal lines mark ±3 residual limits. Use the PNG/SVG export buttons for manuscript figures instead of cropping the PDF report.

Y-Randomization Plot

Run Y-randomization to plot randomized Q² values against the original model Q². Use the PNG/SVG export buttons for manuscript figures instead of cropping the PDF report.

Test Predictions

Actual	Predicted	Residual	Error %	95% prediction interval