Endpoint curation
Separate assay types, units, qualifiers, targets, species, and incompatible measurements before modeling.
Target-bound predictive modeling
Build a model whose limits travel with every prediction.
Curate target-specific activity evidence, document endpoint meaning, compare validation results, define applicability, preserve model versions, and expose uncertainty when predictions enter screening and optimization workflows.
QSARDataset curationModel validationApplicability domainConformal intervalsModel registry
A model is more than an algorithm
Keep endpoint, data, validation, and intended use attached.
MolexIO treats each target-bound model as a versioned evidence object. The source data, curation decisions, molecular representation, split strategy, metrics, applicability, uncertainty method, and intended use remain available when the model scores new candidates.
Molecular identity
Standardize structures, salts, duplicates, stereochemistry, and linked activity records with explicit provenance.
Validation design
Compare cross-validation with scaffold, temporal, or held-out evaluation appropriate to the intended prediction problem.
Applicability domain
Report whether a candidate is supported by relevant training chemistry rather than returning an unqualified number.
Uncertainty
Attach supported intervals or confidence evidence and distinguish model uncertainty from assay variability.
Model registry
Version models with target, dataset, performance, status, visibility, and run-time selection context.
Target-bound modelsA model cannot silently transfer to an unrelated endpoint or target.
Validation matched to useMetrics are interpreted through the split and intended prediction setting.
Out-of-domain visibleUnsupported chemistry is flagged instead of presented with false confidence.
From assay table to deployable evidence
Make each modeling choice reproducible and reviewable.
- 01
Curate
Resolve target and molecular identity, endpoint definition, units, qualifiers, duplicates, and assay comparability.
- 02
Train
Build supported representations and models under a versioned, reproducible configuration.
- 03
Validate
Review predictive error, split-specific behavior, coverage, leakage risk, and applicability.
- 04
Register and monitor
Use qualified versions in workflows and compare future assay feedback with the original prediction.
Scientific boundary
A strong retrospective metric does not guarantee prospective utility.
QSAR performance can be inflated by duplicates, analog leakage, narrow response range, inconsistent assays, biased splits, or test chemistry too similar to training data. Predictions are conditional on a specific endpoint and applicability domain. Prospective testing and monitoring are required before a model is trusted for high-impact decisions.
Questions before use
QSAR model FAQ
How is prospective performance different?
It evaluates compounds and experiments not available during model development and is more relevant to future use than training fit alone.
What is an applicability domain?
It identifies the chemical and response space where the model has adequate support.
Does a QSAR prediction establish potency?
No. It is an endpoint-specific estimate that requires experimental confirmation.
Make predictive models accountable to their evidence.
Connect every candidate prediction to the target, dataset, validation, applicability, uncertainty, and version that produced it.