Explore, predict and compare

The Tools menu has three guided entry points. Advanced map and model controls remain in the analysis dock. Each workflow uses the organism currently open in the main window.

The 0.43 benchmark compares the prediction methods on held-out localization and protein abundance measurements, including source ablations and shuffled-label controls.

Explore a gene

Search an accession, symbol or product description. Double-click a match to select it on the map and inspect its measurements, missing values, evidence status and source links. A gene-table value is often an aggregate: it does not recover the original replicate, time point or uncertainty when those details were not kept. Legacy summaries without an explicit evidence classification say unclassified.

Predict a trait

Choose an observed outcome, its type, a method and the grouping used for evaluation. Orthogroup keeps related proteins together; random gene folds answer an easier question and may overestimate transfer to new families. A study or condition column can be used when it actually identifies the intended held-out groups. Starplast does not invent source dates or study membership for a time/study holdout.

The linear and prior models are useful starting points. Boosted trees, feature neighbours, PCA neighbours, UMAP neighbours and masked factors provide comparisons. The UMAP predictor fits a separate representation inside each training fold; it does not classify the opening three-dimensional map. Masked factors are a simple iterative PCA completion baseline, not an implementation of MOFA+.

For Toxoplasma, Include frozen protein sequence features loads the bundled ESM-2 representations. It does not download a model. Local AF3 features are already present where an exact mapping was available. Sequence or structure evidence is not automatically an improvement; evaluate the additions on the same held-out genes.

Evaluate and predict runs in the shared Jobs panel. Feature exclusions, imputation, scaling and dimensionality reduction are fitted on training genes. The target, its registered relatives and derived inputs are excluded. Labels in the calibration and evaluation groups are not used for fitting those transforms. The Stop button requests cancellation between model fits.

Read the held-out metrics before the candidate table. An unlabelled gene is not a negative example. A gene with insufficient measured features gets no supported call. Classification can also abstain below the selected model-probability cutoff. The cutoff is a declared setting, not a threshold optimized on the test labels.

Classification reports fixed-label accuracy, macro F1, per-class precision/recall and average precision, Brier score and log loss. Unsupported calls count as misses in accuracy and recall; selective accuracy is separately named. Probabilities use temperature scaling on reserved groups when enough calibration labels exist. The table explicitly says when calibration was unavailable. Even a calibrated model can become unreliable on a less-studied population.

Regression reports RMSE, MAE, R², rank correlation and, when available, intervals estimated from reserved-group residuals. Their observed coverage is reported; family or domain shift prevents an unconditional coverage guarantee. Multi-label tasks are available through the API as separate measured binary columns with unknown entries preserved.

Export full run writes predictions.csv, per_class.csv and run.json. The JSON records settings, input identity, exact training/calibration/test gene IDs, excluded features, actual estimators and package versions. Results remain model hypotheses. Comparing methods here is development validation; evaluating a selected winner prospectively requires new independent data.

Compare a screen

Open a gene-level CSV/TSV from spaCR or another measurement pipeline. Choose the gene identifier and measured-value columns. Identifiers are resolved through the current table and its identity catalogue. Ambiguous/unresolved identifiers remain visible; missing or nonnumeric values stay missing, and a measured zero stays zero.

Repeated gene rows, including aliases that converge on one gene, require explicit aggregation before import. Starplast does not silently average guide-level results. The exported comparison includes the existing gene evidence and a JSON record of the source filename, SHA-256 hash and selected columns. Use the advanced importer when you want to add measurements to an analysis table.

Evidence and literature records

starplast.evidence.Observation stores the biological entity, trait, typed value, unit, source/version/location, context, replicate, uncertainty, evidence status, missingness state and derivation. Gene, transcript, protein, residue, allele, pair, host-gene and metabolite identities stay distinct. Contradictory observations from different studies or contexts remain separate content-addressed records.

from_gene_table() wraps existing summaries with their available provenance; it does not claim to reconstruct raw experiments. Write source-level observations directly when loading original tables. Missingness distinguishes unassayed, below-detection, failed-QC, ambiguous and unknown values.

propose_assertions() offers a conservative phrase extractor for source passages. Its output is always pending review, retains the exact passage and negation, and omits sentences with ambiguous gene mappings. A named reviewer must accept a claim before LiteratureAssertion.observation() can expose it as a reviewed literature claim. Acceptance confirms the interpretation of a passage, not experimental truth. Store and review these records through the API; the desktop does not yet contain a bulk literature-curation interface.

Networks and explanations

The network_prediction.run_network() API evaluates weighted propagation on a fixed graph with explicit gene order and edge provenance. Seed labels, layer-weight validation, probability calibration and outer evaluation use disjoint groups. The guided method Weighted networks (fixed graph) uses the bundled domain, coexpression and structural-similarity layers with those builders' declared inputs. Target-derived layers are excluded; disconnected genes abstain. This is explicitly transductive because the fixed graph includes evaluation nodes. It cannot be reported as performance on newly arriving nodes. Custom network fitting uses the API so its source declarations remain explicit. Sequence-feature and measured-feature coverage controls do not apply to network propagation.

prioritization.explain_candidates() reports observed source coverage, training neighbours and, for linear models, feature contributions. It explains unknown-gene calls using the final fitted model; it does not leak that model into held-out evaluation explanations. Model contributions describe association, not mechanism.

prioritization.prioritize() is a generic budget heuristic for separately calibrated classification results, using explicit cost, benefit and an optional uncertainty bonus. It records default assumptions and can limit selection to one item per supplied group. It is not an optimal budget solver, an information-gain model or a prospectively validated experimental design system.