Different strategies read different evidence: measured contacts, co-expression, a trained model. Where they agree, a call is far more likely to be right -- and the self-tests say how much more.
from starplast import strategies as Svote = S.run('layer_vote', target='compartment')vote.tables['source weights']
| source | weight |
|---|---|
| measurements (kNN) | 0.275 |
| xlms | 0.122 |
| coexpression | 0.116 |
| struct | 0.0745 |
| domain | 0.0519 |
| cotranslation | 0.0268 |
| cofitness | 0.0174 |
| ip_ms | 0.00331 |
| orthogroup | 0 |
clf = S.run('supervised_classifier', target='compartment')clf.tables['what drives each class'].groupby('class').head(2).head(12)
| class | measurement | weight |
|---|---|---|
| 19S proteasome | n_crosslink_partners | 1.39 |
| 19S proteasome | n_xlink_partners | 1.29 |
| 40S ribosome | mrna_log2_remaining_4h_actinomycin | 1.16 |
| 40S ribosome | ees_vs_tachyzoite_log2 | 0.82 |
| 60S ribosome | fit_glucose_dependence | 1.16 |
| 60S ribosome | codon_gc3 | 1.14 |
| ER | is_tm | 2.2 |
| ER | n_tm | 2.1 |
| ER 2 | iron_depletion_protein_log2fc | 1.76 |
| ER 2 | has_signal_peptide | 1.68 |
| Golgi | is_tm | 1.85 |
| Golgi | n_tm | 1.8 |
tri = S.test('triangulation', target='compartment')tri
PASS -- precision of calls on hidden genes 0.539 against 0.157 ± 0.011 under 5 runs on shuffled labels (bar 0.171, p 0.167); effect +0.382, needed +0.100; 622 hidden
Hidden: 25% of the compartment labels, whole orthogroups at a time; 622 calls made on 951 hidden genes. Null: 5 runs on shuffled labels.
Scorecard (label calls)
| Accuracy | 0.352 |
| Coverage | 0.654 |
| Precision of calls | 0.539 |
| Macro precision | 0.533 |
| Macro recall (balanced accuracy) | 0.331 |
| Macro F1 | 0.369 |
| Weighted F1 | 0.405 |
| Cohen's kappa | 0.300 |
| Matthews correlation (MCC) | 0.323 |
| Macro AUROC | 0.920 |
| Macro AUPRC | 0.503 |
{k: round(v, 3) for k, v in tri.numbers.items() if 'precision' in k}{'measurements (kNN) precision': 0.406, 'classifier (logistic) precision': 0.461, 'networks (vote) precision': 0.286}under = S.run('understudied_first', target='compartment')under.tables['candidates'].head(10)
| gene_id | product | prediction | support | sources_agreeing |
|---|---|---|---|---|
| TGME49_225745 | hypothetical protein | nucleus - chromatin | 0.546 | 3 |
| TGME49_266010 | phosphatidylinositol 3- and 4-kinase | nucleus - chromatin | 0.546 | 3 |
| TGME49_286270 | hypothetical protein | nucleus - chromatin | 0.546 | 3 |
| TGME49_271145 | hypothetical protein | nucleus - chromatin | 0.546 | 3 |
| TGME49_253870 | hypothetical protein | apicoplast | 0.546 | 3 |
| TGME49_311625 | bromodomain-containing protein BDP6, putative | nucleus - chromatin | 0.546 | 3 |
| TGME49_270720 | hypothetical protein | nucleus - chromatin | 0.546 | 3 |
| TGME49_225960 | STE kinase | nucleus - chromatin | 0.546 | 3 |
| TGME49_314875 | RIC1 protein | nucleus - chromatin | 0.546 | 3 |
| TGME49_273780 | SWI2/SNF2-containing protein | nucleus - chromatin | 0.546 | 3 |