SubcellPert / benchmark /protocols /context_ood_joint.yaml
king159's picture
init
a57b684
Raw History Blame
4.34 kB
schema_version: 1
id: context_ood_joint
generalization_axis: cellular_context
evaluation_mode: joint
name: Cellular-context OOD joint RNA abundance and localization prediction
status: planned
data_tier: strict
observation_unit: context_transition_population
paired_counterfactual: false
required_columns:
- dataset_id
- unit_id
- sample_id
- condition
- cell_id
- cell_type
- gene
- x
- y
split:
group_key: context
task_manifests:
context: benchmark/tasks/context_ood_lung.yaml
expansion: benchmark/tasks/context_ood_expansion.yaml
biological_unit: benchmark/tasks/context_ood_biological_unit.yaml
reaudited: benchmark/tasks/context_ood_reaudited.yaml
reference: scPerturBench-inspired Cellular_context_generalization/o.o.d.; local
full-vocabulary metrics are numerically non-equivalent
rule: jointly train the preregistered known-transition family on all non-test contexts
and hide every target response assigned to the held-out context group
fit_group: independent checkpoint unit; one or more response folds may share it
source_cells_in_test_context_available: true
source_and_target_population_support_gate: source enforced before inference; sealed
target enforced by evaluator
executable_context_keys:
- cell_type
- patient_id
- donor_id
transition_specific: false
transition_family_policy: frozen per fit group
biological_unit_role: leakage_audit_and_clustered_uncertainty
prohibited:
- held_out_context_target_cells_in_fit
- held_out_context_target_rna_in_fit
- target_population_size_at_inference
count_policy:
mode: predicted_distribution
zeros: include every declared source cell without the evaluated gene
target_cell_count_used: false
source_population: held-out-context source-condition cells
intervention_descriptor:
role: known_intervention_covariate_not_an_OOD_identity_prior
applies_to: learned_baselines_only
source: task-bound known-condition descriptor, either a frozen design-only artifact
or deterministic one-hot identity
target_response_data_used: false
foundation_context:
conditioning: source_only
source_cell_policy: all_declared_source_cells_no_subsampling
target_response_data_used: false
coordinate_policy:
eligible_semantics:
- decoded_molecule
- capture_site
capture_site_feature_types:
- dnb
- bead
- polony
- sequencing_cluster
- grid
event_representation: unit_event_rows
unit_event_value: 1
capture_site_gate: deterministic pre-normalization expansion plus artifact hashes,
global lineage, and exact sample/unit/condition/cell/gene reconciliation; no frozen
capture-site importer yet
aggregation: coordinate semantics and capture feature type are explicit non-pooled
strata
required_axes:
- x
- y
optional_axes:
- z
frame: cell_centric
factorization: predict_counts_then_localization
cell_weighting: equal_weight_per_nonempty_cell
metrics:
primary: centroid_rmse_log1p
secondary:
- source_delta_pearson
- marginal_wasserstein_log1p
- multivariate_sliced_wasserstein_log1p
- dependency_rmse
- cell_balanced_spatial_swd
- spatial_target_gene_coverage
ranking: Count and spatial columns remain separate. Report unit and gene coverage;
no cross-dataset composite. Biological-unit uncertainty and global aggregation
are subsequent analyses.
uncertainty:
implementation_status: planned_before_final_reporting
resampling_unit: biological_unit
single_unit_policy: NA_with_unit_coverage
prohibited_pseudoreplicates:
- genes
- cells
- molecules
- folds
- model_seeds
eligible_datasets:
- gse291460_spinal_ko_xenium
- gse269719_kidney_iri_xenium
- gut_microbiota_merfish
- gse280376_cardiac_repair_xenium
- gse296517_breast_fus_cosmx
- gse296577_gbm_ohsv_xenium
- gse325253_lung_drug_xenium
- merfish_sickness
- dss_colitis_merfish
- gse295862_adipose_weightloss_xenium
- uc_vedolizumab_cosmx
candidate_datasets: []
eligible_baselines: []
success_artifacts:
- result_json
- exact_config
- split_manifest
- cache_manifest
- seed
- code_revision
execution_protocol: null
execution_note: Information contract only; no active execution entrypoint for this
setting.
candidate_baselines:
- count_matched_resampling
- gaussian_affine_transport
- radial_quantile_transport
- scfoundation
- scgpt
- scgpt_spatial
- spatialformer