jang1563 commited on
Commit
040093c
·
verified ·
1 Parent(s): 505b7dc

Add benchmark data (16/37): tasks/track_c/C3.json

Browse files
Files changed (1) hide show
  1. tasks/track_c/C3.json +88 -0
tasks/track_c/C3.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task_id": "C3",
3
+ "task_name": "PPI-Guided GNN Gene Classification",
4
+ "task_type": "binary_classification",
5
+ "difficulty_tier": "advanced",
6
+ "track": "C",
7
+ "description": "Node classification on STRING PPI graph for spaceflight DEG prediction. Transductive setting: full graph structure visible during training and testing, only labels are split. Tests whether protein-protein interaction network topology improves gene-level predictions beyond flat tabular classifiers.",
8
+ "data_files": [
9
+ "cfrna_3group_de_noleak.csv"
10
+ ],
11
+ "ground_truth_files": [
12
+ "cfrna_466drr.csv"
13
+ ],
14
+ "graph_data": {
15
+ "source": "STRING v12.0 (Homo sapiens, score > 700)",
16
+ "nodes": 26845,
17
+ "connected_nodes": 14900,
18
+ "isolated_nodes": 11945,
19
+ "edges": 213753,
20
+ "node_features": "29 effect-size statistics (same as B1)"
21
+ },
22
+ "input_spec": {
23
+ "node_features": "29 effect-size statistics per gene",
24
+ "feature_count": 29,
25
+ "edge_source": "STRING PPI v12.0 (combined_score > 700)",
26
+ "gene_mapping": "Gene symbol -> ENSP via STRING aliases (67% coverage)",
27
+ "isolated_node_behavior": "GNN degenerates to MLP for isolated nodes (no message passing)"
28
+ },
29
+ "output_spec": {
30
+ "type": "binary",
31
+ "positive_class": "DRR gene",
32
+ "positive_count": 466,
33
+ "negative_count": 26379,
34
+ "positive_rate": "1.7%"
35
+ },
36
+ "evaluation": {
37
+ "primary_metric": "auprc",
38
+ "secondary_metrics": [
39
+ "auroc",
40
+ "f1"
41
+ ],
42
+ "additional_analysis": [
43
+ "isolated_vs_connected_node_performance",
44
+ "architecture_comparison"
45
+ ]
46
+ },
47
+ "models": [
48
+ {
49
+ "name": "GCN",
50
+ "layers": 2,
51
+ "hidden_dim": 64,
52
+ "dropout": 0.3
53
+ },
54
+ {
55
+ "name": "GAT",
56
+ "layers": 2,
57
+ "hidden_dim": 64,
58
+ "heads": 2,
59
+ "dropout": 0.3
60
+ },
61
+ {
62
+ "name": "GraphSAGE",
63
+ "layers": 2,
64
+ "hidden_dim": 64,
65
+ "aggregator": "mean",
66
+ "dropout": 0.3
67
+ }
68
+ ],
69
+ "training": {
70
+ "epochs": 200,
71
+ "optimizer": "Adam",
72
+ "learning_rate": 0.01,
73
+ "loss": "BCEWithLogitsLoss",
74
+ "pos_weight": "balanced (~56.5)",
75
+ "seed": 42
76
+ },
77
+ "split": "feature_split_B1",
78
+ "setting": "transductive",
79
+ "n_samples": 26845,
80
+ "caveats": [
81
+ "Transductive: full graph structure (all edges) visible during both training and testing",
82
+ "~44% of nodes are isolated (no PPI edges) — GNN degenerates to MLP for these",
83
+ "Stage 2 showed PPI topology features provide zero value (net_only AUPRC=0.019, combined < base_only)",
84
+ "Expected negative result: GNN unlikely to outperform flat XGBoost (AUPRC=0.922)",
85
+ "Gene symbol to ENSP mapping coverage 67% (non-protein-coding genes unmapped)"
86
+ ],
87
+ "status": "design_complete"
88
+ }