File size: 11,949 Bytes
07fcdfe | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 | """Diagnostic probe for DrugRank-Flow weak retrieval.
Dumps hard statistics to results/diagnostics/alignment_diag.json so we can
reason about WHY gap_emb <-> drug_proj alignment is only ~2-3x chance:
1. Transcriptional signal strength by dose: ||target_mean - source_mean||
(is 10nM basically vehicle-noise?)
2. gap_emb collapse: pairwise cosine spread + per-dim std over many conditions
3. drug_proj separation: pairwise cosine among the 189 gallery drugs
4. Same-drug cross-dose consistency of gap_emb (tests dose-confounding)
5. Retrieval Hit@10 / median-rank stratified by dose
6. gap_emb vs drug_proj: for a batch, cosine(true pair) vs cosine(best wrong)
Run:
python scripts/diagnose_alignment.py \
--checkpoint outputs/drug_rank/phase2_best.pt \
--config configs/drug_rank_phase2.yaml
"""
from __future__ import annotations
import argparse
import json
import logging
import os
import sys
from pathlib import Path
import numpy as np
import torch
import torch.nn.functional as F
sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "src"))
from gidflow.data.sciplex_dataset import Sciplex3Dataset
from gidflow.models.population_encoder import PopulationEncoder
from gidflow.models.gap_encoder import GapEncoder
from gidflow.models.drug_encoder import DrugEncoder
from gidflow.models.drug_gene_bridge import DrugGeneBridge
logging.basicConfig(level=logging.INFO, format="%(asctime)s | %(message)s", datefmt="%H:%M:%S")
log = logging.getLogger(__name__)
ANNOTATION_DIR = Path("/data/boom/ICLR/data/annotation")
CELL_LINE_MAP = {"A549": 1, "K562": 2, "MCF7": 3}
def build_and_load(cfg, ckpt_path, device):
m = cfg["model"]
num_proteins = len(json.load(open(ANNOTATION_DIR / "protein_target_vocab.json")))
source_enc = PopulationEncoder(num_genes=m["num_genes"], hidden_dim=m["encoder_hidden"], output_dim=m["encoder_output"]).to(device)
target_enc = PopulationEncoder(num_genes=m["num_genes"], hidden_dim=m["encoder_hidden"], output_dim=m["encoder_output"]).to(device)
gap_enc = GapEncoder(input_dim=m["encoder_output"], hidden_dim=m["gap_hidden"], output_dim=m["gap_output"],
proj_dim=m["gap_proj_dim"], num_cell_lines=m["num_cell_lines"], num_genes=m["num_genes"]).to(device)
drug_enc = DrugEncoder(encoding=m.get("drug_encoder", "morgan"), emb_dim=m["drug_emb_dim"], freeze=True).to(device)
bridge = DrugGeneBridge(num_proteins=num_proteins, drug_emb_dim=m["drug_emb_dim"], hidden_dim=m["bridge_hidden_dim"],
proj_dim=m["bridge_proj_dim"], protein_emb_dim=m["bridge_protein_emb_dim"]).to(device)
ck = torch.load(ckpt_path, map_location=device, weights_only=False)
source_enc.load_state_dict(ck["source_enc"]); target_enc.load_state_dict(ck["target_enc"])
gap_enc.load_state_dict(ck["gap_enc"]); drug_enc.load_state_dict(ck["drug_enc"]); bridge.load_state_dict(ck["bridge"])
for mod in (source_enc, target_enc, gap_enc, drug_enc, bridge):
mod.eval()
return source_enc, target_enc, gap_enc, drug_enc, bridge
@torch.no_grad()
def encode_conditions(dataset, conds, models, device, pass_cell_line):
source_enc, target_enc, gap_enc, drug_enc, bridge = models
X = dataset._X
gaps, dnames, doses, cls, deltas = [], [], [], [], []
for cond in conds:
veh = np.asarray(cond["vehicle_cell_idx"]); drg = np.asarray(cond["drug_cell_idx"])
if len(veh) == 0 or len(drg) == 0:
continue
ns = min(64, len(veh)); nt = min(64, len(drg))
rng = np.random.default_rng(0)
s = rng.choice(veh, ns, replace=False); t = rng.choice(drg, nt, replace=False)
src = torch.from_numpy(np.asarray(X[s], np.float32))[None].to(device)
tgt = torch.from_numpy(np.asarray(X[t], np.float32))[None].to(device)
sm = torch.ones(1, ns, dtype=torch.bool, device=device)
tm = torch.ones(1, nt, dtype=torch.bool, device=device)
z_s = source_enc(src, sm); z_t = target_enc(tgt, tm)
if pass_cell_line:
cl = torch.tensor([CELL_LINE_MAP.get(cond["cell_line"], 0)], device=device)
g = gap_enc(z_s, z_t, cell_line_ids=cl)["gap_emb"]
else:
g = gap_enc(z_s, z_t)["gap_emb"]
gaps.append(g[0].cpu().numpy())
dnames.append(cond["drug_name"]); doses.append(float(cond["dose"])); cls.append(cond["cell_line"])
deltas.append(float(np.linalg.norm(np.asarray(X[t]).mean(0) - np.asarray(X[s]).mean(0))))
return np.array(gaps), dnames, np.array(doses), cls, np.array(deltas)
@torch.no_grad()
def build_gallery(drug_order, smiles_map, drug_enc, bridge, device):
projs = []
for start in range(0, len(drug_order), 32):
names = drug_order[start:start+32]
smis = [smiles_map.get(n, "C") or "C" for n in names]
emb = drug_enc(smis); projs.append(bridge(emb)["drug_proj"].cpu().numpy())
return np.concatenate(projs, 0)
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--checkpoint", default="outputs/drug_rank/phase2_best.pt")
ap.add_argument("--config", default="configs/drug_rank_phase2.yaml")
ap.add_argument("--n-conditions", type=int, default=400)
args = ap.parse_args()
import yaml
cfg = yaml.safe_load(open(args.config))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
models = build_and_load(cfg, args.checkpoint, device)
source_enc, target_enc, gap_enc, drug_enc, bridge = models
dc = cfg["data"]; mc = cfg["model"]
smiles_csv = os.path.join(dc["annotation_dir"], "drug_annotation_master.csv")
dataset = Sciplex3Dataset(h5ad_path=dc["sciplex3_h5ad"], n_hvg=mc["num_genes"],
max_source_cells=64, max_target_cells=64, seed=42,
drug_emb_dim=mc["drug_emb_dim"], preprocessed_path=None,
drug_smiles_csv=smiles_csv)
drug_order = json.load(open(ANNOTATION_DIR / "drug_order.json"))
drug_to_idx = {n: i for i, n in enumerate(drug_order)}
import csv as _csv
smiles_map = {}
with open(os.path.join(dc["annotation_dir"], "drug_annotation_master.csv")) as f:
for row in _csv.DictReader(f):
if row.get("drug_name") and row.get("smiles"):
smiles_map[row["drug_name"].strip()] = row["smiles"].strip()
out = {}
# ---- Signal strength by dose (all conditions) ----
# Cache vehicle means per cell line ONCE (vehicle_cell_idx is shared per line).
veh_mean_cache = {}
def _veh_mean(cell_line, veh_idx):
if cell_line not in veh_mean_cache:
veh_mean_cache[cell_line] = dataset._X[np.asarray(veh_idx)].mean(0)
return veh_mean_cache[cell_line]
by_dose = {}
for cond in dataset._conditions:
veh = np.asarray(cond["vehicle_cell_idx"]); drg = np.asarray(cond["drug_cell_idx"])
if len(veh) == 0 or len(drg) == 0:
continue
d = float(np.linalg.norm(dataset._X[drg].mean(0) - _veh_mean(cond["cell_line"], veh)))
by_dose.setdefault(float(cond["dose"]), []).append(d)
out["signal_by_dose"] = {str(k): {"mean_delta_norm": round(float(np.mean(v)), 4),
"std": round(float(np.std(v)), 4), "n": len(v)}
for k, v in sorted(by_dose.items())}
# vehicle-vehicle noise floor: split vehicle cells of one cell line in half
veh_all = dataset._conditions[0]["vehicle_cell_idx"]
rng = np.random.default_rng(1)
noise = []
for _ in range(20):
perm = rng.permutation(np.asarray(veh_all)); h = len(perm)//2
noise.append(float(np.linalg.norm(dataset._X[perm[:h]].mean(0) - dataset._X[perm[h:2*h]].mean(0))))
out["vehicle_noise_floor"] = round(float(np.mean(noise)), 4)
# ---- Encode a sample of conditions (BOTH with and without cell_line) ----
rng2 = np.random.default_rng(7)
sample = list(rng2.choice(len(dataset._conditions), min(args.n_conditions, len(dataset._conditions)), replace=False))
conds = [dataset._conditions[i] for i in sample]
for tag, pass_cl in [("no_cellline_TRAINMODE", False), ("with_cellline_EVALMODE", True)]:
gaps, dnames, doses, cls, deltas = encode_conditions(dataset, conds, models, device, pass_cl)
gaps_n = gaps / (np.linalg.norm(gaps, axis=1, keepdims=True) + 1e-8)
# gap collapse
sim = gaps_n @ gaps_n.T
off = sim[~np.eye(len(sim), dtype=bool)]
# gallery
gallery = build_gallery(drug_order, smiles_map, drug_enc, bridge, device)
gallery_n = gallery / (np.linalg.norm(gallery, axis=1, keepdims=True) + 1e-8)
scores = gaps_n @ gallery_n.T # [N, 189]
true_idx = np.array([drug_to_idx.get(n, -1) for n in dnames])
valid = true_idx >= 0
ranks = []
for i in np.where(valid)[0]:
order = np.argsort(-scores[i])
ranks.append(int(np.where(order == true_idx[i])[0][0]) + 1)
ranks = np.array(ranks)
# by dose retrieval
vdoses = doses[valid]
dose_ret = {}
for dv in sorted(set(vdoses.tolist())):
rr = ranks[vdoses == dv]
dose_ret[str(dv)] = {"hit@10": round(float((rr <= 10).mean()), 4),
"median_rank": float(np.median(rr)), "n": int(len(rr))}
out[tag] = {
"gap_offdiag_cosine_mean": round(float(off.mean()), 4),
"gap_offdiag_cosine_std": round(float(off.std()), 4),
"gap_perdim_std_mean": round(float(gaps.std(0).mean()), 4),
"true_pair_cosine_mean": round(float(np.mean([scores[i, true_idx[i]] for i in np.where(valid)[0]])), 4),
"overall_hit@10": round(float((ranks <= 10).mean()), 4),
"overall_median_rank": float(np.median(ranks)),
"retrieval_by_dose": dose_ret,
}
# ---- drug_proj separation (shared) ----
gallery = build_gallery(drug_order, smiles_map, drug_enc, bridge, device)
gallery_n = gallery / (np.linalg.norm(gallery, axis=1, keepdims=True) + 1e-8)
gsim = gallery_n @ gallery_n.T
goff = gsim[~np.eye(len(gsim), dtype=bool)]
out["drug_proj_separation"] = {
"pairwise_cosine_mean": round(float(goff.mean()), 4),
"pairwise_cosine_std": round(float(goff.std()), 4),
"pairwise_cosine_max": round(float(goff.max()), 4),
"n_drugs": len(gallery),
}
# ---- Same-drug cross-dose gap consistency (train mode encoding) ----
gaps, dnames, doses, cls, deltas = encode_conditions(dataset, conds, models, device, pass_cell_line=False)
gaps_n = gaps / (np.linalg.norm(gaps, axis=1, keepdims=True) + 1e-8)
from collections import defaultdict
drug_groups = defaultdict(list)
for i, n in enumerate(dnames):
drug_groups[n].append(i)
within, across = [], []
for n, idxs in drug_groups.items():
if len(idxs) >= 2:
for a in range(len(idxs)):
for b in range(a+1, len(idxs)):
within.append(float(gaps_n[idxs[a]] @ gaps_n[idxs[b]]))
# random across-drug pairs
rng3 = np.random.default_rng(3)
for _ in range(2000):
i, j = rng3.integers(0, len(gaps_n), 2)
if dnames[i] != dnames[j]:
across.append(float(gaps_n[i] @ gaps_n[j]))
out["same_drug_gap_consistency"] = {
"within_drug_cosine_mean": round(float(np.mean(within)), 4) if within else None,
"across_drug_cosine_mean": round(float(np.mean(across)), 4) if across else None,
"n_within_pairs": len(within),
"note": "within should be >> across if gap_emb is drug-specific; if similar, dose/noise dominates",
}
os.makedirs("results/diagnostics", exist_ok=True)
outpath = "results/diagnostics/alignment_diag.json"
json.dump(out, open(outpath, "w"), indent=2)
log.info("Saved %s", outpath)
print(json.dumps(out, indent=2))
if __name__ == "__main__":
main()
|