"""Synthetic LD-object dataset generator for demo runs. Builds a 96-well plate (8 drugs x 6 doses x 2 replicates) with ~25 cells per well and ~10 LDs per cell. Drug profiles are encoded as fingerprint targets that vary monotonically with dose. """ from __future__ import annotations import os import numpy as np import pandas as pd # --------------------------------------------------------------------------- # Reference drug fingerprints (target distributions at high dose) # --------------------------------------------------------------------------- DRUG_PROFILES = { # name: (macro_pct, medium_pct, micro_pct, ld_count_factor, plin1_factor, plin5_factor, manders_m1) "vehicle": (0.60, 0.30, 0.10, 1.00, 1.00, 0.30, 0.45), "resmetirom": (0.25, 0.35, 0.40, 0.65, 0.80, 1.40, 0.55), "DGAT2_inhibitor": (0.15, 0.30, 0.55, 0.40, 0.60, 0.50, 0.50), "SCD1_inhibitor": (0.50, 0.30, 0.20, 0.85, 1.10, 0.60, 0.45), "FGF21_analog": (0.40, 0.35, 0.25, 0.75, 0.65, 1.50, 0.55), "ACC_inhibitor": (0.30, 0.35, 0.35, 0.55, 0.70, 0.90, 0.50), "oleic_palmitic_loading_only": (0.70, 0.25, 0.05, 1.20, 1.10, 0.30, 0.40), "perilipin5_inducer": (0.45, 0.30, 0.25, 0.85, 0.85, 1.80, 0.60), } VEHICLE_PROFILE = DRUG_PROFILES["vehicle"] DOSES_UM = [0.01, 0.1, 1.0, 10.0, 100.0, 1000.0] DRUGS = list(DRUG_PROFILES.keys()) REPLICATES = [1, 2] def _wells(n: int) -> list[str]: rows = "ABCDEFGH" cols = list(range(1, 13)) out = [] for r in rows: for c in cols: out.append(f"{r}{c:02d}") return out[:n] def _dose_response_factor(dose_uM: float, ec50: float = 5.0, hill: float = 1.0) -> float: """0 at very low dose, 1 at very high dose.""" if dose_uM <= 0: return 0.0 return 1.0 / (1.0 + (dose_uM / ec50) ** -hill) def _interp(vehicle_val: float, target_val: float, frac: float) -> float: return vehicle_val + (target_val - vehicle_val) * frac def build_synthetic(seed: int = 7, n_cells_per_well: int = 25, n_lds_per_cell: int = 10): rng = np.random.default_rng(seed) n_wells_total = len(DRUGS) * len(DOSES_UM) * len(REPLICATES) # 8*6*2 = 96 well_names = _wells(n_wells_total) # Build plate map plate_rows = [] cell_meta_rows = [] ld_rows = [] # Cell-type assignment cycle cell_types = [ "HepG2", "Huh7", "primary_mouse", "iPSC", "HepaRG", "spheroid", "organoid", "STAM_cryosection", ] idx = 0 for drug in DRUGS: target = DRUG_PROFILES[drug] for dose in DOSES_UM: for rep in REPLICATES: well = well_names[idx] idx += 1 plate_rows.append( {"well": well, "drug": drug, "dose_uM": dose, "replicate": rep} ) ct = cell_types[hash(drug) % len(cell_types)] cell_meta_rows.append( {"well": well, "cell_count": n_cells_per_well, "cell_type": ct} ) # Dose-dependent fraction towards target frac = _dose_response_factor(dose, ec50=5.0, hill=1.2) if drug == "vehicle": frac = 0.0 # vehicle stays at vehicle profile regardless of dose # Mix vehicle <-> target macro_pct = _interp(VEHICLE_PROFILE[0], target[0], frac) medium_pct = _interp(VEHICLE_PROFILE[1], target[1], frac) micro_pct = _interp(VEHICLE_PROFILE[2], target[2], frac) # Normalise to 1 tot = macro_pct + medium_pct + micro_pct macro_pct, medium_pct, micro_pct = ( macro_pct / tot, medium_pct / tot, micro_pct / tot, ) ld_count_factor = _interp(1.0, target[3], frac) plin1_factor = _interp(1.0, target[4], frac) plin5_factor = _interp(1.0, target[5], frac) manders_target = _interp(VEHICLE_PROFILE[6], target[6], frac) # Build LD objects for c in range(n_cells_per_well): cell_id = f"{well}_c{c:03d}" n_lds = max(1, int(rng.normal(n_lds_per_cell * ld_count_factor, 1.5))) # Choose LD subtype per droplet subtypes = rng.choice( ["macro", "medium", "micro"], size=n_lds, p=[macro_pct, medium_pct, micro_pct], ) for k, s in enumerate(subtypes): if s == "macro": diam = rng.normal(7.0, 1.5) elif s == "medium": diam = rng.normal(2.5, 0.8) else: diam = rng.normal(0.6, 0.2) diam = max(0.1, float(diam)) area = float(np.pi * (diam / 2.0) ** 2) circ = float(np.clip(rng.normal(0.92, 0.05), 0.5, 1.0)) max_int = float(max(0.0, rng.normal(180 + 30 * (1 - frac), 25))) int_int = float(max_int * area * rng.normal(1.0, 0.1)) # LD-mito distance: closer when manders_target is higher dist = float(max(0.05, rng.normal(2.5 - manders_target * 1.5, 0.6))) m1 = float(np.clip(rng.normal(manders_target, 0.07), 0.0, 1.0)) m2 = float(np.clip(rng.normal(manders_target * 0.95, 0.07), 0.0, 1.0)) pearson = float(np.clip(rng.normal(manders_target - 0.05, 0.08), -0.2, 1.0)) # PLIN ring intensities plin1 = float(max(0.0, rng.normal(120 * plin1_factor, 18))) plin2 = float(max(0.0, rng.normal(80, 14))) plin3 = float(max(0.0, rng.normal(60, 12))) plin5 = float(max(0.0, rng.normal(40 * plin5_factor, 12))) ld_rows.append( { "well": well, "cell_id": cell_id, "ld_id": f"{cell_id}_l{k:03d}", "area_um2": area, "diameter_um": diam, "circularity": circ, "max_intensity": max_int, "integrated_intensity": int_int, "distance_to_mito_um": dist, "manders_m1": m1, "manders_m2": m2, "pearson": pearson, "plin1_intensity": plin1, "plin2_intensity": plin2, "plin3_intensity": plin3, "plin5_intensity": plin5, } ) plate_df = pd.DataFrame(plate_rows) cell_df = pd.DataFrame(cell_meta_rows) ld_df = pd.DataFrame(ld_rows) # Reference MOA fingerprint table (one row per reference drug) ref_rows = [] for drug, prof in DRUG_PROFILES.items(): macro, med, micro, ld_factor, p1, p5, m1 = prof ref_rows.append( { "drug": drug, "macro_pct": macro, "medium_pct": med, "micro_pct": micro, "total_LD_area_per_cell": 80.0 * ld_factor, "ld_count_per_cell": 10.0 * ld_factor, "manders_m1": m1, "plin1_mean": 120.0 * p1, "plin5_mean": 40.0 * p5, } ) ref_df = pd.DataFrame(ref_rows) return ld_df, plate_df, cell_df, ref_df def write_synthetic_to(data_dir: str, seed: int = 7) -> dict: os.makedirs(data_dir, exist_ok=True) ld, plate, cell, ref = build_synthetic(seed=seed) ld_path = os.path.join(data_dir, "ld_objects.csv") plate_path = os.path.join(data_dir, "plate_map.csv") cell_path = os.path.join(data_dir, "cell_meta.csv") ref_path = os.path.join(data_dir, "reference_moa_fingerprints.csv") ld.to_csv(ld_path, index=False) plate.to_csv(plate_path, index=False) cell.to_csv(cell_path, index=False) ref.to_csv(ref_path, index=False) return { "ld_objects": ld_path, "plate_map": plate_path, "cell_meta": cell_path, "reference_moa": ref_path, }