liver-extratrees-deployment / deployment_common.py
Delower's picture
Upload deployment_common.py with huggingface_hub
ee82dfc verified
Raw History Blame Contribute Delete
16.8 kB
# deployment_common.py — shared code for the liver deployment refits.
# The tox_common block below is copied VERBATIM from the arms' own notebooks
# (kaggle/notebooks/liver_arch3_lazypredict.py, shared programme conventions
# cell), and the graph block (ELEMS / mol_to_graph / GATNet) is copied
# VERBATIM from kaggle/notebooks/liver_arch2_gat.py. Deployment scripts reuse
# the arms' exact identity, featurization and model code — no re-implementation.
import os, json, time
import numpy as np
import pandas as pd
# ---- shared programme conventions (tox_common, inlined verbatim) ----
"""Shared conventions for the organ-toxicity programme (liver + kidney).
Binding rules (PROJECT_SPEC.md + READING_NOTES harmonization protocols):
- classification ONLY; identity = InChIKey connectivity block (first 14 chars)
of the RDKit-standardized parent structure (largest fragment).
- seed-42 scaffold-disjoint (Bemis-Murcko) train/val/test split.
- preprocessing (imputation / scaling) fit on the TRAIN fold only.
- regulatory benchmark compounds are NEVER in training; lockout verified = 0
with the dual key (deposited connectivity UNION standardized-parent
connectivity) against EACH benchmark file separately.
- benchmark scoring is reported PER SET (withdrawn / clinical failures /
pass), never pooled; metrics at full coverage.
"""
import json
import numpy as np
import pandas as pd
SEED = 42
# ---------------------------------------------------------------- identity
def _inchi_key(mol):
"""InChIKey via whichever RDKit API the installed build provides.
DEPLOYMENT COMPATIBILITY FIX (2026-11-15): the arms' verbatim code calls
rdMolDescriptors.CalcInchiKey, which does not exist in RDKit 2024.03.6 or
2026.09.1 (it silently made parent_connectivity return None there).
rdkit.Chem.inchi.MolToInchiKey is the same InChI computation; verified
byte-identical keys across both installed RDKit builds on a sample set
(salts, stereo, charged species)."""
try:
from rdkit.Chem import inchi
return inchi.MolToInchiKey(mol)
except Exception:
from rdkit.Chem import rdMolDescriptors
return rdMolDescriptors.CalcInchiKey(mol)
def parent_connectivity(smiles):
"""Connectivity block of the standardized parent (largest fragment)."""
try:
from rdkit import Chem
from rdkit.Chem.MolStandardize import rdMolStandardize
if smiles is None or (isinstance(smiles, float) and np.isnan(smiles)):
return None
mol = Chem.MolFromSmiles(str(smiles))
if mol is None:
return None
parent = rdMolStandardize.LargestFragmentChooser().choose(mol)
ik = _inchi_key(parent)
return ik[:14] if ik else None
except Exception:
return None
def deposited_connectivity(inchikey):
if inchikey is None or (isinstance(inchikey, float) and np.isnan(inchikey)):
return None
s = str(inchikey).strip()
return s[:14] if len(s) >= 14 else None
def dual_keys(df, smiles_col="canonical_smiles", inchikey_col="inchikey"):
"""Set of connectivity keys: deposited UNION standardized-parent."""
keys = set()
dep = df[inchikey_col].map(deposited_connectivity) if inchikey_col in df else pd.Series([None] * len(df))
par = df[smiles_col].map(parent_connectivity)
for a, b in zip(dep, par):
if a:
keys.add(a)
if b:
keys.add(b)
return keys
def verify_lockout(train_df, bench_dfs, train_smiles="canonical_smiles"):
"""bench_dfs: dict name -> df (with 'smiles' + 'inchikey' columns).
Returns dict of overlap counts; every count must be 0."""
train_keys = dual_keys(train_df, smiles_col=train_smiles)
out = {}
for name, bdf in bench_dfs.items():
bk = dual_keys(bdf, smiles_col="smiles")
out[name] = len(train_keys & bk)
return out
# ---------------------------------------------------------------- loading
def load_train(path):
df = pd.read_csv(path)
df = df.rename(columns={"canonical_smiles": "smiles"})
df["label"] = df["final_label"].astype(int)
df["name"] = df["drug_name"] if "drug_name" in df else ""
df["conn"] = df["inchikey_connectivity"]
keep = ["name", "smiles", "inchikey", "conn", "label"]
for c in ["sources", "sources_present", "final_label_source"]:
if c in df.columns:
keep.append(c)
return df[keep].reset_index(drop=True)
def load_benchmark(path, expected_label):
df = pd.read_csv(path)
name_col = "drug_name" if "drug_name" in df.columns else "compound_name"
out = pd.DataFrame({
"name": df[name_col],
"smiles": df["smiles"] if "smiles" in df else None,
"inchikey": df["inchikey"] if "inchikey" in df else None,
})
if "evidence_grade" in df.columns:
out["evidence_grade"] = df["evidence_grade"]
out["label"] = expected_label
return out
def dedup_benchmark(df):
"""Dedup benchmark rows to unique compounds (deposited connectivity,
falling back to standardized-parent connectivity). Rows with no usable
structure are counted as unscored."""
keys, keep_idx, unscored = {}, [], 0
for i, row in df.iterrows():
k = deposited_connectivity(row.get("inchikey")) or parent_connectivity(row.get("smiles"))
if k is None:
unscored += 1
continue
if k in keys:
continue
keys[k] = True
keep_idx.append(i)
out = df.loc[keep_idx].copy()
out["conn"] = [deposited_connectivity(r.get("inchikey")) or parent_connectivity(r.get("smiles"))
for _, r in out.iterrows()]
return out.reset_index(drop=True), unscored
# ---------------------------------------------------------------- split
def scaffold_split(df, seed=SEED, fracs=(0.80, 0.10, 0.10)):
"""Seed-42 Bemis-Murcko scaffold-disjoint train/val/test split."""
from rdkit import Chem
from rdkit.Chem.Scaffolds import MurckoScaffold
rng = np.random.RandomState(seed)
scaffolds = []
for smi in df["smiles"]:
mol = Chem.MolFromSmiles(str(smi))
try:
scaf = MurckoScaffold.MurckoScaffoldSmiles(mol=mol) if mol is not None else ""
except Exception:
scaf = ""
scaffolds.append(scaf if scaf else f"ACYCLIC::{smi}")
groups = {}
for i, s in enumerate(scaffolds):
groups.setdefault(s, []).append(i)
group_list = list(groups.values())
rng.shuffle(group_list)
group_list.sort(key=len, reverse=True)
n = len(df)
targets = [f * n for f in fracs]
buckets = [[], [], []]
counts = [0, 0, 0]
for g in group_list:
# assign each scaffold group to the bucket furthest below its target
j = int(np.argmax([targets[k] - counts[k] for k in range(3)]))
buckets[j].extend(g)
counts[j] += len(g)
names = ["train", "val", "test"]
split = pd.Series(index=df.index, dtype=object)
for j, idxs in enumerate(buckets):
split.iloc[idxs] = names[j]
# scaffold disjointness assertion
seen = {}
for j, idxs in enumerate(buckets):
for i in idxs:
s = scaffolds[i]
assert seen.get(s, j) == j, "scaffold leaked across splits"
seen[s] = j
return split
# ---------------------------------------------------------------- features
PHYSCHEM_NAMES = ["MolWt", "MolLogP", "TPSA", "NumHDonors", "NumHAcceptors",
"NumRotatableBonds", "RingCount", "NumAromaticRings",
"NumAliphaticRings", "FractionCSP3", "HeavyAtomCount",
"NumValenceElectrons", "BertzCT"]
def featurize(smiles_list):
"""Returns dict of float32 blocks + validity mask. Blocks:
morgan_bits (1024), morgan_counts (1024), maccs (167),
desc (full RDKit descriptor list), physchem (PHYSCHEM_NAMES subset)."""
from rdkit import Chem
from rdkit.Chem import AllChem, MACCSkeys, Descriptors
from rdkit import RDLogger
RDLogger.DisableLog("rdApp.*")
desc_names = [d[0] for d in Descriptors._descList]
phys_idx = [desc_names.index(n) for n in PHYSCHEM_NAMES]
n = len(smiles_list)
blocks = {
"morgan_bits": np.zeros((n, 1024), dtype=np.float32),
"morgan_counts": np.zeros((n, 1024), dtype=np.float32),
"maccs": np.zeros((n, 167), dtype=np.float32),
"desc": np.full((n, len(desc_names)), np.nan, dtype=np.float32),
}
valid = np.zeros(n, dtype=bool)
from rdkit import DataStructs
gen_bits = AllChem.GetMorganGenerator(radius=2, fpSize=1024)
gen_counts = AllChem.GetMorganGenerator(radius=2, fpSize=1024, countSimulation=True)
for i, smi in enumerate(smiles_list):
mol = Chem.MolFromSmiles(str(smi)) if smi is not None and str(smi) != "nan" else None
if mol is None:
continue
valid[i] = True
DataStructs.ConvertToNumpyArray(gen_bits.GetFingerprint(mol), blocks["morgan_bits"][i])
DataStructs.ConvertToNumpyArray(gen_counts.GetCountFingerprint(mol), blocks["morgan_counts"][i])
blocks["maccs"][i] = np.asarray(MACCSkeys.GenMACCSKeys(mol), dtype=np.float32)
# iterate _descList directly: identical values to CalcMolDescriptors
# at ~1/3 the cost (measured on this VM, RDKit 2026.09)
blocks["desc"][i] = np.array([fn(mol) for _, fn in Descriptors._descList], dtype=np.float32)
blocks["physchem"] = blocks["desc"][:, phys_idx].copy()
blocks["desc_names"] = desc_names
return blocks, valid
class DescPrep:
"""Train-fold-fitted descriptor preprocessing: drop columns >50% missing
or near-constant ON TRAIN, median-impute, clip to train 0.1/99.9 pct,
standardize with train mean/std."""
def fit(self, X):
X = np.asarray(X, dtype=np.float64)
miss = np.mean(~np.isfinite(X), axis=0)
std = np.nanstd(np.where(np.isfinite(X), X, np.nan), axis=0)
self.keep_ = (miss <= 0.5) & (std > 1e-8)
Xk = X[:, self.keep_]
self.med_ = np.nanmedian(np.where(np.isfinite(Xk), Xk, np.nan), axis=0)
Xi = np.where(np.isfinite(Xk), Xk, self.med_)
self.lo_ = np.percentile(Xi, 0.1, axis=0)
self.hi_ = np.percentile(Xi, 99.9, axis=0)
Xc = np.clip(Xi, self.lo_, self.hi_)
self.mu_ = Xc.mean(axis=0)
self.sd_ = Xc.std(axis=0) + 1e-9
return self
def transform(self, X):
X = np.asarray(X, dtype=np.float64)[:, self.keep_]
Xi = np.where(np.isfinite(X), X, self.med_)
return ((np.clip(Xi, self.lo_, self.hi_) - self.mu_) / self.sd_).astype(np.float32)
def prep_matrix(blocks, prep, kind):
"""kind: 'fp' (morgan bits) / 'fp_desc' (bits+maccs+desc) /
'desc_maccs_counts' (arch-5 block b) / 'physchem' raw physchem."""
if kind == "fp":
return blocks["morgan_bits"]
if kind == "fp_desc":
return np.hstack([blocks["morgan_bits"], blocks["maccs"], prep.transform(blocks["desc"])]).astype(np.float32)
if kind == "desc_maccs_counts":
return np.hstack([prep.transform(blocks["desc"]), blocks["maccs"], blocks["morgan_counts"]]).astype(np.float32)
if kind == "physchem":
return blocks["physchem"]
raise ValueError(kind)
# ---------------------------------------------------------------- metrics
def metrics_dict(y_true, prob, threshold=0.5):
from sklearn.metrics import (accuracy_score, balanced_accuracy_score,
matthews_corrcoef, roc_auc_score, confusion_matrix)
y_true = np.asarray(y_true, dtype=int)
prob = np.asarray(prob, dtype=float)
pred = (prob >= threshold).astype(int)
out = {"n": int(len(y_true)), "threshold": threshold,
"accuracy": float(accuracy_score(y_true, pred)),
"balanced_accuracy": float(balanced_accuracy_score(y_true, pred))}
tn, fp, fn, tp = confusion_matrix(y_true, pred, labels=[0, 1]).ravel()
out.update({"tn": int(tn), "fp": int(fp), "fn": int(fn), "tp": int(tp),
"sensitivity_recall_pos": float(tp / (tp + fn)) if (tp + fn) else None,
"specificity_recall_neg": float(tn / (tn + fp)) if (tn + fp) else None})
if len(np.unique(y_true)) == 2 and len(np.unique(pred)) == 2:
out["mcc"] = float(matthews_corrcoef(y_true, pred))
else:
out["mcc"] = None
out["mcc_note"] = "single-class ground truth or single-class predictions"
if len(np.unique(y_true)) == 2:
out["auroc"] = float(roc_auc_score(y_true, prob))
else:
out["auroc"] = None
return out
def score_benchmark_sets(predict_proba_fn, bench_sets):
"""bench_sets: dict name -> deduped df (with 'smiles','label','name').
predict_proba_fn(smiles_list) -> P(label=1) array (NaN where unparseable).
Returns (metrics dict per set, predictions dataframe)."""
all_metrics, pred_frames = {}, []
for set_name, bdf in bench_sets.items():
prob = predict_proba_fn(list(bdf["smiles"]))
scored = np.isfinite(prob)
m = metrics_dict(bdf.loc[scored, "label"], prob[scored]) if scored.any() else {"n": 0}
m["n_rows_unique"] = int(len(bdf))
m["n_scored"] = int(scored.sum())
m["n_unscored_invalid_structure"] = int((~scored).sum())
all_metrics[set_name] = m
pf = bdf[["name", "smiles", "label"]].copy()
if "evidence_grade" in bdf.columns:
pf["evidence_grade"] = bdf["evidence_grade"]
pf["set"] = set_name
pf["prob_toxic"] = prob
pf["pred_label"] = np.where(scored, (prob >= 0.5).astype(float), np.nan)
pf["correct"] = np.where(scored, (pf["pred_label"] == pf["label"]).astype(float), np.nan)
pred_frames.append(pf)
return all_metrics, pd.concat(pred_frames, ignore_index=True)
# ---- GAT graph code (verbatim from liver_arch2_gat.py) ----
# Import guard (deployment plumbing only, model code below unchanged): the
# sklearn-side deployment scripts import this module in an env without
# torch_geometric; the graph functions are only called in the pyg env.
try:
import torch
import torch.nn.functional as F
from torch_geometric.data import Data
from torch_geometric.nn import GATConv, global_mean_pool
_TORCH_OK = True
except ImportError:
torch = None
_TORCH_OK = False
ELEMS = [6, 7, 8, 16, 9, 17, 35, 53, 15, 14]
def mol_to_graph(smi):
from rdkit import Chem
HYB = [Chem.rdchem.HybridizationType.SP, Chem.rdchem.HybridizationType.SP2,
Chem.rdchem.HybridizationType.SP3, Chem.rdchem.HybridizationType.SP3D,
Chem.rdchem.HybridizationType.SP3D2]
BOND = {Chem.rdchem.BondType.SINGLE: 0, Chem.rdchem.BondType.DOUBLE: 1,
Chem.rdchem.BondType.TRIPLE: 2, Chem.rdchem.BondType.AROMATIC: 3}
def oh(v, n):
o = [0.0] * n; o[min(max(v, 0), n - 1)] = 1.0; return o
mol = Chem.MolFromSmiles(str(smi)) if smi is not None and str(smi) != "nan" else None
if mol is None or mol.GetNumAtoms() == 0:
return None
xs = []
for a in mol.GetAtoms():
z = a.GetAtomicNum()
f = oh(ELEMS.index(z) if z in ELEMS else len(ELEMS), len(ELEMS) + 1)
f += oh(a.GetDegree(), 6)
f += oh(a.GetFormalCharge() + 2, 5)
f += oh(a.GetTotalNumHs(), 5)
f += oh(HYB.index(a.GetHybridization()) if a.GetHybridization() in HYB else 5, 6)
f += [1.0 if a.GetIsAromatic() else 0.0, 1.0 if a.IsInRing() else 0.0]
xs.append(f)
ei, ea = [], []
for b in mol.GetBonds():
i, j = b.GetBeginAtomIdx(), b.GetEndAtomIdx()
t = BOND.get(b.GetBondType(), 0)
ei += [[i, j], [j, i]]; ea += [oh(t, 4), oh(t, 4)]
return Data(x=torch.tensor(xs, dtype=torch.float),
edge_index=torch.tensor(ei, dtype=torch.long).t().contiguous() if ei else torch.empty((2, 0), dtype=torch.long),
edge_attr=torch.tensor(ea, dtype=torch.float) if ea else torch.empty((0, 4)))
if _TORCH_OK:
class GATNet(torch.nn.Module):
def __init__(self, in_dim, hidden, dropout):
super().__init__()
self.c1 = GATConv(in_dim, hidden, heads=4, dropout=dropout, edge_dim=4)
self.c2 = GATConv(hidden * 4, hidden, heads=1, dropout=dropout, edge_dim=4)
self.head = torch.nn.Sequential(torch.nn.Linear(hidden, hidden // 2), torch.nn.ReLU(),
torch.nn.Dropout(dropout), torch.nn.Linear(hidden // 2, 1))
self.dropout = dropout
def forward(self, data):
x = F.elu(self.c1(data.x, data.edge_index, data.edge_attr))
x = F.dropout(x, p=self.dropout, training=self.training)
x = F.elu(self.c2(x, data.edge_index, data.edge_attr))
x = global_mean_pool(x, data.batch)
return self.head(x).ravel()