ChemiCascade is a self-contained computational cheminformatics, graph neural ADME modeling, empirical 3D protein-ligand docking, PROTAC ternary cooperativity, and de novo molecular optimization platform. Implemented in modular Python with numpy and scipy, it executes a complete structure- and ligand-based drug discovery cascade from SMILES ingestion through multi-objective Pareto lead optimization without native C++/CUDA binary dependencies.
Modern small-molecule and heterobifunctional degrader discovery requires tightly coupled feedback across five quantitative disciplines:
-
Topological Cheminformatics: Daylight SMILES/SMARTS parsing, Morgan-invariant canonicalization, SSSR cycle basis perception, Hückel
$4n+2$ aromaticity, maximum-matching Kekulization, and Bemis-Murcko scaffold extraction. -
Molecular Similarity & SAR Series Mining: Morgan circular
ECFP4/ECFP6& pharmacophoricFCFP4fingerprints with bit-level substructure provenance, topological torsions, vectorized Tanimoto/Dice/Tversky similarity matrices, Taylor-Butina sphere-exclusion clustering, MaxMin diversity picking, and Hussain-Rea Matched Molecular Pair (MMP) property-delta analysis. -
Differentiable Graph Neural ADME & Mechanistic PBPK: An end-to-end differentiable Message-Passing & Graph Attention Network (
MolecularMPNN) in NumPy with analytical backpropagation (verified against central finite differences), AdamW + cosine annealing, MC-Dropout epistemic uncertainty, Integrated Gradients atom attribution, Henderson-Hasselbalch$\log D(\text{pH})$ , well-stirred/parallel-tube hepatic scaling, and a 2-compartment oral PBPK ODE solver. - 3D Distance Geometry, Empirical Docking & PROTAC Biophysics: Triangle-smoothed Crippen-Havel Metric Matrix Distance Geometry embedding, analytical-gradient UFF force-field minimization, Vina-inspired 6-DOF empirical receptor docking with RMSD pose clustering, and exact mass-balance PROTAC ternary complex equilibrium solving with Worm-Like Chain (WLC) linker entropy.
- De Novo Multi-Objective Optimization: Valence-safe chemical graph mutations (substituent grafting, bioisosteric ring heteroatom swaps, leaf pruning) coupled with hybrid UCB1 Monte Carlo Tree Search and NSGA-II Pareto frontier optimization.
flowchart TD
A[SMILES / Compound Series] --> B[smiles.py: Parser, Canonicalizer & SMARTS]
B --> C[topology.py: SSSR, Huckel 4n+2, Kekule, Murcko]
B --> D[descriptors.py: Gasteiger PEOE, Crippen, TPSA, QED, SA]
B --> E[fingerprints.py: ECFP/FCFP, Butina, MaxMin, MMPs]
C --> F[featurizer.py: Node/Edge Tensors & Scaffold Split]
D --> F
F --> G[gnn.py: Differentiable MPNN-GAT + MC-Dropout + IG]
G --> H[adme_pk.py: Henderson-Hasselbalch, CL_H, 2-Comp PBPK ODE]
C --> I[geometry.py: Triangle Smooth DG + UFF BFGS]
D --> J[docking.py: Empirical Vina Scoring, 6-DOF MC/Powell & PROTAC]
I --> J
G --> K[optimizer.py: Valence Mutator + UCB1 MCTS + NSGA-II]
J --> K
K --> L[pipeline.py: DiscoveryCascade + JSON Export]
L --> M[visualizer.py: Strict 7-Bit ASCII Panels & CLI]
flowchart LR
T[types.py] --> S[smiles.py]
S --> TOP[topology.py]
S --> DES[descriptors.py]
S --> FP[fingerprints.py]
TOP --> FEAT[featurizer.py]
DES --> FEAT
FEAT --> GNN[gnn.py]
GNN --> PK[adme_pk.py]
DES --> GEO[geometry.py]
GEO --> DOCK[docking.py]
DES --> DOCK
GNN --> OPT[optimizer.py]
DOCK --> OPT
OPT --> PIPE[pipeline.py]
PIPE --> VIS[visualizer.py]
VIS --> CLI[cli.py]
Each atom
Across damped iteration
Summing implicit-H charges back onto their parent heavy atoms enforces exact formal charge conservation:
-
Wiener Index:
$W = \sum_{i < j} d_{ij}$ where$d_{ij}$ is shortest-path topological distance. -
Zagreb Indices:
$M_1 = \sum_i \deg(i)^2$ ,$M_2 = \sum_{(u,v) \in E} \deg(u)\deg(v)$ . -
Randic Connectivity:
$\chi = \sum_{(u,v) \in E} \left(\deg(u)\deg(v)\right)^{-1/2}$ . -
Balaban Centric Index:
$J = \frac{M}{\mu + 1} \sum_{(u,v) \in E} (s_u s_v)^{-1/2}$ with cyclomatic number$\mu = M - N + C$ and distance sums$s_u = \sum_k d_{uk}$ . - Bickerton QED: Weighted geometric mean over 8 Asymmetric Double Sigmoidal (ADS) desirability curves:
For directed edge
Global attentive + mean pooling feeds a 2-layer SiLU MLP predicting log_s, log_d_74, cl_int_log10, papp_log10, log_bb, and herg_pic50. Atom-level Integrated Gradients attribution is computed via
Microspecies unionized fraction and distribution coefficient at physiological pH
Microsomal intrinsic clearance
Oral pharmacokinetics integrates the 3-state ODE system via SciPy RK45 (A_gut, A_central, A_peripheral):
Triangle-smoothed pairwise bounds
Empirical ligand-receptor binding free energy
Given POI
where effective cooperativity
| Module | Primary Classes / Functions | Mathematical & Domain Invariants |
|---|---|---|
chemi_cascade.types |
Atom, Bond, MoleculeGraph, Conformer3D, ReceptorPocket, DockingPose, ADMEPrediction, PKProfile, ProtacSpec, OptimizationTrace
|
Indexing invariant atoms[i].idx == i; physical constants in A, Da, e, kcal/mol, uM, h |
chemi_cascade.smiles |
parse_smiles, to_canonical_smiles, molecular_formula, substructure_matches, has_substructure
|
Daylight valence H-saturation; Morgan invariant canonicalization; SMARTS backtracking |
chemi_cascade.topology |
connected_ring_systems, perceive_huckel_aromaticity, kekulize, extract_murcko_scaffold
|
Linearly independent SSSR XOR basis; |
chemi_cascade.descriptors |
compute_gasteiger_charges, compute_topological_indices, compute_descriptors, screen_structural_alerts
|
Exact |
chemi_cascade.fingerprints |
compute_morgan_fingerprint, compute_topological_torsion, pairwise_similarity_matrix, butina_clustering, maxmin_diverse_selection, find_matched_molecular_pairs
|
Bit-to-atom provenance; symmetric |
chemi_cascade.featurizer |
featurize_molecule, scaffold_split, GraphTensor
|
32-D node features, 10-D edge features; zero Murcko scaffold overlap between train and test |
chemi_cascade.gnn |
MolecularMPNN, OptionalPyTorchBridge, TrainingHistory
|
Analytical backpropagation verified via central finite differences; AdamW + cosine decay; MC-Dropout; IG attribution |
chemi_cascade.adme_pk |
estimate_dominant_pka, henderson_hasselbalch_logd, scale_hepatic_clearance, simulate_two_compartment_pk, build_calibrated_adme_model
|
|
chemi_cascade.geometry |
build_distance_bounds, triangle_smooth, metric_matrix_embed, embed_conformer, generate_conformer_ensemble, kabsch_rmsd
|
Floyd-Warshall triangle inequality; analytical UFF gradients via BFGS; SO(3)-invariant Kabsch RMSD |
chemi_cascade.docking |
build_kinase_hinge_pocket, score_ligand_pose, dock_ligand, wlc_linker_entropy_penalty, solve_ternary_equilibrium, simulate_protac_hook_effect
|
6-DOF MC + Powell pose search; RMSD clustering; exact ternary mass balance |
chemi_cascade.optimizer |
generate_neighbors, non_dominated_sort, compute_pareto_hypervolume_2d, run_pareto_optimization
|
Valence-safe round-trip mutations; NSGA-II crowding distance; UCB1 MCTS parent selection |
chemi_cascade.benchmark |
run_benchmark_suite, sensitivity_pk_clearance, BenchmarkReport
|
Charge conservation, GNN gradient, scaffold-split RMSE/Spearman, docking ROC-AUC, PROTAC mass balance |
chemi_cascade.pipeline |
DiscoveryCascade, CascadeResult, export_cascade_json
|
End-to-end orchestration and structured JSON serialization |
chemi_cascade.visualizer |
render_descriptor_panel, render_similarity_matrix, render_pk_curve, render_protac_hook_curve, render_docking_breakdown, render_pareto_frontier
|
Strict 7-bit ASCII (codes 0-127) output safe on Windows cp1252 and headless CI |
chemi_cascade.cli |
build_parser, main
|
7 subcommands (analyze, cluster, adme, dock, optimize, benchmark, demo) |
git clone https://github.com/ZengLLQ/chemi-cascade.git
cd chemi-cascade
python -m pip install -e .[dev]Optional PyTorch bridge (not required for any core algorithm or test):
python -m pip install -e .[ml]| Subcommand | Purpose | Example |
|---|---|---|
analyze |
Canonical SMILES, Murcko scaffold, Gasteiger Q, descriptors & alerts | chemi-cascade analyze --smiles "CC(=O)Oc1ccccc1C(=O)O" |
cluster |
ECFP4 Tanimoto heatmap, Butina clusters, MaxMin & MMPs | chemi-cascade cluster --threshold 0.55 |
adme |
Calibrate MPNN-GAT, predict 6 ADME endpoints + 2-comp PBPK curve | chemi-cascade adme --smiles "CN1C=NC2=C1C(=O)N(C(=O)N2C)C" --epochs 20 |
dock |
Embed 3D UFF conformer, dock into kinase hinge & PROTAC hook curve | chemi-cascade dock --smiles "Cc1ccc(NC(=O)c2ccccc2)cc1" --seed 42 |
optimize |
Run MCTS + NSGA-II multi-objective de novo lead optimization | chemi-cascade optimize --seed-smiles "c1ccc(Nc2ncnc3ccccc23)cc1" --generations 4 |
benchmark |
Run quantitative verification gates & clearance sensitivity sweep | chemi-cascade benchmark --seed 42 |
demo |
Execute complete end-to-end discovery cascade + optional JSON export | chemi-cascade demo --seed 42 --out-json report.json |
from chemi_cascade import (
DiscoveryCascade,
compute_descriptors,
dock_ligand,
build_kinase_hinge_pocket,
parse_smiles,
murcko_scaffold_smiles,
)
# 1. Parse, canonicalize, and compute physicochemical profile
mol = parse_smiles("c1ccc(Nc2ncnc3ccccc23)cc1", name="Quinazoline_Seed")
desc = compute_descriptors(mol)
print(f"MW={desc.molecular_weight}, cLogP={desc.clogp}, QED={desc.qed}, Scaffold={murcko_scaffold_smiles(mol)}")
# 2. Dock into the synthetic kinase ATP-hinge pocket
pocket = build_kinase_hinge_pocket()
poses = dock_ligand(mol, pocket, n_starts=6, n_poses=2, seed=42)
print(f"Best pose DeltaG = {poses[0].affinity_kcal:.3f} kcal/mol (Kd = {poses[0].predicted_kd_um:.2f} uM)")
# 3. Run full multi-stage discovery cascade
cascade = DiscoveryCascade(seed=42)
result = cascade.run_series(opt_generations=3, opt_pop_size=10, adme_epochs=25)
print(f"Pareto front size: {len(result.optimization_trace.pareto_front)}")==========================================================================
PROTAC TERNARY COMPLEX HOOK-EFFECT DOSE-RESPONSE
==========================================================================
Peak [ABP]=22.74 nM | EC50=4.23 nM | Hook IC50=1972.7 nM | eff_alpha=8.920 | dG_WLC=0.18 kcal/mol
--------------------------------------------------------------------------
22.74nM| ##### |
20.47nM| ## ## |
18.19nM| # ## |
15.92nM| # # |
13.64nM| ## ## |
11.37nM| # # |
9.10nM| # ## |
6.82nM| ## ## |
4.55nM| ## ### |
2.27nM| ### #### |
0.00nM|###### #####################|
+----------------------------------------------------------+
10^-1.0 nM 10^+5.0 nM
Running python -m chemi_cascade.cli benchmark --seed 42 checks all mathematical invariants and model generalization gates:
| Verification Gate | Criterion | Typical Measured Result |
|---|---|---|
| Gasteiger PEOE Charge Conservation | $\max | \sum q_i - Q_{\text{tot}} |
| MPNN-GAT Analytical vs. Finite-Diff Gradients | ||
| Murcko Scaffold-Split ADME Test RMSE | Bounded generalization error |
|
| Kinase Docking Virtual Enrichment ROC-AUC |
|
|
| PROTAC Ternary Mass Conservation Error | $\max | \Delta C |
| De Novo MCTS + NSGA-II MPO Progression | $\text{MPO}{\text{final}} \ge \text{MPO}{\text{seed}}$ | Positive multi-objective gain |
Run all five self-contained examples directly after installation:
python examples/01_smiles_topology_and_descriptors.py
python examples/02_fingerprints_clustering_and_mmp.py
python examples/03_gnn_adme_and_pbpk_kinetics.py
python examples/04_3d_conformer_docking_and_protac.py
python examples/05_denovo_mcts_pareto_cascade.py-
Pure-Python / NumPy Graph Engine: Keeping
chemi_cascadedependency-light (numpy,scipy) guarantees deterministic installation and execution across Linux, macOS, and Windows without C++ compiler toolchains, while targeting single-series to mid-size virtual screening cohorts ($N \le 10{,}000$ compounds) rather than billion-compound ultra-large screening. - 2D-to-3D Conformer & Docking Approximation: Distance-geometry embedding and UFF minimization operate over heavy-atom graphs with implicit hydrogen steric radii; receptor pockets are treated as rigid pharmacophore grids rather than fully flexible induced-fit protein dynamics.
-
Synthetic ADME Calibration Labels: Built-in ADME calibration uses physically grounded QSAR equations (ESOL-like
$\log S$ , Henderson-Hasselbalch$\log D$ , Clark$\log BB$ ) for reproducible out-of-the-box calibration and testing; production medicinal chemistry deployments should supply experimental assay labels toMolecularMPNN.fit(). - Research Scope Disclaimer: ChemiCascade is a computational chemistry research and simulation toolkit and is not validated or certified for clinical decision-making or regulatory submission.
MIT License. See LICENSE.