Six-Stage Filtering Exposes Weak Drug Generators

Molecular generators are often evaluated on validity, novelty, uniqueness, distributional similarity, or one downstream score. Those measures are useful for detecting broken generators, but they do not match how early drug discovery triages compounds. A molecule that is valid SMILES and has a favorable docking score can still be chemically implausible, hard to synthesize, structurally undesirable, or geometrically inconsistent in the binding pocket.

This paper introduces HEDGEHOG, a six-stage benchmark for asking a stricter question: after a generator proposes molecules, how many survive a medicinal-chemistry workflow that becomes more expensive and target-specific at each step? The authors instantiate the benchmark on KRAS G12D in the switch-II pocket and evaluate 23 generators across unconditional, ligand-based, and protein-based classes.

Core Contribution

The main contribution is not another molecule generator. It is an evaluation cascade that makes failure visible at the stage where it occurs. HEDGEHOG starts with cheap standardization and descriptor checks, then applies structural alerts, synthesis-feasibility tests, docking and affinity estimation, and final 3D pose checks. That ordering matters: retrosynthesis, docking, and affinity prediction are slow enough that running them on every raw output would waste compute on molecules already disqualified by basic chemistry.

Figure 1 shows the workflow: preprocessing, 21 physicochemical descriptors, structural filters, synthesis feasibility, molecular docking, and final 3D filtration.

Technical Approach

The benchmark first cleans generated SMILES with RDKit and DataMol, removing salts and solvents, standardizing molecules, retaining the largest organic fragment, and enforcing a restricted element set. It then computes 21 two-dimensional descriptors, including molecular weight, logP, topological polar surface area, QED, hydrogen-bond counts, ring statistics, and an MCE-18 complexity term. The descriptor thresholds are not a single Lipinski-style rule; the authors combine literature cutoffs with manually chosen bounds where no accepted threshold exists.

The structural stage aggregates public SMARTS-based alert collections and graph-level checks: PAINS, Dundee, BMS, Glaxo, Lilly MedChem rules, NIBR filters, Bredt-rule violations, protecting groups, ring infractions, stereochemistry constraints, and excessive halogenicity. Synthesis feasibility then combines SA score, RA score, SYBA score, and AiZynthFinder route search. The default thresholds include SA ≤ 4.5, RA ≥ 0.5, SYBA ≥ 0, and a successful retrosynthetic route within the configured search.

The final two stages make the benchmark target-aware. Surviving molecules are docked into KRAS G12D with smina, GNINA, and Matcha, and binding affinity is estimated with Boltz-2. The docking threshold is −6.5 kcal/mol, and the Boltz-2 cutoff corresponds to predicted IC50 ≤ 100 µM. The last stage checks whether the pose stays in the intended pocket, passes PoseCheck Fast geometry tests, avoids excessive conformer deviation, and satisfies a KRAS-specific Asp12 interaction requirement.

Results and Analysis

The headline result is severe attrition. Across 230,000 generated molecules, only 1,490 pass all six stages, or 0.65% of the starting pool. Figure 2 makes the drop-off clearer than a single aggregate: all three model classes lose most candidates before the final target-aware filters, and high early retention does not reliably predict final survival.

The class-level counts show different failure modes. Unconditional models start with 80,000 molecules and end with 609 survivors, or 0.76%. Ligand-based models start with 70,000 and end with 396 survivors, or 0.57%. Protein-based models start with 80,000 and end with 485 survivors, or 0.61%. Descriptor filtering is the first major bottleneck: after Stage 2, unconditional, ligand-based, and protein-based classes are reduced to 24.93%, 28.54%, and 24.27% of their initial molecules. After structural filters, those fall to 5.82%, 5.90%, and 3.62%.

The model rankings are also more nuanced than a simple class comparison. Dragonfly is the top protein-based model with 345 final molecules, nearly twice the second-best ligand-based model, REINVENT4 (V), at 182. The best unconditional model, REINVENT4, retains 163. Yet protein-based models as a class do not win end to end, because their early-stage losses are large and their final count is driven disproportionately by Dragonfly. The authors’ interpretation is convincing: target conditioning helps only when it preserves chemically plausible scaffolds and synthesis-compatible structures.

Caveats

The evidence is strongest as a computational benchmark, not as proof that the surviving molecules are hits. The paper does not estimate variance across independent generation seeds for all models, because some baselines lacked consistent seed control. The benchmark instance is centered on one target, KRAS G12D, one switch-II pocket setup, and a target-specific Asp12 interaction rule, so rankings may change for other targets or interaction hypotheses. The pipeline also omits long-term ADMET, pharmacokinetics, clinical-success prediction, uncertainty-aware generation, and prospective validation. Those omissions do not weaken the central claim, but they define the boundary: HEDGEHOG is a stricter triage benchmark, not an experimental discovery result.