📄 chemistry

ReliaMol-3D: Auditable Post-Generation Reliability Reranking for Pocket-Conditioned 3D Molecular Generation

ReliaMol-3D is an auditable, generator-agnostic reranking workflow that significantly improves the reliability of pocket-conditioned 3D molecular generation by utilizing score-free evidence from ligand chemistry, geometry, and context to prioritize high-quality candidates without relying on docking scores or direct failure indicators.

Original authors: Shenkuang Wu, Jing Zhang, Lianming Zhang, Pingping Dong

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shenkuang Wu, Jing Zhang, Lianming Zhang, Pingping Dong

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a hiring manager for a very specific job: finding a key that fits perfectly into a complex, unique lock (a protein pocket in the body).

Recently, AI tools have become incredibly good at generating thousands of potential keys (molecules) in seconds. But here is the problem: just because a key looks good on its own doesn't mean it will actually fit into the lock without breaking, jamming, or falling out.

The paper introduces ReliaMol-3D, which isn't a new key-maker. Instead, it's a super-strict, auditable quality inspector that sorts through the pile of AI-generated keys to find the ones that are actually ready for the next step.

Here is how it works, using simple analogies:

1. The Problem: The "Good Looking" Trap

Imagine an AI generates 10,000 keys.

Old Method (QED/SA): This method looks at the key's material. Is it gold? Is it the right size? It's a good check, but it doesn't tell you if the key's teeth are bent or if it will jam the lock.
Docking Scores: This is like a machine that tries to force the key into the lock and gives it a score based on how hard it pushes. Sometimes, the machine says "Great score!" even if the key is actually cracked or the lock is damaged.
Physical Checks (PoseBusters): This is a safety gate. It checks if the key is made of solid metal or if it's melting. It's a pass/fail test, but it doesn't tell you which of the passing keys is the best.

The result? You have a huge pile of keys, and it's very hard to pick the top 10 that won't fail when you actually try to use them.

2. The Solution: ReliaMol-3D (The "Fit-Check" Inspector)

ReliaMol-3D is a new workflow that acts as a re-ranker. It doesn't try to make new keys; it takes the pile the AI made and sorts them based on how well they fit the specific lock.

It uses three main "eyes" to judge the keys, but it deliberately ignores the "force scores" (docking scores) and the "safety gates" (physical checks) during its main sorting process. This keeps the inspection honest and separate from the other tools.

Eye 1: The Chemist's Eye (Ligand Chemistry): Does the key look like a real, usable key? Is it too complex to build? Is it made of weird materials?
Eye 2: The Sculptor's Eye (Internal 3D Geometry): Is the key bent? Are the teeth twisted? Is the metal warped? It checks the shape of the key itself.
Eye 3: The Locksmith's Eye (Ligand-Pocket Context): This is the most important part. It looks at the key inside the specific lock. Does it sit deep enough? Is it touching the right walls? Is it crashing into the lock's frame?

3. The "Exact Frame" Rule

The paper makes a very specific point about coordinates.
Imagine you are trying to fit a key into a lock, but you are looking at a blurry, low-resolution photo of the lock. You might think the key fits, but in reality, it's off by a millimeter.
ReliaMol-3D insists on using the exact, high-definition blueprint of the lock (the "exact-CrossDocked" coordinates). If you use a blurry photo (public approximations), the inspector fails. If you use the exact blueprint, the inspector becomes incredibly accurate.

4. The Results: Sorting the Gold from the Dust

When the researchers tested this on 10,000 keys generated by a popular AI:

The Old Way (QED/SA): If you picked the top 10 keys, about 3 or 4 of them would actually be broken or useless when you tried to use them.
ReliaMol-3D: If you picked the top 10 keys, almost all of them (9.99 out of 10) were perfectly usable.

It didn't just find the "best" keys; it found the keys that were reliable in the specific context of the lock.

5. The "Two-Stage" Safety Net

The paper suggests a smart workflow:

Stage 1: Use ReliaMol-3D to sort the pile and pick the top candidates.
Stage 2: Run those top candidates through the "Safety Gate" (PoseBusters) to make sure they are physically valid.
Stage 3: Run them through the "Force Machine" (Docking) to see how strong the bond is.

By separating these steps, scientists can trust that the list they are working with is solid, without mixing up the different types of checks.

Summary

Think of ReliaMol-3D as a quality control supervisor in a factory.

The AI Generator is the machine that churns out products.
The Docking Score is a sales pitch.
The Physical Check is a safety inspector.
ReliaMol-3D is the manager who looks at the product in its specific environment, checks if it's built right, and picks the top 10% that are actually ready for the customer, ignoring the sales pitch and the safety badge until later.

The paper claims this method is auditable (you can see exactly why it picked a key), generator-agnostic (it works no matter which AI made the keys), and highly reliable, especially when the AI makes mistakes or generates "weird" shapes. It turns a messy pile of 10,000 possibilities into a clean, trustworthy shortlist of 10.

Technical Summary: ReliaMol-3D

Problem Statement

Pocket-conditioned three-dimensional (3D) molecular generators can produce numerous ligand candidates for a specific receptor. However, a critical bottleneck remains in the post-generation triage phase: determining which candidates from a generated pool are sufficiently usable within the intended receptor coordinate frame to justify expensive downstream processes like docking refinement, synthesis planning, or expert inspection.

Existing evaluation signals are insufficient for this specific decision:

Ligand-only metrics (e.g., QED, Synthetic Accessibility) capture chemical desirability but ignore receptor placement.
Docking scores (e.g., Vina, GNINA) provide affinity-like estimates but can coexist with strained geometries, clashes, or coordinate inconsistencies.
Physical-validity checks (e.g., PoseBusters) verify internal plausibility but do not rank surviving candidates or account for receptor-specific context.

The central challenge is to rank a fixed pool of generated 3D candidates by computational receptor-frame usability without conflating selection with affinity prediction or embedding external oracles (like docking scores) as hidden inputs.

Methodology: ReliaMol-3D

ReliaMol-3D is an auditable, generator-agnostic post-generation reranking workflow designed to prioritize candidates based on score-free evidence. The workflow separates the roles of generation, selection, and validation.

1. Coordinate Governance and Auditing

The workflow begins with strict receptor-coordinate provenance auditing. Candidates are evaluated in the exact receptor coordinate frame (specifically, exact-CrossDocked heavy-receptor coordinates) rather than approximate public structures. This prevents coordinate-provenance errors from being misidentified as generator failures.

2. Score-Free Feature Extraction

The primary model uses a 41-feature vector derived from three chemically interpretable families, explicitly excluding docking scores, generator identity, direct failure indicators, and PoseBusters outcomes:

Ligand Chemistry: Molecular weight, logP, TPSA, QED, rotatable bonds, ring counts, heteroatom counts, formal charge, rare-element counts, and atom-type fractions.
Internal 3D Geometry: Pose atom counts, pose radius, and nearest-neighbor distance quantiles (5%–90%).
Ligand–Pocket Context: Ligand-to-pocket minimum-distance quantiles and shell occupancy fractions (0–2Å, 2–4Å, etc.) relative to the receptor pocket.

3. Reliability Endpoint Definition

The target variable is a binary computational receptor-reliability label ( $y_i$ ). A candidate is deemed reliable only if it passes five pre-specified failure indicators simultaneously:

Chemical validity ( $z_{chem}$ )
Internal geometry plausibility ( $z_{geom}$ )
Pocket contact presence ( $z_{pocket}$ )
Absence of receptor clash ( $z_{clash}$ )
Synthetic complexity limits ( $z_{syn}$ )

This label is distinct from binding affinity or biological activity; it serves as an auditable endpoint for "usability in the receptor frame."

4. Reranking Models

Primary Model (RF-ReliaMol): A Random Forest classifier trained to predict the probability of passing the reliability criteria. It is transparent, reproducible, and designed for post-generation triage.
Variants: HGB-ReliaMol (Histogram Gradient Boosting) for performance confirmation; MLP-ReliaMol and RCEF (Reliability-Calibrated Evidence Fusion) as diagnostic models to test feature-group sensitivity and calibration.
External Tools: Docking scores (Vina/GNINA) and PoseBusters (PB) are treated as separate comparison baselines or optional strict eligibility gates, not as inputs to the primary reranker.

5. Selection Protocol

Candidates are ranked within each pocket based on the predicted reliability probability ( $\hat{p}_i$ ). A fixed-budget shortlist (e.g., top 10%) is selected. An optional strict PB-gated policy can be applied, where only PB-passing candidates are reranked, without backfilling if the pool is insufficient.

Key Results

Main Benchmark (DiffSBDD)

Evaluated on 10,000 candidates from 100 exact-CrossDocked pockets:

Reliability: At 10% coverage, RF-ReliaMol achieves 0.9990 top-10% reliability (zero standard deviation across seeds), compared to 0.9660 for the QED/SA baseline. HGB-ReliaMol reaches 0.9993.
Ablation: The gain is driven primarily by internal 3D geometry and ligand-pocket context. Ligand descriptors alone improve reliability to 0.9727, but adding pocket geometry raises it to 0.9990.
Robustness: The model maintains high reliability (0.9959–0.9988) across 20% and 50% coverage levels and remains stable under coordinate perturbations (0.25 Å) and noisy evidence.

Stress Tests and Transfer

Generator Shift: On a held-out CVAE generator split (known for high failure rates), RF-ReliaMol improves reliability from 0.7868 (Vina-minimize) to 0.9800.
Leave-Generator-Out: In cross-generator evaluations, RF-ReliaMol achieves 0.9950 reliability, demonstrating transferability across generator families (TargetDiff, Pocket2Mol, AR).
Controlled Failure (DUD-E): On a DUD-E stress test with injected failures, MLP-ReliaMol achieves an ROC-AUC of 0.9981, significantly outperforming Vina (0.5966).

Chemical Utility and Diversity

ReliaMol improves reliability without collapsing chemical diversity:

Scaffold Diversity: RF-ReliaMol retains 9.88 unique Murcko scaffolds per pocket (vs. 9.90 for QED/SA).
Medicinal Chemistry: Selected candidates maintain high Lipinski/Veber pass rates and low PAINS alert rates.
PB Gating: When combined with a strict PoseBusters gate, the selected set achieves 1.0000 reliability and 1.0000 PB pass rate.

Significance and Claims

The paper positions ReliaMol-3D not as a new molecular generator or affinity predictor, but as a reproducibility-focused, role-separated triage layer for structure-based drug design (SBDD).

Auditable Triage: It defines a specific "receptor-frame reliability" endpoint distinct from affinity, allowing for a clear separation between candidate selection and downstream validation.
Score-Free Evidence: By excluding docking scores and generator identity from the primary model, it prevents circular reasoning and ensures the selection is based on interpretable chemical and geometric evidence.
Generator Agnosticism: The workflow is effective across diverse generator families, including those with higher failure rates (e.g., CVAE), making it a robust filter for heterogeneous generated pools.
Practical Workflow: It provides a concrete protocol for prioritizing candidates before expensive rescoring or synthesis, bridging the gap between high-throughput generation and expert inspection.

The authors conclude that while modern generators produce plausible candidates, the downstream bottleneck of selecting the usable subset is best addressed by a dedicated, auditable reranking layer that explicitly models receptor-frame usability through score-free, chemically interpretable features.