跳转至

中文文档

Case: Cryptic Pocket Discovery Workflow

1. Case Concept and Scientific Questions

1.1 Background and Challenges

In drug discovery, traditional Structure-Based Drug Design (SBDD) workflows typically rely on a static protein structure (usually apo state or a complex with a known ligand). However, many important drug targets such as KRAS G12C, SHP2, and BTK possess a special binding site — the Cryptic Pocket.

Cryptic pockets are characterized by: - Being absent or fully exposed in static protein structures (especially apo state). - Pocket formation resulting from ligand-induced fit or rare conformational states in protein dynamics. - Traditional docking algorithms assuming a rigid protein, therefore systematically missing these important druggable sites.

1.2 Scientific Questions

This case aims to address the core scientific question:

How to build a computational, verifiable closed-loop workflow to predict, validate, and exploit cryptic pockets in proteins?

Specifically, we explore: 1. Conformational Ensemble Generation: Can the OpenFold3 AI model (an open-source reproduction of AlphaFold3) generate an ensemble of "openable" protein conformations through template perturbation and ligand induction strategies, serving as candidate sources for cryptic pockets? 2. Physical Validation: Can molecular dynamics (MD) simulations, particularly enhanced sampling techniques, physically validate whether these candidate conformations can truly open pockets? 3. Ligand-Induced Closed Loop: Once a pocket is validated in physical simulations, can ligand generation and complex prediction be performed on this "open" conformation, with binding advantages quantified by methods such as FEP?

1.3 Research Value

  • Scientific Value: Addresses the core question in AI for Science of whether AI-predicted conformations have physical authenticity.
  • Engineering Value: Provides an automated toolchain that transforms cryptic pocket discovery from "reliance on expert intuition and manual experimentation" to "repeatable, quantifiable computational workflow".
  • Application Value: Provides new strategies for first-in-class drug design targeting KRAS, MYC, and other targets.

2. Workflow Design and Task Decomposition

To realize the above concept, we designed a closed-loop workflow consisting of four core stages. The current WA-DD platform supports this workflow through the existing OpenFold3, GROMACS, PocketXMol, FEP, and standard asset-management stack. Pocket analysis is exposed as part of the GROMACS pocket_discovery protocol and can emit standard reusable pocket assets.

2.1 Workflow Data Flow

graph TD
    A["Input: APO protein structure"] --> B{"Stage 1: OpenFold3 Conformational Ensemble Generation"};
    B -->|"Conformational Ensemble (PDBs)"| C{"Stage 2: MD Physical Validation (Gromacs)"};
    C -->|"MD Trajectory + Representative OPEN conformation"| D{"Stage 3: Pocket and Water Network Analysis"};
    D -->|"Validated OPEN conformation (PDB)"| E{"Stage 4: Ligand Generation and Complex Prediction (Reuse existing)"};
    E -->|"Active ligands + Complex structures"| F(("Scientific Conclusion: Cryptic Pocket Druggability Validation"));

2.2 Task Decomposition and Component Status

Stage Task Description WA-DD Component Status
1. OpenFold3 Conformational Ensemble Generation Generate multiple candidate protein conformations using the OpenFold3 model through template perturbation and ligand induction strategies. wa-dd-openfold3 ✅ Integrated
2. MD Physical Validation Perform enhanced sampling MD simulations on candidate conformations using GROMACS to validate pocket opening. wa-dd-gromacs (pocket_discovery / aMD / analysis) ✅ Integrated
3. Pocket and Water Network Analysis Extract candidate pocket events, pocket-volume curves, and candidate pocket structures from MD outputs, then register reusable standard pocket assets; assemble cluster representative structures into a pocket_ensemble (dynamic pocket ensemble) asset for fan-out docking. Built-in pocket analyzer + pocket ensemble in wa-dd-gromacs ✅ Integrated (pocket_ensemble and ensemble docking live since 2026-08-25); deeper fpocket/MDAnalysis scoring remains planned
4. Ligand Generation and Complex Prediction Generate ligands on validated OPEN conformations, predict complex structures via OpenFold3, and validate with FEP. wa-dd-molecule-gen, wa-dd-openfold3, wa-dd-fep ✅ Integrated

2.3 OpenFold3 Conformational Ensemble Generation Strategies

OpenFold3 does not have built-in MSA subsampling, but generates conformational diversity through two complementary strategies:

Strategy Principle Implementation
Template Perturbation OpenFold3 supports template input; different templates induce different conformations Provide 5 templates for KRAS G12C: APO (4OBE), GTP state (5VQ2), GDP state (4TQ9), Switch II OPEN (6GJ8), randomly perturbed APO
Ligand Induction OpenFold3 can directly input ligands to predict complexes; different ligands induce different conformations Input KRAS + known active ligands (ARS-1620, MRTX849) + fragment library, predict ~10 conformations

Combined, these two strategies can generate approximately 15-20 candidate conformations.

3. Validation Process (Single-Server Progressive Approach)

We will use KRAS G12C as the model target to validate this workflow. The Switch II region of KRAS G12C contains a classic cryptic pocket that opens when bound to covalent inhibitors (such as ARS-1620).

Given the current available resources being a single server (tc232/server6), we adopt a "pyramid-style progressive validation" strategy, ensuring each step can be completed in the short term and produce deterministic evidence.

3.1 Four-Level Progressive Validation Plan

Level Goal Computational Task Single-Server Duration Validation Output
L0: Baseline Validation Can FEP distinguish Open vs APO conformations? Calculate ΔΔG of ARS-1620 analogs using KRAS G12C known OPEN (6GJ8) and APO (4OBE) structures 1-2 days FEP ΔΔG result table (showing OPEN conformation ΔG significantly better than APO)
L1: AI Prediction Can OpenFold3 predict the OPEN conformation? OpenFold3 template perturbation + ligand induction, generate ~15 candidate conformations Half a day Conformational ensemble + RMSD comparison with 6GJ8
L2: Physical Validation Are AI-predicted conformations physically stable? Run 5ns aMD on AI optimal conformation + baseline conformation, compare RMSD drift 1-2 days MD trajectory + RMSD evolution curve
L3: Closed-Loop Discovery Can AI-discovered conformations induce ligand binding? On AI-validated conformations, PocketXMol generates 10 molecules, OpenFold3 predicts complex structures, Top 3 selected for FEP 3-5 days ΔΔG < 0 for new molecules (validating druggability)

3.2 Detailed Steps for Each Level

L0: Baseline Validation (Day 1-2)

  • Goal: Prove that WA-DD's FEP pipeline can physically distinguish OPEN from APO conformations.
  • Input: KRAS G12C APO (PDB: 4OBE), KRAS G12C OPEN (PDB: 6GJ8), ARS-1620 analog ligands.
  • Steps:
    1. Process 4OBE and 6GJ8 with wa-dd-protein-prep.
    2. Prepare ARS-1620 analogs with wa-dd-ligand-prep.
    3. Calculate binding ΔG of ligands in both conformations with wa-dd-fep.
  • Pass Criteria: OPEN conformation ΔG is -1.0 kcal/mol or more superior to APO conformation.

L1: AI Prediction (Day 3-4)

  • Goal: Generate candidate conformations representing different states of the Switch II region.
  • Input: KRAS G12C APO (PDB: 4OBE), using the OpenFold3 structure-prediction capability already integrated into the platform.
  • Steps:
    1. Run OpenFold3 template perturbation (5 templates) on 4OBE with wa-dd-openfold3, generating 5 conformations.
    2. Run OpenFold3 ligand induction (5 ligands) on 4OBE with wa-dd-openfold3, generating 5 conformations.
    3. Calculate RMSD between each conformation and the Switch II region of 6GJ8.
    4. Screen candidate conformations with RMSD < 3.0Å (as input for L2).
  • Pass Criteria: Find at least 1 conformation with RMSD < 3.0Å from 6GJ8.

L2: Physical Validation (Day 5-8)

  • Goal: Validate the physical stability of AI-predicted conformations.
  • Input: AI conformations screened from L1 + 6GJ8 baseline conformation.
  • Steps:
    1. Run 5ns simulations on 3 conformations each with the pocket_discovery / aMD protocol in wa-dd-gromacs.
    2. Use the built-in GROMACS pocket analyzer to output pocket_analyzer_report.json, pocket_events.csv, pocket_volume.csv, and a standard pocket asset.
    3. Compare RMSD drift between AI conformations and 6GJ8.
  • Pass Criteria: AI conformation RMSD drift < 2.0Å (comparable to 6GJ8).

L3: Closed-Loop Discovery (Day 9-14)

  • Goal: Validate that AI-discovered conformations can be used for ligand design.
  • Input: AI conformations validated from L2.
  • Steps:
    1. De novo generate 10 molecules on AI conformations with wa-dd-molecule-gen (PocketXMol).
    2. Predict complex structures of "protein + new ligand" with OpenFold3 (replacing original Uni-Dock docking).
    3. Calculate ΔΔG of Top 3 with wa-dd-fep.
  • Pass Criteria: At least 1 new molecule has ΔΔG < 0 (stronger binding than baseline ligand).

3.3 Expected Scientific Discovery Signals

  1. AI Prediction Capability Validation: What proportion of OpenFold3 conformations generated through template perturbation and ligand induction can stably exist in MD simulations, and can pocket opening be observed.
  2. Dynamic Entropy Contribution: Quantify the contribution of conformational changes to binding free energy by comparing FEP results across different conformations.
  3. Water Molecule Role: Analyze water molecule entry/exit patterns during pocket opening in MD trajectories, identifying potential "unhappy water".
  4. OpenFold3 Complex Prediction Accuracy: Compare the consistency between OpenFold3-predicted complex structures and FEP physical validation results.

4. Development Progress

4.1 Completed

  • Scientific Question Definition: Clarified the scientific value and technical path of cryptic pocket discovery.
  • Workflow Design: Completed the full data flow design from OpenFold3 conformation generation to FEP validation, as well as the single-server progressive validation plan.
  • Existing Component Integration: openfold3, gromacs, molecule-gen, fep, and related components now share one asset chain.
  • Pocket Analyzer MVP: GROMACS pocket_discovery can register md_result, candidate pocket assets, and a complete downloadable result package.
  • FEP Engine Hardening, Full-Chain Production Run (2026-08-23, server6): Fixed a three-layer failure chain (receptor clashes entering propagation, over-deviant mapped atoms entering the hybrid topology, openmmtools FIRE minimization stalling on strained systems) plus receptor guards (cofactor precheck, terminal rebuild, unresolved-sidechain ALA truncation). The first full production RBFE completed on PI3KA H1047R (job 42bdf7f8, ΔΔG = 19.02 ± 10.60 kcal/mol); fixes are committed (bfba633, 1775440, 378060f) and deployed to both amd and thor images.
  • Dynamic Pocket Ensemble and Ensemble Docking (2026-08-25, 133acb5): GROMACS pocket_discovery now parses the cluster.log table, splits cluster representative structures, recomputes per-conformer pocket residue composition, and registers a pocket_ensemble asset (cluster populations, per-pocket representative and pocket PDBs). The docking page accepts the ensemble directly in its pocket picker; Uni-Dock fans out over "pocket x ligand" runs, the pose library records WA_DD_POCKET_* properties, and the report adds cross-pocket consensus ranking. MD-derived md_structure assets are also accepted as docking/generation receptors. amd/thor images are built and deployed on server6.

4.2 In Progress

  • OpenFold3 Conformation-Generation Validation (next step): The AssetFile.size worker bug is fixed (a43b816, images after 2026-08-20 include it); template-perturbation and ligand-induced predictions are ready to resubmit on server6.

4.2.1 L0 Baseline Completed (2026-08-24)

Both formal RBFE jobs ran end to end on server6 (APO 6273dcd5, OPEN 484407ba; reproducible procedure in section 5.1 of the Chinese execution reference). The Adagrasib edge gives ΔΔΔG = -1.91 kcal/mol (OPEN favored over APO — direction consistent with the cryptic-pocket hypothesis; combined uncertainty ±7.6, not statistically significant, needs longer sampling or a closer ligand pair). The Sotorasib edges diverged in both runs (-99 / 3e+13) and motivated the low-reliability annotation (7a55400, deployed).

4.3 Planned

  • Pocket Analyzer Depth: Cluster representative structures → pocket_ensemble are delivered (see 4.1); continue to add fpocket, MDAnalysis, water-network analysis, and per-frame pocket-event scoring without changing the standard pocket asset contract.
  • Case Evidence: Fill the KRAS G12C L0-L3 example with real outputs, screenshots, and threshold evidence.
  • Protein Preparation Source Fix: Default prepared proteins still carry chain-internal termini (no H1/H2/H3, no OXT); FEP rebuilds them at run time, a source-side fix is pending.
  • ABFE Evaluation: FEP currently supports RBFE only; cross-conformation comparisons rely on paired RBFE differences. Introducing an ABFE protocol would enable absolute ΔG.

4.4 Milestones

Timeline Milestone Status
Phase 1 (Week 1) Run L1 conformation generation and candidate screening with wa-dd-openfold3. Not started (bug fixed, ready to retry)
Phase 2 (Week 2) Run L0 baseline FEP and L2 physical validation with wa-dd-gromacs. ✅ L0 done (2026-08-24, both formal RBFE jobs end to end; ΔΔΔG = -1.91 kcal/mol, right direction, large uncertainty); L2 waits on L1
Phase 3 (Week 3) Use the GROMACS pocket analyzer to emit standard pocket assets and complete L3 closed-loop discovery. Not started
Phase 4 (Week 4) Add fpocket/MDAnalysis scoring, water-network analysis, and complete case evidence. Not started
(Inserted 2026-08-23) FEP engine hardening + full PI3KA chain. ✅ Done

5. Platform UI Operations and Asset Flow

This case targets an already deployed WA-DD instance. Regular users do not need to install OpenFold3 manually, download model weights, write Dockerfiles, or run commands inside containers; those are platform-operations concerns. The L0-L3 workflow should be completed through the Web UI as much as possible, with every step registered as project assets.

5.1 UI Entry Points

Work Page Main Inputs Main Outputs
Structure preparation Protein Processing PDB ID or uploaded PDB/CIF protein / prepared_protein assets
Ligand preparation Ligand Processing SDF, SMILES, table, or drawn molecule ligand / prepared_ligand assets
Conformation generation Structure Prediction Protein asset, template/ligand-induced parameters OpenFold3 structure-result assets
MD and pocket analysis GROMACS / MD Protein/complex/topology/trajectory assets, pocket_discovery protocol md_result, pocket_analyzer_report.json, pocket_events.csv, pocket_volume.csv, standard pocket assets, dynamic pocket_ensemble asset
Molecule generation Molecule Generation Standard pocket asset and generation constraints Candidate-molecule SDF assets
FEP validation FEP / Analysis Congeneric ligands, complexes, or docking/generation outputs FEP result tables and downloadable result assets

5.2 Standard Asset Chaining

  • GROMACS pocket_discovery writes pocket-analysis outputs into the same task result and creates a standard pocket asset when a representative structure is available.
  • pocket_discovery also registers a pocket_ensemble (dynamic pocket ensemble) asset; the Docking page pocket picker accepts it directly without a separate receptor — docking fans out over each pocket's cluster representative structure and tags poses with pocket origin and cluster population.
  • Standard pocket assets can be selected directly in the Docking and Molecule Generation pages; users do not need to manually copy center coordinates or file paths.
  • The task-output button in the upper-right task panel lists registered output assets; "download all outputs" packages all result files for that task.
  • Individual result files remain openable or downloadable from asset details, including json, csv, xvg, pdb, gro, and log outputs.

5.3 Human Review Still Required

The platform handles the data flow and compute scheduling, but the cryptic-pocket case still requires scientific judgment:

  • Choose APO, OPEN, ligand-induced, or template-perturbed conformations as controls.
  • Decide whether RMSD, pocket volume, candidate residues, and FEP ΔΔG support the conclusion that a pocket is druggable.
  • Manually review the current MVP pocket analyzer's geometry-defined candidate pockets; deeper fpocket, MDAnalysis, water-network, and multi-frame event scoring remain planned enhancements.