Virtual Drug Screening

Druggable-pocket screening on the PfSnf2L ATPase domain

1. What makes this target hard

Three features dictate the method choice:

• No experimental structure. You have an AF3 model only. Predicted models are biased toward a single, usually ligand-competent ground state.
• Snf2 ATPases are conformational machines. The catalytic core is two RecA-like lobes joined by a flexible linker; the nucleotide site forms only at the closed lobe interface, and the ISWI-specific AutoN/NegC/brace elements dock onto the lobes in a nucleosome- and ATP-state-dependent way. Any pocket worth having is likely state-dependent, i.e. cryptic. A single-structure pocket scan will systematically miss them.
• Selectivity is the whole game. PfSnf2L is ISWI-related, with human SMARCA1/SMARCA5 orthologues [1]. The orthosteric ATP pocket is a conserved P-loop — it will score as "druggable" and is almost useless for selective antimalarials. The screen must therefore be built to rank non-orthosteric and parasite-divergent pockets, using the ATP site only as an internal positive control.

Conclusion: the state-of-the-art answer is not "run fpocket on the AF3 model." It is an ensemble-based, consensus-detected, ligandability-validated, selectivity-filtered cascade.

2. Recommended pipeline

Stage 0 — Define the folded ATPase domain (input curation)

• Pull the AF3 model + PlasmoDB/UniProt entry; annotate with InterProScan/HMMER for Pfam SNF2_N (PF00176) and Helicase_C (PF00271).
• Parse domains from the PAE matrix (PAE-clustering / Chainsaw / DPAM) rather than by eye — this is what separates rigid-body units from linkers — and cross-filter with pLDDT ≥ 70.
• Define the excised construct as: SNF2_N lobe 1 → linker → Helicase_C lobe 2, including the brace helix, trimming low-pLDDT/low-PAE-confidence insertions. Plasmodium proteins carry long low-complexity asparagine insertions; internal ones that sit inside a lobe must be excised and the backbone closed, not just truncated at the ends.
• Deliverable: explicit residue ranges (domain start–end, plus each excised internal insertion), a per-residue pLDDT table, and structural validation by superposition onto experimental ISWI/Snf2–nucleosome structures (TM-score + RMSD over the RecA folds).

Stage 1 — Generate a conformational ensemble, not a structure

• MSA subsampling with AlphaFold2/3: repeatedly subsample the MSA to sample open/closed lobe states. This is the cheapest way to reach alternative states, and it demonstrably uncovers cryptic pockets — in 6 of 10 known cryptic-pocket cases AlphaFold samples the open state; for plasmepsin II, a malaria aspartic protease, only a partial opening was captured, and simulations seeded from the AlphaFold ensemble were needed to complete it [4].
• Generative ensemble models: BioEmu and/or AlphaFlow for a broader, physics-flavoured ensemble.
• AF3 holo-modelling: additionally model the domain with ADP, ATP·Mg²⁺ and with nucleosomal DNA, to obtain the closed/engaged states that apo prediction will not give you.
• Seeded MD: 3–5 × 500 ns explicit-solvent MD launched from the diverse ensemble members (apo and ATP-bound). If a candidate site opens only partially, escalate to enhanced sampling (metadynamics with data-driven CVs, e.g. AlphaFold-SFA [11]).

Stage 2 — Consensus pocket detection across the ensemble

• fpocket on every ensemble member and mdpocket across MD trajectories — mdpocket is the tool built precisely for transient/allosteric site tracking [6,7].
• P2Rank for ML-based, template-free ligandability ranking [8].
• PocketMiner / AE-PocketMiner for residue-level cryptic-site prediction from single structures; AE-PocketMiner additionally predicts the allosteric coupling of each predicted cryptic site, which is exactly the prioritisation signal you need for an enzyme where you want to block the ATPase cycle from a distance [2,3].
• Score by consensus: a site found by ≥2 orthogonal detectors, persistent in >20–30 % of MD frames, with fpocket DrugScore ≥ 0.5.

Stage 3 — Ligandability, not just geometry

Pocket detection tells you a cavity exists; it does not tell you a drug-like ligand can bind it. Run mixed-solvent MD (MixMD / SILCS) on the top 5–8 consensus sites: competitive water-vs-probe occupancy with full protein flexibility, which maps binding hot spots and distinguishes genuinely druggable surfaces from crystal-packing artefacts by requiring higher probe occupancy than the packing faces [10]. This is the quantitative gate — everything upstream is hypothesis generation.

Stage 4 — Selectivity and genetics triage

• Run the identical pipeline on AF models of human SMARCA1 and SMARCA5 ATPase domains; compute per-pocket sequence divergence and shape dissimilarity.
• Overlay MalariaGEN/PlasmoDB polymorphism data — pockets lined by variable residues are pre-loaded resistance liabilities.
• Cross-check pocket-lining residues against the conserved ISWI catalytic motifs and the AutoN/NegC elements to build a mechanistic hypothesis for each site.

Stage 5 — Prospective test

Ensemble docking (and/or Boltz-2 co-folding) of a focused library into the two or three surviving pockets, with decoy/DUD-E controls, to confirm the site can discriminate. Only then is a virtual screen justified.

Built-in controls (non-negotiable): the ATP site must come out as top-ranked druggable; a known cryptic-pocket protein (e.g. TEM β-lactamase or Ebola VP35) should be run through the same pipeline to calibrate thresholds; report per-pocket persistence and variance across ensemble members, not a single number.

3. Why this is the best approach

• Ensembles beat single structures, and this is now measured. The 2026 Bowman-lab benchmark [5] shows AI and MD methods reliably predict the direction of change in cryptic-pocket opening but none reliably predict absolute opening probabilities, and all degrade badly for pockets open <1 % of the time. The correct inference is to use ensembles for ranking and hypothesis generation, and to treat any single method's absolute score as untrustworthy — which is exactly why Stage 2 is a consensus and Stage 3 is an independent physical assay.
• Single-structure holo-trained predictors overestimate their own performance on apo/predicted inputs. CryptoBench was built to expose this: current methods trained and evaluated on ligand-bound conformations give unrealistic performance expectations for cryptic sites [9]. An AF3 apo model is precisely the failure mode.
• Speed where speed is free, physics where it matters. PocketMiner runs >1,000-fold faster than simulation-based detection at ROC-AUC 0.87 [2], so it costs nothing to use as a filter; MixMD is expensive, so it is applied only to the survivors.
• It answers the question you actually have. A geometric pocket list is not a drug-discovery result. A ranked set of sites with persistence, ligandability hot-spot maps, allosteric coupling and a human-orthologue selectivity delta is.

Rejected alternatives: fpocket/P2Rank on the single AF3 model (fast, but misses every cryptic site and returns ~10–15 pockets with no way to rank them); SiteMap/Site Finder alone (useful as focus-point finders but commercial and still single-structure [ref. Ge et al. 2024]); brute-force long MD from the AF3 model without AlphaFold seeding (demonstrably fails to sample openings that seeded simulations reach [4]); pure ML cryptic-site prediction without physical validation (qualitative only [5]).

4. Paste this into Claude Science

Target: PfSnf2L, PF3D7_1104200 (P. falciparum 3D7), ISWI-related Snf2
ATPase. Goal: identify and rank druggable pockets in the folded ATPase
domain, prioritising sites that are NOT the conserved ATP pocket and that
differ from human SMARCA1/SMARCA5.

Work in stages and stop after each for my approval. Do not proceed on
assumptions; if a value is not in a file or database you queried, say so.

STAGE 0 — Domain definition
1. Retrieve the PF3D7_1104200 sequence (PlasmoDB/UniProt) and the
   AlphaFold3 model I provide (or generate one and state which).
2. Annotate Pfam SNF2_N (PF00176) and Helicase_C (PF00271) via InterProScan
   or HMMER against Pfam.
3. Parse structural domains from the PAE matrix (PAE clustering), filter by
   pLDDT >= 70, and define the folded ATPase core = lobe 1 + linker + lobe 2
   + brace.
4. Identify low-complexity / low-pLDDT internal insertions inside the lobes
   and propose an excised construct.
5. OUTPUT: a table of residue ranges (domain start-end, each excised
   insertion), mean pLDDT per segment, a pLDDT plot, the trimmed PDB, and a
   structural superposition (TM-score, RMSD) onto experimental ISWI/Snf2
   structures from the PDB. State explicitly which residue numbering the
   coordinates refer to.

STAGE 1 — Ensemble (after my approval)
Generate 100-200 conformers by AlphaFold MSA subsampling; add BioEmu and/or
AlphaFlow ensembles; additionally model ATP-Mg and ADP bound states. Cluster,
and report lobe-closure distribution.

STAGE 2 — Pocket detection
Run fpocket on all conformers, mdpocket on any MD trajectories, P2Rank, and
PocketMiner / AE-PocketMiner. Build a consensus table: site ID, lining
residues, volume, fpocket DrugScore, P2Rank rank, PocketMiner score,
AE-PocketMiner allosteric coupling, fraction of ensemble in which the site
is open. Flag the ATP site explicitly as the positive control.

STAGE 3 — Ligandability
Mixed-solvent MD (MixMD-style, e.g. 5% v/v acetonitrile / isopropanol /
pyrimidine probes) on the top 5-8 consensus sites; report probe occupancy
hot-spot maps and rank sites by hot-spot strength.

STAGE 4 — Selectivity
Run stages 0-2 identically on human SMARCA1 and SMARCA5 ATPase domains.
Report per-pocket sequence identity and shape similarity to PfSnf2L, and
overlay PlasmoDB/MalariaGEN polymorphism data on pocket-lining residues.

STAGE 5 — Report
A ranked shortlist of 2-3 candidate pockets with: residue lists, mechanistic
rationale relative to the Snf2 ATPase cycle and ISWI AutoN/NegC regulation,
selectivity delta vs human orthologues, and the main caveats. Cite literature
with numbered brackets tied to PMIDs, and list references at the end with PMID,
full citation, and three sentences on the key findings of each.

References (retrieved via PubMed)

[1] Plasmodium blood stage development requires the chromatin remodeller Snf2L. Nature (2025). PMID: 39972139 — DOI

Identifies PfSnf2L (PF3D7_1104200) as an ISWI-related ATPase that actively repositions P. falciparum nucleosomes in vitro. Conditional knockout shows the enzyme is essential for asexual blood-stage development and controls just-in-time transcription. The domain architecture is divergent from canonical ISWI, with altered autoregulatory and HAND–SANT–SLIDE regions — the basis for hoping a parasite-selective pocket exists.

[2] Meller A, et al. Predicting locations of cryptic pockets from single protein structures using the PocketMiner graph neural network. Nat Commun 14:1177 (2023). PMID: 36859488 — DOI

A graph neural network trained on MD-observed pocket openings predicts cryptic-pocket locations from one static structure. On 39 experimentally confirmed cryptic pockets it reaches ROC-AUC 0.87, over 1,000-fold faster than simulation-based detection. Proteome-wide application suggests more than half of apparently pocketless proteins harbour cryptic sites.

[3] Zhang S, et al. AE-PocketMiner uses attention to simultaneously predict cryptic pockets and their allosteric coupling. bioRxiv (2026). PMID: 42239053 — DOI

Adds an attention mechanism so that cryptic-pocket location and allosteric coupling are predicted together from a single structure. It outperforms earlier cryptic-pocket predictors and recapitulates known allosteric interactions. Newly predicted pockets and pocket-controlling mutations were confirmed experimentally, making it the current best tool for prioritising rather than merely listing sites.

[4] Meller A, Bhakat S, Solieva S, Bowman GR. Accelerating cryptic pocket discovery using AlphaFold. J Chem Theory Comput 19:4355–4363 (2023). PMID: 36948209 — DOI

AlphaFold ensembles generated by stochastic MSA subsampling sampled the open cryptic state in 6 of 10 benchmark proteins. For plasmepsin II only a partial opening was captured, but MD seeded from the AlphaFold ensemble reached the full opening where MD from the apo crystal structure did not. Markov state models from those seeded simulations reproduced metadynamics free-energy landscapes at much lower cost.

[5] Zhang S, Miller JJ, Bowman GR. How well can AI and physics-based simulations predict the probability a cryptic pocket is open? J Chem Theory Comput 22:3839–3850 (2026). PMID: 41945898 — DOI

Benchmarks AlphaFlow, BioEmu, PocketMiner, CryptoBank and MD against experimentally characterised pocket populations in Ebola VP35 and TEM β-lactamase. Most methods correctly predict whether a mutation raises or lowers opening probability, but none predict absolute probabilities reliably, and all fail for pockets open <1 % of the time. MD comes closest for wild-type proteins — the empirical justification for using AI for triage and physics for validation.

[6] Le Guilloux V, Schmidtke P, Tuffery P. Fpocket: an open source platform for ligand pocket detection. BMC Bioinformatics 10:168 (2009). PMID: 19486540 — DOI

Introduces the Voronoi/alpha-sphere geometric algorithm that remains the field's workhorse for cavity detection. It recovers 94 % (holo) and 92 % (apo) of known pockets within the top three ranks, outperforming contemporaneous tools while being faster. Its open modular design allows custom descriptors and druggability scoring, which is why downstream tools build on it.

[7] Schmidtke P, Le Guilloux V, Maupetit J, Tufféry P. fpocket: online tools for protein ensemble pocket detection and tracking. Nucleic Acids Res 38:W582–W589 (2010). PMID: 20478829 — DOI

Adds mdpocket, which tracks pockets along MD trajectories, and hpocket for homologous structure families. This turns static cavity detection into a tool for transient and allosteric site discovery. It is the direct route from a conformational ensemble to a per-site persistence value.

[8] Krivák R, Hoksza D. P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure. J Cheminform 10:39 (2018). PMID: 30109435 — DOI

Predicts ligandability of local chemical neighbourhoods on the solvent-accessible surface using random forests, template-free. It outperforms Fpocket, SiteHound, MetaPocket 2.0 and DeepSite on standard benchmarks. Sub-second runtimes and a stand-alone CLI make it ideal for scanning hundreds of ensemble members.

[9] Škrhák V, Novotný M, Feidakis CP, Krivák R, Hoksza D. CryptoBench: cryptic protein–ligand binding sites dataset and benchmark. Bioinformatics 41(1) (2024). PMID: 39693053 — DOI

Builds a 1,107-structure apo–holo benchmark specifically for cryptic sites, with predefined cross-validation splits. It shows that evaluating on holo conformations produces unrealistically optimistic performance estimates. A sequence-based protein-language-model baseline outperformed both PocketMiner and P2Rank on AUC, AUPRC, MCC and F1 — evidence that no single detector should be trusted alone.

[10] Ung PMU, Ghanakota P, Graham SE, Lexa KW, Carlson HA. Identifying binding hot spots on protein surfaces by mixed-solvent molecular dynamics: HIV-1 protease as a test case. Biopolymers 105:21–34 (2016). PMID: 26385317 — DOI

Establishes MixMD with full protein flexibility and explicit water–probe competition for hot-spot mapping. Low probe concentration (5 % v/v) gave markedly better signal-to-noise and fewer spurious hot spots than 50 % w/w. It recovered the catalytic site plus experimentally confirmed allosteric sites, and defined an occupancy cut-off separating druggable surfaces from crystal-packing faces.

[11] Vats S, Bobrovs R, Söderhjelm P, Bhakat S. AlphaFold-SFA: accelerated sampling of cryptic pocket opening, protein–ligand binding and allostery. PLoS One 19:e0307226 (2024). PMID: 39190764 — DOI

Uses unsupervised slow feature analysis on short MD launched from AlphaFold ensembles to extract slow collective variables automatically. Metadynamics along those CVs opened a deep cryptic pocket within hundreds of nanoseconds, beyond the reach of microsecond unbiased MD. It gives a concrete escalation path when Stage 1 yields only partial openings.

[12] Ge Y, Pande V, Seierstad MJ, Damm-Ganamet KL. Exploring the application of SiteMap and Site Finder for focused cryptic pocket identification. J Phys Chem B 128:6233–6245 (2024). PMID: 38904218 — DOI

Tests two widely used commercial pocket finders against 136 known cryptic pockets from the PocketMiner and CryptoSite sets. They work well as "focus point" detectors that nominate regions for subsequent enhanced sampling. This is the practical solution to the chicken-and-egg problem of needing a reaction coordinate before you know where the pocket is.