# AlphaFold Confidence Scores for Drug Discovery By the end of this article you will be able to download an AlphaFold model, extract its per-residue and pairwise confidence values, decide whether a candidate binding site is trustworthy enough for docking, compute a pocket-level confidence summary, and run a first docking pass against the model. You will need a computer with Python 3, the Biopython and NumPy packages, fpocket for pocket detection, the AutoDock Vina Python package, and a browser for the AlphaFold [Protein Structure](/blog/guides/protein-structure) Database or the AlphaFold server. Everything shown here runs on a standard laptop. No GPU is required for the confidence analysis itself. ## Why Confidence Scores Decide Whether a Predicted Binding Site Is Usable Structure-based drug discovery predicts molecules that bind tightly to a three-dimensional protein structure. Its usefulness has always been limited by the number of experimentally solved structures available, especially for poorly characterized disease targets. AlphaFold changed that by producing high-confidence predicted structures for a large fraction of the known proteome, although prediction confidence and accuracy vary considerably between targets and between regions within a single target [1]. The catch is that drug discovery does not need the whole protein to be right. It needs the binding site to be right. Docking and molecular dynamics predictions become unreliable when the geometry of the target binding site is uncertain, which is why careful binding-site characterization before any ligand calculation is treated as a foundational step rather than an afterthought [2]. A model can have a beautifully resolved kinase domain and a floppy, meaningless loop sitting exactly where you want to place your ligand. Confidence scores are how you tell the difference. Every AlphaFold output ships with them, and learning to read them is the single highest-leverage skill in this workflow. A related family of methods, binding-site comparison tools that search for similar pockets across proteins, assumes you know where the pocket is and how reliable it is, which makes the confidence assessment upstream of nearly everything else you will do, including drug repurposing and polypharmacology analyses [3]. ## The Confidence Numbers You Will Actually Use AlphaFold and its successors emit three families of scores. You will touch all three in any serious drug discovery project. **pLDDT (per-residue local distance difference test).** A number from 0 to 100 assigned to every residue, stored in the B-factor column of the PDB file and in the B_iso_or_equiv field of the mmCIF. It measures how well the local structure around each residue agrees with the reference. High pLDDT means the residue's position relative to its neighbors is trustworthy. It says nothing about whether the whole domain is positioned correctly relative to other domains. **PAE (predicted aligned error).** An N by N matrix in angstroms. Entry (i, j) answers this question: if I align the predicted and true structures on residue j, how far off do I expect residue i to be? Low PAE between two residues means their relative arrangement is reliable. This is the score that tells you whether a pocket on one domain is positioned correctly relative to a partner chain or a neighboring domain. **pTM and ipTM.** pTM predicts the TM-score, a global fold similarity measure where values above roughly 0.5 indicate the same fold. ipTM is the interface analog used by AlphaFold-Multimer and AlphaFold 3 for complexes. The values above roughly 0.8 for ipTM and below roughly 0.6 come from the AlphaFold-Multimer work and are now the widely used reading convention: high ipTM means a well-modeled interface, low ipTM means the interface arrangement is probably wrong. The AlphaFold Protein Structure Database documentation, maintained by Google DeepMind together with EMBL-EBI, describes the standard pLDDT interpretation bands used in the viewer color scheme. pLDDT above 90 is very high accuracy, 70 to 90 is confident for the backbone with side chains less certain, 50 to 70 is low confidence, and below 50 typically flags an intrinsically disordered region rather than a structured but poorly modeled segment. | Score | Where it lives | Scale | What it measures | Binding-site use | |--|--|--|--|--| | pLDDT | B-factor column of the PDB, B_iso_or_equiv in mmCIF | 0 to 100 per residue | Local accuracy around each residue | Gate which pocket residues you trust for docking | | PAE | Separate JSON download or mmCIF quality metrics | Angstroms, pairwise matrix | Expected positional error of residue i when aligned on residue j | Check that pocket residues and partner domains are mutually consistent | | pTM | Summary metrics with the prediction | 0 to 1 | Global fold similarity (TM-score) | Quick screen of whether the model is worth using at all | | ipTM | Summary metrics for complexes (Multimer, AF3) | 0 to 1 | Interface accuracy | Judge receptor and partner or peptide interfaces before pocket analysis | A note on newer tools. Benchmarks comparing AlphaFold 2, AlphaFold 3, Chai-1, and related deep learning models on GPCR and peptide systems show that the ranking behavior differs between methods, so the confidence semantics you learned for AlphaFold 2 do not transfer blindly to every new predictor [4]. Always check which scores your specific tool emits. ## What the Evidence Says About AlphaFold Models at Binding Sites Before running commands, it helps to know where predicted structures succeed and fail in published benchmarks, because that tells you which confidence checks actually predict downstream docking success. The most directly relevant study examined AlphaFold 2 models for GPCRs, the largest class of drug targets. The finding has two halves. First, AF2 models capture binding pocket structures far more accurately than traditional homology models, with pocket errors nearly as small as the differences between experimental structures of the same protein crystallized with different ligands bound. Second, and this is the half people forget, the accuracy of ligand binding poses predicted by docking into AF2 models was not significantly higher than docking into traditional homology models, and was much lower than docking into experimental structures of the same protein [5]. Pocket geometry and dockable pose accuracy are different things, and high pLDDT at the pocket guarantees only the first one. Peptide and complex work tells a similar nuanced story. In a benchmark across multiple deep learning structure predictors, AF2 reproduced the correct binding modes in nearly all of 67 recently solved complexes, about 94 percent, outperforming both AF3 and Chai-1 on that measure, and structure-aware models reached an area under the curve of 0.86 when classifying true ligands against 1240 decoy peptides. Rescoring predicted structures on local interactions improved recovery of the correct principal ligand, while purely deep-learning-based scoring did not [4]. The lesson for binding-site work is that local interaction detail, exactly what pLDDT and PAE describe, matters more than global model identity. On the success side, an end-to-end AI drug discovery pipeline used AlphaFold structures of targets with no experimental structures, combined with a biocomputational target identification platform and a generative chemistry platform, to reach a confirmed hit against cyclin-dependent kinase 20 (CDK20). The first round produced a molecule with a binding constant Kd of 9.2 ± 0.5 μM within 30 days of target selection after synthesizing only 7 compounds, with a second generation round delivering a more potent analogue [6]. That was achieved on predicted structures despite varying confidence across the target, which is exactly the scenario the confidence-reading workflow in this article supports. Review coverage of AlphaFold-driven structure-based drug discovery now spans target identification, binding-site prediction, [molecular docking](/knowledge/bioinformatics/docking-algorithms-autodock-glide-and-beyond), virtual screening, lead optimization, and structure-activity relationship work, and consistently stresses the same caveat: confidence and accuracy vary by target and by region within a target, so per-region inspection is mandatory rather than optional [1]. ## Step-by-Step: Reading AlphaFold Confidence at a Binding Site Before You Dock This walkthrough doubles as a compact AlphaFold tutorial for the docking-aware reader. We use human EGFR (UniProt P00533) as the example because its kinase pocket is well characterized, but the commands are identical for any target. The biological logic: extract per-residue confidence, define the pocket, summarize confidence over pocket residues only, then dock. ### Step 1. Get the model from the AlphaFold database or generate one with the AlphaFold server If your protein is already in the AlphaFold Protein Structure Database, download the model directly: ```bash curl -O https://alphafold.ebi.ac.uk/files/AF-P00533-F1-model_v4.pdb ``` Expected output is a silently created file. Check it: ```bash head -2 AF-P00533-F1-model_v4.pdb ``` You should see HEADER and REMARK lines mentioning the AlphaFold database. The B-factor column (columns 61 to 66 of ATOM records) carries the pLDDT. If your protein is not in the database, or you need it in a complex with a ligand, peptide, nucleic acid, or metal ion, use the AlphaFold server, which is Google DeepMind's free web interface for AlphaFold 3 predictions on non-commercial work. You submit sequences for each component, optionally specify bonds or modifications, and receive a structure with pLDDT in the B-factors plus a PAE matrix and interface metrics. For a receptor and peptide ligand, the ipTM and the PAE cross-chain block are the numbers to look at first, because a low ipTM means the modeled interface is likely wrong before you ever discuss a pocket. A practical rule for choosing between the two sources: database models are fast, versioned, and reproducible, which matters when you need to document exactly which model version a virtual screening campaign used. The server gives you flexibility with cofactors, ions, and complexes, at the cost of outputs that depend on the submission. ### Step 2. Extract per-residue pLDDT and filter the disordered tail Full-length AlphaFold models include low-confidence disordered termini that generate phantom pockets. Strip them before pocket detection. ```python from Bio.PDB import PDBParser import numpy as np parser = PDBParser(QUIET=True) structure = parser.get_structure("EGFR_AF", "AF-P00533-F1-model_v4.pdb") residues = [] for residue in structure.get_residues(): if residue.id != " ": # skips waters and HETATM records continue if "CA" not in residue: continue residues.append((residue.id[7], residue["CA"].bfactor, residue["CA"].coord)) ids = np.array([r for r in residues]) plddt = np.array([r[7] for r in residues]) coords = np.array([r[3] for r in residues]) structured = plddt > 70 print(f"total residues: {len(residues)}") print(f"pLDDT > 70 residues: {structured.sum()}") print(f"pLDDT < 50 residues: {(plddt < 50).sum()}") print(f"mean pLDDT: {plddt.mean():.1f}") ``` Expected output looks like this (your exact numbers depend on the model version): ``` total residues: 1210 pLDDT > 70 residues: 962 pLDDT < 50 residues: 141 mean pLDDT: 78.3 ``` The high count of residues below 50 is normal for a full-length receptor with unstructured regions. Those residues are almost certainly not part of a druggable pocket, and leaving them in confuses every downstream pocket detector. ### Step 3. Download and inspect the PAE matrix The AlphaFold database model page provides a predicted aligned error download for each entry, typically a JSON file. Fetch it from the same page you got the PDB from and load it: ```python import json with open("AF-P00533-F1-pae.json") as handle: data = json.load(handle) pae = np.array(data["pae"]) # N x N matrix in angstroms print(pae.shape) ``` Expected output: ``` (1210, 1210) ``` Interpretation for binding-site work: pick the row and column indices of your candidate pocket residues. If the PAE between every pair of pocket residues is in the low single-digit angstrom range, the pocket holds together as a rigid unit. If a subset of pocket residues shows PAE of 10 Å or more against the rest of the domain, those residues are positioned unreliably relative to the pocket, and docking will treat their coordinates as facts when they are guesses. ### Step 4. Define the binding site two ways and compare The most defensible definition combines a known ligand-bound experimental structure, when one exists for the protein or a close homolog, with automated pocket detection on the model. **Way one, ligand-derived.** Take a co-crystal structure, align it on the AlphaFold model, and collect every model residue with any atom within 8 Å of the bound ligand. If your target has no structures, a ligand from a homologous protein after structural alignment is the standard fallback. **Way two, geometric.** Run fpocket on the low-confidence-filtered model: ```bash fpocket -f AF-P00533-F1-model_v4_clean.pdb ``` Expected output is a directory named `AF-P00533-F1-model_v4_clean_out/` containing pocket sub-PDB files and a pocket descriptor text file. fpocket ranks pockets by a volume and druggability score. Cross-check that the top-ranked geometric pocket overlaps the ligand-derived pocket when you have both. Binding-site-driven modeling approaches, such as methods that take the 3D structure of the pocket itself as input and predict the ligand preferences that site supports, depend on this pocket definition being correct, which is why the cross-check matters. ### Step 5. Compute a pocket-level confidence summary ![Worked calculation of mean pocket pLDDT, minimum pLDDT and mean internal PAE for a 31-residue kinase pocket using the article's example output.](/article-images/illustrations/alphafold-confidence-drug-discovery-binding-site-pocket-confidence-summary.webp) This is the number you will actually quote to collaborators. ```python pocket_mask = np.linalg.norm(coords - ligand_center, axis=1) <= 8.0 pocket_ids = ids[pocket_mask] pocket_plddt = plddt[pocket_mask] rows = pocket_ids - 1 pocket_pae_internal = pae[np.ix_(rows, rows)] print(f"pocket residues: {len(pocket_ids)}") print(f"mean pocket pLDDT: {pocket_plddt.mean():.1f}") print(f"min pocket pLDDT: {pocket_plddt.min():.1f}") print(f"residues pLDDT < 70: {(pocket_plddt < 70).sum()}") print(f"mean internal PAE (A): {pocket_pae_internal.mean():.1f}") ``` Expected output for a well-behaved kinase pocket: ``` pocket residues: 31 mean pocket pLDDT: 91.4 min pocket pLDDT: 68.2 residues pLDDT < 70: 2 mean internal PAE (A): 2.3 ``` Decision guidance that follows from the published benchmarks rather than from optimism. A mean pocket pLDDT in the 90s with low internal PAE means the pocket geometry is about as trustworthy as it gets from a predicted model, and the Karelina benchmark says you should still expect pose predictions worse than what an experimental apo structure would give you [5]. A mean pocket pLDDT in the 70s with a couple of low-confidence side chains is workable for hit identification but demands experimental validation of every prediction. A mean in the 50s to 60s means treat the pocket coordinates as a hypothesis, not a structure. ### Step 6. Build the docking inputs and run a first docking pass AlphaFold models have no hydrogens and standard docking engines want PDBQT format. Prepare the receptor: ```bash prepare_receptor4.py -r AF-P00533-F1-model_v4_clean.pdb -o EGFR_AF.pdbqt -A hydrogens ``` Define the search box from the pocket residue coordinates computed in step 5, with a margin of roughly 4 to 6 Å on each side of the pocket centroid, then dock using the Vina Python API: ```python from vina import Vina box_center = coords[pocket_mask].mean(axis=0) box_size = np.array([22.0, 22.0, 22.0]) v = Vina(sf_name="vina") v.set_receptor("EGFR_AF.pdbqt") v.set_ligand_from_file("ligand.pdbqt") v.compute_box(center=box_center.tolist(), box_size=box_size.tolist()) v.dock(exhaustiveness=32, n_poses=20) v.write_poses("EGFR_AF_docked.pdbqt", n_poses=5, overwrite=True) print("best affinity (kcal/mol):", v.energies(n_poses=1)) ``` Expected output: ``` best affinity (kcal/mol): -9.7 ``` Interpret the score against a reference, never in isolation. The correct reference is a docking of the same ligand into an experimental structure of the same or a homologous protein, docked with the same protocol. A score of minus 9.7 kcal/mol into a predicted structure means nothing until you know what the same ligand scores into a crystal structure under identical settings. ### Step 7. Sanity checks before you believe anything Three checks catch most failures. **Redock a known ligand.** If any ligand-bound experimental structure exists for your target family, dock that ligand into both the experimental structure and the AlphaFold model and measure the pose RMSD between the two best poses. Large divergence, more than about 3 Å, is a red flag on the model's usability for pose work, consistent with the published finding that docking accuracy to AF2 GPCR models did not beat homology models [5]. **Check the low-confidence residues in the pocket.** List them. If a residue with pLDDT below 70 sits where your ligand's key hydrogen bond would go, that interaction is a guess. **Check PAE from pocket to partner domains.** For an allosteric pocket near a domain interface, high PAE between the pocket domain and the partner domain means the relative geometry that defines the pocket shape may be wrong even if every pocket residue has pLDDT of 95. ## Tips That Make the Difference in Practice **Never dock against the full-length raw model.** Strip residues with pLDDT below 50 first. They are usually disordered tails and they generate fake pockets that waste weeks. **Treat pLDDT as a spatial filter, not a quality certificate.** A residue with pLDDT of 95 in a domain whose PAE against the pocket domain is 12 Å is confidently in the wrong place relative to the pocket. **Use the model version in your records.** AlphaFold database models are versioned, and pocket-adjacent predictions can change between versions. Screen results are only reproducible if you log the exact filename and version. **For complexes, lead with ipTM and cross-chain PAE.** When you model a receptor and a peptide or a protein partner on the AlphaFold server, ipTM below the 0.6 threshold commonly used in the AlphaFold-Multimer literature means you should not trust interface residues regardless of their individual pLDDT. **Dock into multiple states when you can.** A single predicted model represents one conformational snapshot, usually resembling an apo state. Where a ligand-bound experimental structure exists, dock against both and note which poses are consistent. **Follow low-confidence regions to opportunity, not just to failure.** Regions flagged by MD-based ensemble methods as harboring hidden pockets, like the additional pockets beyond the active site revealed by molecular dynamics simulations on Trypanosoma cruzi spermidine synthase, are exactly the regions a static AlphaFold model cannot show on its own [7]. The confidence scores tell you where the static model is honest, and the honest gaps tell you where to spend simulation time. ## Where Confidence-Low Regions Become Drug Discovery Opportunities The static AlphaFold model is one sample from a conformational ensemble. Cryptic binding sites, defined as pockets that exist in the ligand-bound state but not in the apo form, are invisible to any single static model, including a perfectly accurate one. Current approaches to finding them fall into molecular dynamics families such as Markov state models, enhanced sampling, and cosolvent MD, and machine learning families built on support vector machines, random forests, and neural networks, with ML methods generally more cost- and time-effective but limited by scarce training data [8]. The PPM1D case shows how confidence-aware use of AlphaFold feeds this pipeline directly. PPM1D phosphatase lacked crystal structures, so researchers started from an AlphaFold-predicted structure, ran molecular dynamics simulations initiated from it, and built a Markov state model that revealed a cryptic pocket at the flap and hinge interface, the region where known allosteric inhibitors bind. Deep learning was then used to predict pose quality for docked compounds [9]. The AlphaFold model was the starting point, and the confidence context was what made the team treat it as a starting point rather than an endpoint. Binding site and dynamics information also plugs into affinity and interaction prediction. Dual-modality neural networks that fuse sequence and graph structure features with a binding-site-focused graph construction have shown strong generalization on completely unseen drugs and targets, with improvements of over 8 percent over existing methods in drug-target affinity prediction [10]. Simpler deep learning models that feed motif-rich binding pocket subsequences, rather than full-length sequences, have shown better or comparable performance on MSE and AUPR metrics with lower computational cost [11]. Both lines of work reinforce the same principle: the binding site, and by extension its confidence, carries most of the predictive signal. Water complicates everything. Thermodynamic calculations of water behavior in binding sites, as implemented in tools such as WaterMap, quantify which waters are displaced favorably by a ligand and which are stable and should be respected in design. These calculations assume an accurate binding-site geometry, which routes straight back to confidence assessment before any water analysis is attempted [12]. Allosteric sites add another layer. Hidden allosteric pockets in GPCRs that only become apparent after ligand binding, and allosteric communication between protomers in oligomers, represent non-classical sites that a single apo-state prediction cannot reveal [13]. A distinct class of allosteric sites sits at the receptor-lipid bilayer interface, where lipid modulators bind, adding a membrane context that soluble-protein predictions do not capture [14]. Low local confidence near a predicted interface or at a membrane-facing surface is sometimes the model honestly telling you that the relevant conformational state is absent. Sequence-level AI can also nominate binding regions before any structure is inspected. A deep learning model predicting binding regions from patterns of protein and ligand interactions was used to nominate a novel putative binding site on the P2X3 receptor, which was converted into pharmacophore features and used for virtual screening, yielding compounds with low micromolar IC50 values and a hit rate reported as 10-fold higher than conventional approaches [15]. Interaction fingerprint analysis across hundreds of RNA-dependent RNA polymerase structures similarly distilled binding-site characteristics into major subpockets that guided screening of a drug dataset of 7894 compounds [16]. ## Common Mistakes and Limitations ![Three-question triage routing a pocket to docking, residue scrutiny or MD refinement based on mean pLDDT, low-confidence side chains and PAE.](/article-images/illustrations/alphafold-confidence-drug-discovery-binding-site-pocket-confidence-triage.webp) **Mistake 1, reading the global mean pLDDT as a binding-site verdict.** A model can average 85 while the pocket region averages 55. Always compute the pocket-restricted summary from step 5. **Mistake 2, assuming accurate pockets mean accurate poses.** The benchmark on GPCRs found that docking into AF2 models produced binding poses no better than docking into traditional homology models, and clearly worse than docking into experimental structures [5]. High pLDDT at the pocket licenses pocket-level hypotheses, not pose-level conclusions. **Mistake 3, docking into the apo snapshot and calling it induced fit.** AlphaFold models have no ligand bound, no induced-fit adjustment, and no bound cofactors, ions, or structured waters. Docking ignores the fact that binding sites move. Treat the AF model as one member of a conformational ensemble and use MD or ensemble docking to expand it [7]. **Mistake 4, trusting pocket residues in the 50 to 70 pLDDT band.** Side chains in that band are explicitly flagged as less reliable by the database color convention. If your ligand makes a specific contact with such a residue, that contact is speculative until validated. **Mistake 5, ignoring PAE across domains and chains.** pLDDT is local. A pocket at a domain interface, at an oligomer interface, or at a receptor and partner interface is governed by PAE and ipTM. Many allosteric and PPI-targeting projects die from reading only pLDDT. **Mistake 6, transferring thresholds between tools.** AlphaFold 3, Chai-1, and other recent predictors emit related but not identical scores, and their relative accuracy on binding modes differs by system, as the GPCR and peptide benchmark demonstrated [4]. Check the documentation of the exact tool and version you ran. **Mistake 7, docking to a model with a cofactor-bound state.** Many enzymes, especially kinases and phosphatases, adopt distinct conformations depending on bound cofactors or activation state. A predicted model of the unbound protein may not resemble the drug-relevant state. Where a ligand-bound experimental structure exists for the state you care about, prefer it or at minimum validate against it. **Mistake 8, treating disordered regions as structure.** pLDDT below 50 usually means the region is intrinsically disordered, not that AlphaFold failed to see a structure that exists. There may be nothing to model. **The honest overall limitation.** Every published benchmark converges on the same picture. AlphaFold models are excellent maps of binding-site geometry and poor guarantees of binding-pose accuracy, and they capture one state of a dynamic ensemble [5]. Use them to prioritize and to generate hypotheses, and use experimental structures, molecular dynamics, and biochemical assays to confirm. One final scope note. This article is a computational workflow guide. Any individual research or clinical decision that follows from it needs domain expertise and experimental validation specific to that case. ## Frequently Asked Questions ### What pLDDT cutoff should I demand for binding-site residues? There is no official cutoff, and the most defensible practice is to report the distribution rather than pick a single threshold. Use the database color bands as a vocabulary: above 90 is very high, 70 to 90 is confident for backbone with less certain side chains, 50 to 70 is low, and below 50 usually means disorder. In practice, pockets averaging above 90 are strong candidates for docking, pockets in the 70s need residue-level scrutiny, and pockets below 70 on average should push you toward MD refinement or experimental structures. ### Can I dock small molecules directly into an AlphaFold model? Yes, and it is a legitimate first screen, especially for targets with no experimental structures, as the CDK20 hit discovery work demonstrated. Just dock with the knowledge that predicted pose accuracy is the weakest link, and validate any pose that matters against an experimental structure or simulation. ### What is the difference between pLDDT and PAE? pLDDT is a single number per residue describing local accuracy, how well that residue and its neighbors are placed. PAE is a matrix describing how confidently two residues are positioned relative to each other, which governs whether domains and chains are oriented correctly. Pocket geometry reads pLDDT, pocket context reads PAE. ### Does high confidence mean a pocket is druggable? No. Confidence scores measure how accurately the structure is predicted, not whether a pocket can bind a drug-like molecule with useful affinity. A perfectly predicted shallow, solvent-exposed groove is accurately not druggable. Druggability requires separate geometric and physicochemical analysis, plus water thermodynamics for serious campaigns. ### Should I use the AlphaFold server or the AlphaFold database models? Use the database when your protein alone is already modeled, because the models are versioned, instantly available, and reproducible. Use the AlphaFold server when you need a complex with a ligand, peptide, nucleic acid, ion, or modification, or when your protein is absent from the database. For documented screening campaigns, the versioned database models are easier to cite. ### Why do my docking results against AlphaFold models look worse than I hoped? Because the benchmark evidence says exactly this: pocket geometry in AF2 models is far better than homology models, yet ligand pose accuracy from docking was not better than with homology models and was much lower than with experimental structures. The model is good, the pose problem is real, and the fix is validation against experiment and ensemble or dynamics-aware approaches, not a longer exhaustiveness setting. ### How do I handle a low-confidence loop sitting inside an otherwise good pocket? Three options, in increasing order of effort. Exclude the loop residues from the docking box so the ligand is not scored against unreliable atoms. Refine the loop with a loop-modeling or MD protocol and dock against the refined ensemble. Or treat the loop region as a design variable and wait for experimental structures. Never let the docking engine treat an unreliable loop as fact. ### Can AlphaFold reveal cryptic or allosteric binding sites? Not by itself, since it predicts a single static state that usually resembles the apo form. But AlphaFold models are excellent starting points for the simulations that do reveal cryptic pockets, as shown for PPM1D where simulations started from the predicted structure exposed the cryptic flap and hinge pocket that explains its allosteric inhibitors. Run dynamics from the model, look for pockets that open, and return to the confidence scores to judge which parts of the starting model you could trust in the first place. ## Sources 1. [AlphaFold-driven structure-based drug discovery: Capabilities, benchmarks and limitations for medicinal chemistry.](https://pubmed.ncbi.nlm.nih.gov/42906387/) 2. [Unveiling Dynamic Hotspots in Protein-Ligand Binding: Accelerating Target and Drug Discovery Approaches](https://pubmed.ncbi.nlm.nih.gov/40362212/) 3. [Binding site comparisons for target-centered drug discovery](https://pubmed.ncbi.nlm.nih.gov/30856011/) 4. [Deep learning in GPCR drug discovery: benchmarking the path to accurate peptide binding](https://pubmed.ncbi.nlm.nih.gov/40285358/) 5. [How accurately can one predict drug binding modes using AlphaFold models?](https://pubmed.ncbi.nlm.nih.gov/38131311/) 6. [AlphaFold accelerates artificial intelligence powered drug discovery: efficient discovery of a novel CDK20 small molecule inhibitor](https://pubmed.ncbi.nlm.nih.gov/36794205/) 7. [Discovery of a Hidden Trypanosoma cruzi Spermidine Synthase Binding Site and Inhibitors through In Silico, In Vitro, and X-ray Crystallography](https://pubmed.ncbi.nlm.nih.gov/37521650/) 8. [Recent computational advances in the identification of cryptic binding sites for drug discovery](https://pubmed.ncbi.nlm.nih.gov/40799497/) 9. [Discovery of a cryptic pocket in the AI-predicted structure of PPM1D phosphatase explains the binding site and potency of its allosteric inhibitors](https://pubmed.ncbi.nlm.nih.gov/36993233/) 10. [Dual modality feature fused neural network integrating binding site information for drug target affinity prediction](https://pubmed.ncbi.nlm.nih.gov/39875637/) 11. [Deep Learning-Based Modeling of Drug-Target Interaction Prediction Incorporating Binding Site Information of Proteins](https://pubmed.ncbi.nlm.nih.gov/36967455/) 12. [Calculating Water Thermodynamics in the Binding Site of Proteins - Applications of WaterMap to Drug Discovery.](https://pubmed.ncbi.nlm.nih.gov/28413953/) 13. [Unmasking allosteric-binding sites: novel targets for GPCR drug discovery](https://pubmed.ncbi.nlm.nih.gov/35649692/) 14. [Allosteric binding sites at the receptor-lipid bilayer interface: novel targets for GPCR drug discovery.](https://pubmed.ncbi.nlm.nih.gov/33301977/) 15. [AI-based prediction of new binding site and virtual screening for the discovery of novel P2X3 receptor antagonists.](https://pubmed.ncbi.nlm.nih.gov/35849939/) 16. [Structural Insights into the Binding Modes of Viral RNA-Dependent RNA Polymerases Using a Function-Site Interaction Fingerprint Method for RNA Virus Drug Discovery](https://pubmed.ncbi.nlm.nih.gov/32946692/)