# Interpreting AlphaFold-Multimer Confidence Scores: pLDDT and PAE AlphaFold-Multimer reports two headline confidence metrics for every predicted protein complex: pLDDT, a per-residue score from 0 to 100 that estimates how well each residue's local structure is predicted, and PAE (predicted aligned error), a residue-pair matrix in Ångströms that estimates how much uncertainty remains about the relative position of two residues once the model is aligned. Interpreting these two outputs correctly, together with the interface scores ipTM and pTM, is the difference between a prediction you can publish on and a model that merely looks convincing in a molecular viewer. This matters because AlphaFold-Multimer (AFM) has become the default first step for anyone asking whether two proteins form a complex and what that complex looks like. The method transformed the field. Benchmarks of protein-peptide docking report a median DockQ of 0.47 across 112 complexes, a leap over classical docking approaches, and AFM also ranked best among tested methods for separating interacting from non-interacting peptide-protein pairs, with a ROC-AUC of 0.78 [1]. Yet the same literature is full of warnings that AFM succeeds on only a fraction of targets, with reported accurate prediction of protein complexes in up to 43% of benchmark cases, and far less for antibody-antigen pairs [2]. The raw coordinates tell you almost nothing about which category your prediction falls into. Only the confidence scores do that, and only if you read them correctly. ## The Two Scores at a Glance ![Comparison of pLDDT, PAE and ipTM by scale, what each measures, and the blind spot each one leaves.](/article-images/illustrations/alphafold-multimer-plddt-pae-interpretation-score-roles.webp) Before going deeper, here is how pLDDT and PAE relate to each other and to the scores most often confused with them. | Metric | Scale | Answers the question | Granularity | Common failure mode | |--|--|--|--|--| | pLDDT | 0 to 100 per residue | Is this residue's local backbone placed correctly? | Single residue | High pLDDT on both chains says nothing about whether the chains touch correctly | | PAE | 0 to ~32 Å per residue pair | If I align the true structure on residue A, how far off is residue B likely to be? | Residue pair | A mostly blue matrix can hide one bad interface block | | pTM | 0 to 1 per model | How confident is the model that the whole assembly folds correctly? | Whole complex | Includes within-chain packing, so it can look fine while the interface is wrong | | ipTM | 0 to 1 per model | How confident is the model about the interface specifically? | Interface only | Can be high for a plausible-looking but biologically wrong assembly | The single most common mistake in alphafold-multimer plddt pae interpretation is judging a complex by pLDDT alone. A two-chain model can carry an average pLDDT above 85 and still have its two subunits docked in an entirely wrong orientation, because pLDDT only evaluates local geometry around each residue. The cross-chain block of the PAE matrix and the ipTM score are what carry the interface information. ## pLDDT: What It Measures and How to Read It ### Definition and computation pLDDT stands for predicted local distance difference test. It is a per-residue confidence estimate that mimics the lDDT metric used in structure-assessment experiments: the model compares, for each residue, the distances from that residue's atoms to all other atoms within a fixed local neighborhood, against what the correct structure would give. A pLDDT of 90 means the network expects roughly 90% of those local distance checks around that residue to be preserved. The score is written into the B-factor column of the output PDB file, which is why most molecular viewers show AlphaFold models pre-colored from blue (high) through green and yellow to orange (low). Two properties of pLDDT follow directly from this definition and shape all practical interpretation: 1. pLDDT is local. It says the helix, sheet, or loop around that residue is internally well defined. It says nothing about where that helix sits relative to another chain, or even relative to a distant domain of the same chain. 2. pLDDT is calibration-sensitive. The score is generated by a network head trained to reproduce agreement statistics, so it estimates the accuracy the model expects, not a guarantee. Systematic overconfidence and underconfidence both occur, particularly for proteins far from the training distribution. ### Interpreting pLDDT bands The standard color bands used across AlphaFold DB and most viewers encode widely accepted interpretation thresholds: - **Above 90 (dark blue).** Very high local accuracy. Backbone geometry in this region is typically reliable enough for mechanistic reasoning, and side-chain placement is often usable. - **70 to 90 (light blue to cyan).** Confident backbone. Individual residue positions may wobble, but secondary structure elements and domain cores are usually right. - **50 to 70 (yellow).** Low confidence. Treat the topology as a rough hypothesis. Loops and termini often fall here even in otherwise excellent models. - **Below 50 (orange).** The model frequently expects these regions to be disordered. AlphaFold documentation itself treats this band as a disorder signal, and many practitioners cut these segments before downstream modeling. One subtlety matters enormously for complex work: a low-pLDDT region can still be biologically meaningful. Intrinsically disordered segments fold upon binding, and short peptides are exactly such segments. Work on protein-peptide complex prediction shows that AFM and related models can predict peptide binding geometry well, but that performance depends heavily on the peptide [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) and on training-data bias toward previously seen structures [3], [4]. So a peptide drawn with a low pLDDT in isolation may still dock confidently when its receptor is present in the same prediction job, which is one reason full-complex prediction outperforms dock-into-fixed-monomer workflows for flexible partners. A second subtlety comes from docking studies. Because pLDDT reflects local structural definition, it can be repurposed as a proxy for rigidity. Harmalkar and colleagues exploited exactly this, using AlphaFold confidence measures to estimate protein flexibility and docking accuracy for multimers within their replica-exchange docking pipeline, AlphaRED [2]. High pLDDT regions are treated as rigid bodies while low pLDDT hinges and termini become the flexible parts that reorient during docking. That dual use, accuracy estimate and flexibility estimate, is worth keeping in mind when you look at a per-residue plot. ## PAE: The Matrix That Actually Describes a Complex ### Definition and computation PAE, predicted aligned error, answers a different question than pLDDT. Imagine you know the true structure. Align that true structure so that residue A sits perfectly on the model's residue A. How far, in Ångströms, is the model's residue B from where it should be? That expected displacement is the PAE value for the pair (A, B). The output is a square matrix, one row and one column per residue, so a 500-residue complex produces a 500 by 500 matrix. Typical values range from near 0 Å (the two residues' relative placement is essentially fixed) up to a saturating maximum of about 31.75 Å in the standard output (effectively "no idea"). PAE is not symmetric for a given pair, because the error depends on which residue you align on. In practice the matrix is rendered with a blue (low) to orange (high) colormap, and the standard output includes a pre-rendered PAE plot alongside the raw numbers in a JSON file. ### Reading the PAE matrix for a multimer For a two-chain prediction with chains of length N and M, the matrix has four blocks, and each block carries distinct information: 1. **Top-left block, chain A against chain A.** Low values here mean chain A's internal fold is pinned down. This overlaps with what pLDDT tells you. 2. **Bottom-right block, chain B against chain B.** The same statement for chain B. 3. **Top-right and bottom-left blocks, chain A against chain B.** These are the blocks that matter for complex prediction. Low (dark blue) values across these blocks mean the network is confident about where every residue of chain A sits relative to every residue of chain B, which is a statement about the interface and the relative orientation of the two subunits. The diagnostic pattern for a failed multimer prediction is easy to recognize once you have seen it: both diagonal blocks are dark blue while both cross-chain blocks are washed out in pale colors. The network folded each chain beautifully but is telling you, in Ångströms, that it has essentially no information about how they fit together. pLDDT will still look excellent, which is exactly why judging complexes on pLDDT alone fails. For higher-order assemblies the matrix becomes a mosaic of chain-by-chain blocks, and this block structure is precisely what allows large complexes to be tackled pairwise. The AFM-RL method for complexes of six to twenty chains starts from AFM-predicted pairwise models and then assembles them with a reinforcement learning framework, precisely because full-complex prediction degrades as chain count grows [5]. When you inspect such a multi-block matrix, you are asking the same question pairwise: is the cross-chain block for subunits i and j dark or pale? ### Quantitative reading of PAE values There is no single official threshold that converts a PAE value into "correct" or "wrong," because PAE is an expectation of displacement, not a binary verdict. Working practice, however, converges on rough qualitative bands: - Cross-chain PAE values of roughly 0 to 3 Å between interface residues indicate that the relative orientation is pinned down well enough for mechanistic discussion, provided the model itself is a plausible state (more on that caveat later). - Values in the 5 to 10 Å range signal ambiguity about the orientation or the register of the interface. The prediction may still be usable as a hypothesis, especially if multiple independent runs converge on the same arrangement. - Values saturating near the maximum mean the network is signaling ignorance. Do not interpret geometry from that block at all. One more practical point: PAE describes relative placement given an alignment, so it is insensitive to a whole-assembly rotation. Two models that differ only by a rigid-body rotation of chain B will show very similar PAE, which is why PAE should be read alongside actual structural clustering of your five predicted models rather than instead of it. ## ipTM and pTM: The Numbers That Rank Your Models AlphaFold-Multimer output includes two scalar confidence estimates that condense the whole picture into one number each, and for complex prediction they are arguably more useful than either pLDDT or PAE for the first pass of triage. **pTM**, predicted template modeling score, estimates the TM-score the model structure would achieve against the true assembly, computed over all residues of all chains. TM-score is a global fold-similarity metric that is length-normalized, so pTM summarizes confidence in the overall architecture of the assembly. **ipTM**, interface pTM, applies the same TM-score formula but restricted to the interface region between chains. It therefore isolates exactly the quantity a complex predictor is supposed to get right: the relative placement of subunits. In AFM output, models are ranked by a weighted combination of ipTM and pTM in which ipTM carries the dominant weight, so ranked_0.pdb is the model the network itself considers most confident at the interface. The widely used working heuristic, drawn from the original AFM evaluation, is that successful complex predictions cluster at ipTM above roughly 0.8, while most failures sit below roughly 0.6, with the region between treated as ambiguous. Recent work treats interface confidence with more nuance. A systematic screen of oxysterol-binding protein-related proteins against phosphoinositide phosphatases assessed interface confidence across roughly 200 protein pairs using the weighted ipTM plus pTM metric, the newer actifpTM and ipSAE scores, and FoldSeek-Multimer clustering, then filtered candidates by subcellular localization and interface plausibility before trusting them [6]. That layered approach, score first, biology second, is the current best practice, and it exists because scalar scores alone cannot separate a true interface from a convincing artifact. The distinction between ipTM and pTM also explains a confusing output pattern. A complex can have a respectable pTM driven by well-folded individual subunits while ipTM sits in the failure zone. When the two disagree, trust ipTM for interface questions and pTM for questions about the assembled fold. ## A Practical Walkthrough: Running Predictions and Finding the Numbers ![Decision path that checks average pLDDT, then ipTM, then the cross-chain PAE block, ending in a confident interface or a failed prediction.](/article-images/illustrations/alphafold-multimer-plddt-pae-interpretation-interface-triage.webp) This section works as a compact alphafold tutorial for the two most common routes, the local or notebook route and the official AlphaFold Server, and shows exactly where each score appears in the output. ### Running predictions locally with ColabFold ColabFold wraps AFM with sensible defaults and writes every confidence metric you need into its output directory. A minimal command for a two-chain prediction looks like this: ``` colabfold_batch --num-models 5 --model-type alphafold2_multimer_v3 dimer.fasta ./out ``` The FASTA file contains both chains separated by a header break, one header per chain. The output directory contains, among other files: - **ranked_0.pdb through ranked_4.pdb**, the five models ordered by ranking confidence, with per-residue pLDDT in the B-factor column. - **ranked_0_scores.json**, containing plddt, ptm, iptm, max_predicted_aligned_error, and the full PAE matrix as paes. - **Pre-rendered PAE plots as PNG images**, with chains labeled along both axes. For a quick triage, extract the scalars from the JSON: ``` python -c "import json; d=json.load(open('out/ranked_0_scores.json')); print(d['plddt'], d['ptm'], d['iptm'])" ``` Running multiple models is not optional decoration. Because AFM is stochastic, different seeds and model versions can land in different assemblies, and the spread across replicates is itself diagnostic. A screen that combined five-replicate AF2-Multimer predictions with a large interaction screen and then cross-checked with AlphaFold3 and additional scoring methods found that consistency across replicates and across tools was a key filter for candidate interactions [6]. If your five models agree, confidence goes up. If they split between two orientations, you have an ambiguous interface, not a confident one. ### Using the AlphaFold Server The AlphaFold Server, DeepMind's web interface, accepts protein sequences, [nucleic acids](/blog/guides/nucleic-acids), and common ligands as entities and returns predicted structures with an interactive PAE plot and downloadable confidence data. The interface colors the PAE matrix by chain, so the cross-chain blocks are visually obvious, and it reports the scalar confidence measures alongside. For a first pass on a new interaction, the server is the fastest route, and many published screens use it as the entry point before local replication. Interkingdom interaction screens, for example pathogen-secreted inhibitors of host immune hydrolases, describe exactly this pattern of server-based screening followed by curation and control-based filtering, and note that platform constraints, sequence depth, and compute time shape what such screens can deliver [7]. One limitation worth planning around: the server does not let you tune seeds or model versions the way local runs do, so reproducibility across large screens requires care in how predictions are queued and recorded. ### A worked interpretation Suppose you predict a 180-residue protein against a 120-residue partner. The output gives an average pLDDT of 88, ipTM of 0.86, pTM of 0.79, and a PAE matrix whose cross-chain block shows values of 1 to 4 Å between residues 40 to 90 of chain A and residues 10 to 60 of chain B, with values above 15 Å elsewhere in the cross-chain block. Reading in order: 1. pLDDT of 88 says both chains fold well internally. It says nothing yet about the interface. 2. ipTM of 0.86 falls above the confident band. The network expects the interface geometry itself to be roughly right. 3. The cross-chain PAE block confirms the localization. A tight dark patch spanning a specific residue range means the confidence is concentrated at a defined interface, not smeared across a vague overall association. 4. The pale remainder of the cross-chain block means the distal parts of the two chains dangle with uncertain relative position, which is normal and not a red flag as long as the interface core is dark. Contrast this with the failure pattern: pLDDT of 90, ipTM of 0.45, cross-chain PAE uniformly above 15 Å. Each chain is well folded, the interface is guesswork, and the model should be reported as a prediction of two folds, not a prediction of an interaction. ## How Confidence Scores Are Used in Real Research Pipelines Confidence metrics are not just report decorations. They drive decisions at every stage of modern complex modeling pipelines, and the published literature shows three recurring uses. **Triage and ranking.** Every AFM run produces five ranked models, and every downstream tool consumes or recomputes those rankings. A comprehensive evaluation of AFM, AlphaFold3, and ColabFold for protein-peptide complexes compared template-based and template-free prediction pools and assessed a battery of scoring functions for re-ranking; the built-in AlphaFold scoring function performed best among those tested, while FoldX-Stability and HADDOCK-type contact scores added complementary value, and the authors suggested consensus scoring across many prediction models as the most robust strategy [8]. In other words, the built-in confidence scores are the best single ranking signal available, but combining them with physics-based scores squeezes out further accuracy. **Rescue of hard targets.** When AFM's own scores reveal a weak prediction, hybrid pipelines take over. AlphaRED treats AFM output as a starting point, uses pLDDT as a flexibility guide, and refines with physics-based replica-exchange docking, recovering acceptable-quality CAPRI predictions for 63% of benchmark targets that include AFM failure cases, and raising antibody-antigen success from roughly 20% under AFM alone to 43% [2]. The logic only works because pLDDT is informative about where the model is uncertain, which is a direct payoff of correct confidence interpretation. Similar integration appears in CAPRI-casino practice, where a team progressively integrated AlphaFold into its HDOCK-based pipeline and refined its sampling strategy across rounds 47 to 55 to improve prediction accuracy [9], and in restrained docking frameworks such as ColabDock, which folds experimental restraints into the AlphaFold prediction objective and outperformed HADDOCK and ClusPro on restrained benchmarks [10], [11]. Sparse experimental data such as covalent labeling mass spectrometry measurements can similarly distinguish nativelike from non-nativelike complex models when combined with AlphaFold-derived subunits in Rosetta [12]. **Interaction discovery at scale.** High-level screens use confidence scores as a coarse filter before any structural scrutiny. A study of predation-essential hypothetical proteins in Bdellovibrio combined protein network clustering with AFM and reported high-confidence protein-protein interactions and two novel complexes among candidate attack-phase proteins, with confidence scoring as the primary gate [13]. The caveat from that literature is consistent: screen-level confidence is a hypothesis generator, and each candidate needs separate validation, ideally experimental. Work on interkingdom interaction discovery emphasizes control sequences, careful sequence curation, and reuse of sequence alignments as the practical levers that keep such screens honest [7]. A final, less obvious use: confidence scores tell you when alternative conformations are being suppressed. A study of the mitochondrial ADP/ATP carrier found that both AF2 and AF3 bias predictions toward the cytoplasmic-facing state, while a ColabFold implementation generated a credible model of the matrix-facing state, with the authors pursuing the alternative conformation through modeling and molecular dynamics rather than accepting the highest-confidence output as the only state [14]. A confident model is a confident model of one state. Low-confidence signals elsewhere in the PAE matrix can be the first hint that a protein samples more than one arrangement. ## What Confidence Scores Cannot Tell You Several hard limits follow from how these scores are computed, and each has tripped up published analyses at least once. **High confidence does not certify a biological interaction.** AFM was trained on known complexes, and like all deep-learning structure predictors it shows bias toward interface geometries seen during training, with documented struggles on novel target proteins and binding sites [3], [4]. A dark PAE block can therefore reflect a memorized geometry applied to an unrelated pair. Cross-validation with evolutionary plausibility, localization, and experimental evidence remains mandatory. The PeSTo interface-prediction tool and ConSurf conservation analysis are among the complementary checks used in recent integrative protocols [6]. **The scores describe the model relative to a hypothetical truth, not the truth itself.** PAE tells you the model's self-estimated uncertainty. If the network's training distribution never included your interaction class, that self-estimate can be miscalibrated in both directions. **MSA depth modulates everything.** For peptide partners and orphan proteins, the paired and unpaired multiple sequence alignments are shallow by nature. Analysis of how AFM and related models use these alignments found weak evidence for use of paired coevolutionary signal but confirmed that both the protein and peptide unpaired MSAs contribute to accuracy [4]. Practically, a confident prediction with deep MSAs on both sides carries more weight than an equally confident prediction built on a single sequence and a shallow profile. **Stochastic variation is real.** Running the same job twice changes the numbers. Protocols that rely on single runs, or that compare scores across different numbers of replicates, are comparing noise as much as signal. Five replicates is the de facto standard in careful pipelines. ## Common Mistakes and Limitations - **Ranking a complex by average pLDDT.** This is the defining error of new users. pLDDT evaluates local geometry and is blind to inter-chain orientation. A complex with pLDDT of 90 and ipTM of 0.4 is a failed prediction wearing a healthy-looking average. - **Reading only the pre-rendered PAE image without the chain labels.** The matrix only carries meaning when you know which blocks correspond to which chain pairs. For a three-chain prediction, a dark block between chains 1 and 2 and a pale block between chains 2 and 3 describes a complex with one confident interface and one guess. - **Treating a single model as the answer.** Five stochastic predictions are the minimum useful sample. Agreement across models and seeds raises confidence, disagreement means the interface is ambiguous or state-dependent. - **Ignoring the fold-state caveat.** A high-confidence model of a conformation that is not the biologically relevant state is still a wrong answer for your question, as the ADP/ATP carrier study demonstrated for transporter conformations [14]. - **Applying monomer intuition to antibody-antigen systems.** AFM is weakest exactly where many users most want it, on antibody-antigen interfaces, with reported success around 20% under AFM alone [2]. Low ipTM on an antibody complex is expected behavior, not proof the interaction is absent. Hybrid pipelines exist precisely for this regime,. - **Comparing scores across different tools as if they were interchangeable.** AlphaFold3, Boltz-1, and Chai-1 report related but not identical confidence schemes, and their calibration differs. Cross-tool agreement at the level of interface geometry is meaningful, cross-tool score comparison is not,. - **Using pLDDT below 50 as evidence of disorder without checking context.** The orange band usually signals disorder, but for peptides that fold on binding, low isolated-confidence regions can become confident once the partner is included in the same job [1]. - **Forgetting that ipTM is a model-internal estimate.** It correlates with interface accuracy on benchmarks, but an ipTM of 0.9 on a novel-type interface, far from training data, deserves more skepticism than the same number on a familiar fold-family interaction [3]. ## Quick Review 1. pLDDT is a per-residue local accuracy score from 0 to 100, written into the PDB B-factor column, with widely used interpretation bands at 90, 70, and 50. 2. PAE is a residue-pair matrix in Ångströms that estimates relative-position uncertainty, and only its cross-chain blocks speak to the interface. 3. The classic failed-complex signature is dark diagonal blocks and pale cross-chain blocks, paired with high average pLDDT and low ipTM. 4. ipTM is the interface-specific scalar and the dominant term in model ranking, with successful complex predictions typically clustering above roughly 0.8 and failures below roughly 0.6. 5. Newer interface scores such as ipSAE and actifpTM, combined with replicate consistency and tool cross-checks, form the current best practice for interface confidence [6]. 6. AFM succeeds on a minority of hard targets, up to about 43% of general benchmark cases and around 20% for antibody-antigen pairs, so low scores are common and honest, not always a user error [2]. 7. Scores are self-estimates subject to training bias, MSA dependence, and conformational state bias, so every confident prediction needs biological and, where possible, experimental corroboration [3], [14]. ## Frequently Asked Questions ### What pLDDT value is high enough to trust a model? There is no single cutoff, but the standard color bands give working guidance. Residues above 90 carry very high local confidence, 70 to 90 usually means the backbone is right even if side chains wobble, 50 to 70 should be treated as low confidence topology, and below 50 often marks disorder. For a complex, average pLDDT is the least informative of all the available numbers, so never gate a prediction on it alone. ### What does a high ipTM mean for my predicted complex? ipTM estimates how accurate the interface between chains is likely to be, on a 0 to 1 scale, and it is the number AFM uses to rank models. Values above roughly 0.8 correspond to the confident band in benchmark evaluations, values below roughly 0.6 correspond mostly to failed interface predictions, and the interval between is ambiguous. A high ipTM means the network is confident about the relative subunit arrangement, not that the interaction exists in the cell. ### Why is my cross-chain PAE high when both chains have excellent pLDDT? Because the two metrics measure different things. pLDDT evaluates local geometry within each residue's neighborhood, so each chain can fold beautifully while the network has no information about their relative orientation. Pale cross-chain blocks in the PAE matrix with dark diagonal blocks are the signature of this failure, and the correct response is to look at ipTM, run more replicates, and consider hybrid docking rather than accepting the coordinates. ### Can I compare pLDDT and PAE numbers between AlphaFold 2, AlphaFold 3, and other tools? Treat cross-tool score comparison as unreliable. The metrics share names and scales but are produced by differently trained networks with different calibration, and benchmark comparisons show the tools differ in what they get right even when scores look similar. The robust cross-tool check is agreement of interface geometry, for example through structural clustering of models from each tool, rather than agreement of the numbers themselves. ### Where do I find the PAE values in my output files? In ColabFold output, the full PAE matrix sits in each ranked model's scores JSON file under the key for PAE values, alongside scalar entries for plddt, ptm, iptm, and max_predicted_aligned_error, and pre-rendered PAE plots ship as PNG images. On the AlphaFold Server, the interactive PAE viewer and downloadable files carry the same information. In PDB files, only pLDDT is embedded, in the B-factor column. ### Does a confident AlphaFold-Multimer model prove the interaction exists in the cell? No. A confident model is a well-supported structural hypothesis, not evidence of a biological interaction. Training bias toward known interface geometries, shallow sequence alignments, and conformational-state effects can all produce confident models of assemblies that never form in vivo. Published discovery pipelines treat high-confidence predictions as candidate hypotheses and require orthogonal evidence, such as localization, conservation, interface plausibility analysis, or experiments, before claiming an interaction. ## Sources 1. [Improving peptide-protein docking with AlphaFold-Multimer using forced sampling](https://pubmed.ncbi.nlm.nih.gov/36304330/) 2. [Reliable protein-protein docking with AlphaFold, Rosetta, and replica-exchange](https://pubmed.ncbi.nlm.nih.gov/37546760/) 3. [How AlphaFold and related models predict protein-peptide complex structures](https://doi.org/10.1101/2025.06.18.660495) 4. [Training bias and sequence alignments shape protein-peptide docking by AlphaFold and related methods](https://pubmed.ncbi.nlm.nih.gov/41084281/) 5. [AFM-RL: Large Protein Complex Docking Using AlphaFold-Multimer and Reinforcement Learning](https://doi.org/10.1101/2024.01.20.576386) 6. [AlphaFold-driven discovery of oxysterol-binding protein-related protein-phosphoinositide 3-, 4-, and 5-phosphatase interactions using new generation confidence scores](https://pubmed.ncbi.nlm.nih.gov/41983709/) 7. [Using AlphaFold Multimer to discover interkingdom protein-protein interactions](https://pubmed.ncbi.nlm.nih.gov/39152709/) 8. [Comprehensive Evaluation of AlphaFold-Multimer, AlphaFold3 and ColabFold, and Scoring Functions in Predicting Protein-Peptide Complex Structures](https://doi.org/10.1101/2024.11.11.622992) 9. [Protein Complex Structure Prediction With AlphaFold-Enhanced HDOCK in CAPRI Rounds 47-55.](https://pubmed.ncbi.nlm.nih.gov/40167205/) 10. [Integrated structure prediction of protein-protein docking with experimental restraints using ColabDock](https://doi.org/10.1038/s42256-024-00873-z) 11. [ColabDock: inverting AlphaFold structure prediction model for protein-protein docking with experimental restraints](https://doi.org/10.1101/2023.07.04.547599) 12. [Protein complex prediction using Rosetta, AlphaFold, and mass spectrometry covalent labeling](https://pubmed.ncbi.nlm.nih.gov/36543826/) 13. [Using AlphaFold-Multimer to study novel protein-protein interactions of predation essential hypothetical proteins in Bdellovibrio](https://pubmed.ncbi.nlm.nih.gov/40297267/) 14. [Modeling the different conformations of the human mitochondrial ADP/ATP carrier using AlphaFold and molecular dynamics simulations of the protein-ligand complexes](https://pubmed.ncbi.nlm.nih.gov/40225838/)