# AlphaFold-Multimer vs. AlphaFold 3 for Complex Prediction AlphaFold-Multimer predicts structures for complexes made of protein chains only, while AlphaFold 3 predicts complexes that can also include [nucleic acids](/blog/guides/nucleic-acids), ligands, ions, and chemical modifications, and it replaces the earlier attention-based structure module with a diffusion-based one. The practical rule of thumb is simple. If your complex is proteins only and you need local control, reproducibility, bulk sampling, or integration with docking and scoring pipelines, AlphaFold-Multimer remains the workhorse. If your complex involves RNA, DNA, small molecules, metal ions, or modified residues, or if you want the single best-shot prediction without building a pipeline, AlphaFold 3 is the better starting point. On head-to-head benchmarks of protein-protein docking, AlphaFold 3 generally posts higher success rates than AlphaFold-Multimer, but neither tool is close to solved for antibody-antigen targets, large assemblies, or complexes with large conformational changes, which is exactly where hybrid pipelines built around AlphaFold-Multimer still earn their keep [1](#ref-4)[2](#ref-15). This article breaks down the difference between AlphaFold-Multimer and AlphaFold 3 for protein complexes, compares them on the criteria that actually decide a project, and gives you worked scenarios, command-level practical guidance, and the common mistakes that waste compute and mislead interpretations. ## What Each Tool Actually Is ### AlphaFold-Multimer AlphaFold-Multimer (often abbreviated AF-Multimer or AFm) is an extension of AlphaFold 2 that was retrained to handle multiple protein chains as a single folded unit. It takes, for each chain, a [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) (MSA, a table of evolutionarily related sequences that reveals which positions co-vary across partners) and optional structural templates, then produces 3D coordinates for every chain in the assembly together with per-residue and pairwise confidence metrics. Because it inherits the AlphaFold 2 codebase, it runs on your own hardware, accepts full control over MSA depth, number of recycles, and random seeds, and produces as many models as you are willing to run. A typical local setup runs five model presets with five seeds each for 25 predictions per complex, which matters because complex prediction is far more stochastic than monomer prediction and ranking across models is half the job. The confidence metrics you will use constantly are ipTM (interface predicted TM-score, an estimate of interface accuracy that roughly separates reliable interfaces above about 0.8 from unreliable ones below about 0.6), pTM (predicted TM-score for the whole assembly), pLDDT (per-residue confidence), and PAE (predicted aligned error, a pairwise matrix whose cross-chain blocks show how well positioned each chain is relative to the others). These metrics are estimates, not measurements, and the antibody-antigen cases discussed later show why you should never trust them blindly [1](#ref-4)[3](#ref-18). ### AlphaFold 3 AlphaFold 3 is a redesign rather than an increment. Its architecture replaces the AlphaFold 2 Evoformer and structure module with a paired MSA module and a diffusion module, and its input representation generalizes across molecule types. You can feed it protein chains, RNA, DNA, small molecules, ions, and covalent modifications in one job, and it outputs coordinates plus confidence metrics tailored to the new architecture (pLDDT and PAE carry over conceptually, with an interface-level metric called pTM and a ligand-specific metric called PDE for pocket alignment). Two access modes exist. The AlphaFold Server is a free web interface hosted by Google DeepMind for non-commercial use, and the model code and weights are available to the research community under a license that permits local inference for non-commercial applications. The diffusion design choice has a direct practical consequence. Each prediction is a sampled denoising trajectory, so identical inputs with different seeds can produce visibly different poses, particularly for flexible regions and weak interfaces. This makes multi-seed sampling and ranking just as necessary for AlphaFold 3 as for AlphaFold-Multimer, which surprises researchers who expected the newer model to give a single deterministic answer. ## Side-by-Side Comparison | Criterion | AlphaFold-Multimer | AlphaFold 3 | |--|--|--| | Molecule types | Protein chains only | Proteins, DNA, RNA, ligands, ions, covalent modifications | | Architecture | AlphaFold 2 Evoformer plus IPA structure module, retrained for chains | Paired MSA module plus diffusion module | | Access | Open source, runs locally on your own GPUs | Free web AlphaFold Server for non-commercial use, local code under a research license | | Control over inputs | Full control over MSAs, templates, seeds, recycles, model presets | Server input is sequences plus optional custom MSA; local version gives more control | | Batch throughput | Unlimited locally, scriptable | Server has daily job limits and token limits per job | | Reproducibility | Fully reproducible with fixed seeds and pinned databases | Server outputs are not seed-controlled by the user, local runs are | | Protein-protein accuracy | Strong on rigid heterodimers and homomers, weaker on antibodies and large assemblies | Generally higher top-5 success rates on docking benchmarks [2](#ref-15) | | Antibody-antigen | Low success rate, around 20% in one benchmark [1](#ref-4) | Most accurate method in that benchmark family, but still far from reliable [2](#ref-15) | | Protein-RNA | Not applicable | Top-5 success rate of 87.0% on a 235-complex benchmark [4](#ref-1) | | Protein-lipid | Not applicable | 76.1% success at RMSD under 2 Å, dropping to 47.2% for targets not seen in training [5](#ref-8) | | Large assemblies | Handles moderate sizes, needs stepwise assembly for very large complexes [6](#ref-9) | Better per-chain accuracy but token limits constrain very large jobs [7](#ref-6) | | Confidence metrics | ipTM, pTM, pLDDT, PAE | pLDDT, PAE, pTM, PDE for ligands | | Output format | PDB or mmCIF per model, ranking JSON | mmCIF download per model from the server | | Best use case | Protein-only complexes where you need bulk sampling, scoring, and pipeline integration | Complexes containing nucleic acids, ligands, ions, or modifications, or when you want a strong single prediction fast | ## How They Differ on Real Prediction Tasks ### Protein-protein complexes For straightforward heterodimers and homomers with well-populated MSAs, both tools are strong and AlphaFold 3 is usually somewhat better. A systematic evaluation of 11 docking methods on DockingBenchmark 5.5, the standard curated set of protein-protein complexes with bound and unbound reference structures, found that AlphaFold 3 achieved a top-5 success rate of 77.98% when docking against apo (unbound) structures, while the best classical method, HDOCK, reached only 12.84% in that setting even though it hit 85.24% when given bound (holo) structures [2](#ref-15). That gap is the whole story of the field in one number. End-to-end co-folding methods like AlphaFold-Multimer and AlphaFold 3 read the sequences and fold the complex directly, so they do not suffer the way classical docking does when the unbound starting structures have drifted apart from the bound conformation. AlphaFold-Multimer itself, evaluated in its original publication window, succeeded on up to 43% ofDocking Benchmark 5.5 targets depending on the strictness of the criteria, and that figure is still quoted as the baseline that hybrid methods try to beat [1](#ref-4). The gap between AlphaFold-Multimer and AlphaFold 3 on protein-protein targets is real but modest for easy cases, and it widens for hard cases. A recent pre-training strategy called split and merge proxy (SMP), which generates pseudo-dimers by splitting monomeric proteins and training models to merge them back together, produced predictions more accurate than both AlphaFold-Multimer and AlphaFold 3 on several CASP15 dimer targets, which tells you two things. First, the ceiling on dimer accuracy is not architectural destiny, it is partly a data problem. Second, both production models still leave recoverable accuracy on the table for hard dimers [8](#ref-2). ### Antibody-antigen complexes This is where the choice of tool matters least in absolute terms because both struggle. Antibodies pair with antigens through CDR loops that are generated by V(D)J recombination, so antibody chains lack the deep co-evolutionary signal that powers co-folding on naturally evolving complexes. In the ReplicaDock 2.0 study, AlphaFold-Multimer succeeded on only about 20% of antigen-antibody targets, and the AlphaRED protocol, which uses AlphaFold to generate starting structures and then runs physics-based replica exchange docking in Rosetta, raised that to 43% on the same subset [1](#ref-4). In the broader 11-method benchmark, AlphaFold 3 was the most accurate method for antibody-antigen docking, but the absolute success rates remained low relative to general protein-protein docking [2](#ref-15). CAPRI, the community-wide blind prediction experiment, tells the same story from the inside. Teams that combined AlphaFold-based model generation with classical docking, scoring, and refinement consistently outperformed either approach alone, and several groups noted that AlphaFold could produce near-native antibody-antigen models that their confidence metrics then misranked [9](#ref-5)[3](#ref-18). The practical takeaway for antibody work is that you should treat both tools as model generators, not as answer machines. Generate many models with multiple seeds, rescore with an independent energy or machine-learning-based function, and, if you have any experimental restraints from mutagenesis or cross-linking, use them to filter. ### Protein-RNA and protein-nucleic acid complexes AlphaFold-Multimer cannot model RNA at all, so this comparison is one-sided. A benchmark of 235 protein-RNA complexes curated from PDB structures deposited up to July 2024 compared five template-free docking web servers against the AlphaFold Server. The best docking server, HDOCK, achieved success rates of 31.1% and 44.7% within the top 1 and top 5 predictions by CAPRI metrics. AlphaFold 3 outperformed all docking servers with an overall top-5 success rate of 87.0%, though it failed in nine cases where at least one docking server succeeded, which is a useful reminder that the co-folding approach has failure modes docking does not [4](#ref-1). For any protein-RNA question, including RNA-binding protein specificity and ribonucleoprotein assembly, AlphaFold 3 through the AlphaFold Server or local installation is the default choice, with docking servers as a complement rather than a baseline. ### Protein-ligand and protein-lipid complexes AlphaFold-Multimer cannot model lipids or small molecules, so again only AlphaFold 3 competes. The LiPP benchmark, a curated set of 311 nonredundant nonannular lipid-protein complex structures, evaluated three AI co-folding tools (AlphaFold 3, Chai-1, RoseTTAFold AA), one AI docking tool (DiffDock-L), and one physics-based docking tool (AutoDock Vina). AlphaFold 3 had the highest success rate at 76.1% under an RMSD threshold of 2 Å, but the rate dropped to 47.2% for structures not seen during training [5](#ref-8). That 29-point drop is the clearest published illustration of a general AlphaFold 3 property you must budget for. The model memorizes training-set chemistry remarkably well and generalizes less well to unfamiliar complexes. When you predict a complex involving an unusual ligand, check whether similar ligands appear in the PDB and treat the confidence metrics with extra suspicion if it does not. ### Large assemblies Both tools degrade as assembly size grows, for different reasons. AlphaFold-Multimer has no hard token limit in principle but runs out of practical memory, and very large complexes need stepwise assembly. A study modeling five complete cytokine-receptor-JAK2 signaling complexes, ranging from 3,220 to 4,074 residues, had to build the models piece by piece from smaller fragments, select and validate each fragment against published experimental data, and then assemble them, because a single-shot prediction of the full complex was not tractable [6](#ref-9). AlphaFold 3 constrains input size through a token budget, so very large assemblies hit the wall in a different place. The most interesting recent result on large complexes is HDOCK-Multimer, a framework that integrates docking with combinatorial assembly. On three benchmarks covering 35 large heteromeric complexes, 172 large protein complexes, and 7 CASP15 targets, HDOCK-Multimer substantially outperformed state-of-the-art methods including CombFold, MoLPC, AlphaFold-Multimer, and AlphaFold 3, and it can even infer stoichiometry when the user does not provide it [7](#ref-6). The lesson is that end-to-end co-folding is not the only path to big assemblies, and for complexes of more than roughly ten chains you should seriously consider assembly-based pipelines that use pairwise interaction predictions as input rather than trying to fold everything at once. ### Weak and transient interactions Both AlphaFold-Multimer and AlphaFold 3 were trained predominantly on tight, well-ordered complexes from the PDB. Weak, transient, and highly flexible interactions are underrepresented, and the review literature identifies protein flexibility, reliance on co-evolutionary signals, and intrinsically disordered regions as the standing challenges for the entire AI-driven generation of methods [10](#ref-17). If your biological question involves a transient signaling encounter or a disordered segment that folds on binding, expect low-confidence output and plan experimental validation or restrained modeling rather than trusting a single prediction. ## When to Use AlphaFold-Multimer Knowing when to use AlphaFold-Multimer comes down to four situations where it beats or complements AlphaFold 3 in practice. **You need pipeline integration.** AlphaFold-Multimer is Python code you own. You can wrap it in a Snakemake or Nextflow workflow, feed it MSA variants generated by different tools, run it on hundreds of complex candidates overnight, and pipe its outputs directly into scoring functions, molecular dynamics setup, or docking refinement. The pyDock group's CAPRI work shows exactly this pattern, generating AlphaFold-Multimer models and then scoring them with a combination of pyDock energy and AlphaFold confidence, which improved discrimination between correct and incorrect interfaces [11](#ref-16). You cannot reproduce that workflow against a web server. **You need diverse sampling.** Because complex prediction is stochastic, running one model preset with one seed is a common and costly error. AlphaFold-Multimer lets you sweep model presets (model_1_multimer through model_5_multimer), seeds, MSA depths, and recycle counts. The MULTICOM system, which was blindly tested in CASP15, improved on standard AlphaFold-Multimer simply by sampling diverse MSAs and templates, generating many predictions, and ranking them with complementary metrics, raising the average TM-score of first submissions to about 0.76 from about 0.72 for standard AlphaFold-Multimer, with best-of-5 rising to about 0.80 [12](#ref-7). Nearly all of that gain came from sampling and ranking, not from touching the network itself. That headroom is only available when you control the run. **You want to validate predicted interactomes at scale.** Large interactome studies routinely use AlphaFold-Multimer as a triage step. The Arabidopsis TOR complex study used proximity labeling to map over a hundred candidate interactors and then applied an AlphaFold-Multimer approach to validate which of them form structurally plausible complexes with the core TORC proteins [13](#ref-12). That kind of high-throughput structural triage, running AlphaFold-Multimer over dozens or hundreds of candidate pairs and filtering on ipTM and PAE, is a natural fit for a locally scripted tool and an awkward fit for a rate-limited web server. **Your complex is protein-only and well-behaved.** For a rigid heterodimer of two annotated proteins with deep MSAs, AlphaFold-Multimer and AlphaFold 3 frequently agree, and the older tool gives you reproducibility and unlimited reruns for free. ## When to Use AlphaFold 3 **Any complex containing non-protein components.** RNA, DNA, ligands, ions, glycans, and modified residues are out of scope for AlphaFold-Multimer by construction. The protein-RNA benchmark result of 87.0% top-5 success makes AlphaFold 3 the default first attempt for ribonucleoprotein questions [4](#ref-1). **Protein-protein complexes where you want the strongest single shot.** On docking benchmarks, AlphaFold 3 posts the highest success rates among the evaluated methods, including against apo structures where classical docking collapses [2](#ref-15). If you have one complex, no pipeline, and no time to build one, running it through the AlphaFold Server is the fastest route to a defensible model. **You want a quick structural hypothesis before committing compute.** The AlphaFold Server returns results in minutes to hours with no local setup. For a ten-second sanity check on whether two proteins plausibly interact and through which domains, it is the right tool, and you can follow up locally with AlphaFold-Multimer sampling if the answer matters. ## Practical Workflows ### Using the AlphaFold Server The AlphaFold Server requires no installation. You create an account, enter sequences in FASTA format for each chain, optionally add entities such as ions or ligands by name or SMILES string, and submit. Job size is bounded by a token budget that counts every chain, residue, modification, and ligand, so a large multi-chain job with ligands may exceed the limit where the same complex without ligands would not. Results arrive as mmCIF files with per-model confidence data. Because the server does not let you set seeds or choose among many independent runs, treat each submission as one sample, and submit variants (for example, with and without a ligand, or with alternate stoichiometries) when you want to compare hypotheses. A short alphafold server walkthrough for a protein-RNA complex looks like this. Paste the [protein sequence](/blog/guides/protein-sequence) and the RNA sequence as two entities, submit, then inspect the returned PAE matrix. Look specifically at the cross-chain blocks. Dark blocks between the protein domain and the RNA segment indicate a well-determined interface, pale blocks indicate the relative orientation is a guess. Compare the PAE pattern against any biochemical footprinting or mutagenesis data you have before you invest in downstream modeling. ### Running AlphaFold-Multimer locally ![Worked calculation showing five model presets times five seeds gives 25 structures per complex, and a serious project needs 25 to 100 predictions.](/article-images/illustrations/alphafold-multimer-vs-alphafold-3-for-protein-complexes-sampling-budget.webp) The standard path is the open-source AlphaFold repository with the multimer parameter set, or a maintained containerized distribution. The core steps of a minimal alphafold tutorial-style run look like this. 1. Install prerequisites. You need a CUDA-capable GPU with at least 16 GB of memory for small complexes and substantially more for large ones, Docker or Apptainer with NVIDIA container support, and a large reference disk, typically on the order of 2 TB for the full genetic databases. 2. Download the genetic databases (UniRef90, UniRef300, MGnify, BFD, Uniclust30, PDB70 or the small molecule-free template sets as appropriate for your version) and the AlphaFold-Multimer parameter files. 3. Prepare a multi-chain FASTA where each chain header starts with a unique chain identifier. 4. Run the prediction. A typical invocation is: ```bash python run_alphafold.py \ --fasta_paths=my_complex.fasta \ --output_dir=/data/afm_output \ --db_preset=full_dbs \ --model_preset=multimer \ --num_multimer_predictions_per_model=5 \ --use_gpu_relax=true ``` This produces five predictions per model preset, 25 structures in total, each with a JSON file of pLDDT, PAE, pTM, and ipTM values. 5. Rank and inspect. Sort models by ipTM for interface quality, then look at the cross-chain PAE blocks to confirm the interface geometry is internally consistent rather than a high average masking a misoriented partner. For researchers who want a middle path, ColabFold provides a notebook and command-line interface that wraps AlphaFold-Multimer with its own MSA generation against public databases, runs on Colab GPUs or your own hardware, and handles the database setup for you. It is the fastest way to a legitimate multimer prediction without managing a 2 TB database mirror. ### The full sampling and ranking workflow For a complex that matters, a single run is never the deliverable. A defensible workflow looks like this. 1. Generate 25 to 100 models by sweeping model presets, seeds, and, where justified, MSA variants from different search depth or tools. 2. Rank with ipTM combined with at least one independent metric. EquiRank, a protein-protein interface quality estimator built on E(3)-equivariant graph neural networks and ESM-2 protein language model embeddings, outperformed existing interface quality methods on diverse datasets, including the top CASP15 method VoroIF_GNN and the self-assessment module of AlphaFold-Multimer repurposed for scoring [14](#ref-13). DeepUMQA-PA takes a different angle, using Voronoi tessellation-derived contact area and orientation features plus a fused graph neural network to estimate residue-wise model accuracy for complexes, and it outperformed the previous state of the art DeepUMQA3 on the CASP15 test set [15](#ref-3). Adding one of these independent scorers guards against the well-documented failure mode where AlphaFold confidence metrics rank a wrong model above a right one. 3. Cluster the top models by interface RMSD to see whether sampling converged on one interface or is split between alternatives. A split result is itself information, usually indicating a weak or ambiguous interface. 4. Optionally refine top models with physics-based methods. AlphaRED, the AlphaFold-initiated replica exchange docking protocol, rescued 63% of Docking Benchmark 5.5 targets that AlphaFold failed on, and the ClusPro team's CASP16 work showed that physics-based sampling combined with AlphaFold-based refinement produced high-accuracy models in cases where AlphaFold alone failed [1](#ref-4)[16](#ref-11). Energy-based rescoring with pyDock against AlphaFold-generated models follows the same logic and is documented as a practical protocol [17](#ref-10). ## Worked Scenarios ### Scenario 1: A signaling complex with a dimerized receptor, a ligand, and a kinase You want the structure of a class 1 cytokine receptor homodimer bound to its ligand and associated with JAK2, roughly 3,500 residues total. A single AlphaFold-Multimer job on the full assembly is not feasible. The published approach, which is the sensible template, is to model each subcomplex separately (extracellular domains plus ligand, transmembrane dimer, intracellular domain plus JAK2), validate each against experimental data such as cross-linking or functional mutagenesis, and assemble stepwise [6](#ref-9). AlphaFold 3 could model some of the smaller subcomplexes directly, but the assembly logic and the validation against experiment remain yours to do. This is a pipeline problem, and AlphaFold-Multimer's scriptability earns its place. ### Scenario 2: An antibody against a novel antigen You have a lead antibody sequence and an antigen with an AlphaFold-quality monomer model. Neither tool will hand you a reliable complex. Run AlphaFold 3 for its somewhat better raw success rate on this class [2](#ref-15), run AlphaFold-Multimer with aggressive sampling and MSA masking as a second opinion, then apply independent rescoring, for example with pyDock energy or a learned interface scorer [17](#ref-10)[14](#ref-13). If the two tools converge on the same epitope with high confidence, that is meaningful. If they disagree, plan an experimental mapping experiment before trusting either pose. In CAPRI practice, groups using exactly this hybrid strategy got acceptable models for antibody targets where AlphaFold alone had failed, and repeatedly emphasized that confidence metrics must be treated as one input among several during model selection [9](#ref-5)[3](#ref-18). ### Scenario 3: Triaging a protein-protein interactome Proximity labeling or co-IP gives you 80 candidate interactors for a bait protein. Model all 80 pairs with AlphaFold-Multimer in a batch, filter on ipTM and interface PAE, and take the top candidates for experimental follow-up. This is the TOR complex interactome workflow [13](#ref-12). AlphaFold 3 cannot be run this way through the server at that volume, and the local AlphaFold 3 license is restricted to non-commercial research, so for industrial triage or very high throughput, AlphaFold-Multimer (or a purpose-built docking method such as GeoDock, which reaches a 43% top-1 success rate on the DIPS test set without requiring MSAs) may be the only option that fits the throughput and licensing constraints [18](#ref-14). ## Common Mistakes and Limitations **Trusting a single prediction.** Complex prediction is stochastic. One model, one seed, no comparison, is a guess with a confidence score attached. Run at least five seeds and check whether the interface is consistent across models before you interpret anything. **Trusting ipTM as ground truth.** ipTM correlates with interface quality but fails predictably on antibody-antigen targets, where high-confidence wrong interfaces are common, and on complexes whose biology involves induced fit or disorder-to-order transitions [1](#ref-4)[3](#ref-18). The ClusPro and pyDock CAPRI teams both found that rescoring AlphaFold models with orthogonal physics- or energy-based functions recovered correct models that AlphaFold's own confidence had buried [16](#ref-16). **Forgetting the training-data boundary.** AlphaFold 3's performance on lipid-protein complexes dropped from 76.1% to 47.2% for structures not seen during training [5](#ref-8). Novel chemistry, unusual stoichiometries, and fold combinations absent from the PDB are exactly where both tools are weakest, and neither tool tells you when you have stepped outside its competence. You have to check. **Feeding bad stoichiometry.** Both tools predict the assembly you specify. If the biological unit is a trimer and you submit a dimer, you will get a confident-looking dimer. Check stoichiometry against size-exclusion chromatography, native mass spectrometry, or PDB biological assembly annotations before modeling. HDOCK-Multimer's ability to predict stoichiometry rather than assume it exists precisely because this input error is so common [7](#ref-6). **Assuming the tools replace docking.** The systematic docking evaluation showed HDOCK achieving an 85.24% top-5 success rate when given bound structures, a regime where the docking machinery itself is excellent and only the unbound-to-bound gap hurts it [2](#ref-15). When you have experimental structures of the individual partners, or when AlphaFold's confidence is low, classical docking plus scoring remains a legitimate and sometimes superior route. The fields are converging, with hybrid AlphaFold-plus-docking pipelines now standard practice in CAPRI [9](#ref-5)[16](#ref-11)[11](#ref-16). **Interpreting interface side chains literally.** Both tools frequently place interface side chains in approximate rotamers. The backbone topology and the identity of the contacting surfaces are usually the trustworthy part. Individual hydrogen bonds and salt bridges within a predicted interface should be treated as hypotheses for mutagenesis, not facts, a caution that applies equally to the detailed mechanistic residue-contact claims built on AlphaFold-Multimer models in the literature [19](#ref-20). **Confusing the server and the model.** The AlphaFold Server is a product with input limits, job limits, and a non-commercial term of service. AlphaFold 3 the model, run locally, is a different artifact with different constraints. Results from the two are not interchangeable in a methods section. State which you used. **Ignoring model quality estimation as a field.** A growing set of tools exists purely because AlphaFold's self-reported metrics are insufficient for complexes. DeepUMQA-PA [15](#ref-3), EquiRank [14](#ref-13), and the MULTICOM ranking metrics [12](#ref-7) all exist because researchers kept finding that ranking is where predictions fail. If the complex matters, use one. ## Frequently Asked Questions ### Is AlphaFold 3 always better than AlphaFold-Multimer for protein complexes? No. AlphaFold 3 posts higher success rates on most docking benchmarks, including a top-5 success rate of 77.98% against apo structures [2](#ref-15), but AlphaFold-Multimer remains competitive on straightforward protein-only complexes, and hybrid pipelines built around AlphaFold-Multimer with sampling, independent scoring, and refinement can outperform either tool alone [1](#ref-4)[12](#ref-7). For large assemblies, dedicated assembly frameworks have beaten both [7](#ref-6). ### Can I run AlphaFold 3 locally? Yes, within limits. DeepMind released the AlphaFold 3 model code and weights under a license that permits non-commercial use, so academic groups can run local inference with full control over seeds and inputs. Commercial use is not permitted under that license, and the free AlphaFold Server is likewise restricted to non-commercial work. If your project is commercial, AlphaFold-Multimer and other open alternatives are the practical route. ### When should I use AlphaFold-Multimer instead of AlphaFold 3? Use AlphaFold-Multimer when your complex is proteins only, when you need batch throughput or pipeline integration, when you need reproducible runs with controlled seeds, when you want to sweep MSAs and model presets for better sampling, or when licensing rules out AlphaFold 3. Use AlphaFold 3 whenever nucleic acids, ligands, ions, or modifications are part of the complex, since AlphaFold-Multimer cannot model them at all [4](#ref-1). ### Which confidence metrics should I check for a predicted complex? For AlphaFold-Multimer, check ipTM for interface quality, pTM for overall assembly accuracy, pLDDT for per-residue reliability, and the cross-chain blocks of the PAE matrix for whether each chain's position relative to the others is well determined. For AlphaFold 3, the analogous outputs are pLDDT, PAE, pTM, and PDE for ligand pockets. Treat all of them as estimates, and cross-check important models with an independent scorer [15](#ref-3)[14](#ref-13). ### How well do these tools handle antibody-antigen complexes? Poorly, relative to their performance on general protein-protein complexes. AlphaFold-Multimer succeeded on only about 20% of antigen-antibody targets in one benchmark, and a physics-based refinement protocol raised that to 43% [1](#ref-4). AlphaFold 3 is the most accurate co-folding method for this class but still far from reliable [2](#ref-15). The lack of co-evolutionary signal across V(D)J-recombined antibody chains is the root cause, so plan for sampling, rescoring, and experimental validation. ### Do I need MSAs, and how long does a prediction take? Both tools use MSAs, and MSA quality strongly influences results, though AlphaFold 3's paired MSA module extracts more from shallow alignments than AlphaFold 2 did. A local AlphaFold-Multimer run on a small dimer typically takes tens of minutes on a modern GPU once databases are set up, with MSA generation on CPUs often the slowest stage for the first run. An AlphaFold Server job usually returns within minutes to a few hours depending on queue load and job size. For a real project, budget for 25 to 100 predictions per complex plus ranking time, not for one run. ## Sources 1. [Reliable protein-protein docking with AlphaFold, Rosetta, and replica exchange.](https://pubmed.ncbi.nlm.nih.gov/40424178/) 2. [Revisiting Protein-Protein Docking: A Systematic Evaluation Framework.](https://pubmed.ncbi.nlm.nih.gov/40966105/) 3. [AlphaFold and Docking Approaches for Antibody-Antigen and Other Targets: Insights From CAPRI Rounds 47-55.](https://pubmed.ncbi.nlm.nih.gov/39831331/) 4. [Evaluation of protein-RNA Docking Web Servers for Template-Free Docking and Comparison with the AlphaFold Server.](https://pubmed.ncbi.nlm.nih.gov/41882507/) 5. [The LiPP Benchmark Set for Modeling Lipid-Protein Complexes: Comparison of Co-Folding and Docking Methods.](https://pubmed.ncbi.nlm.nih.gov/42268396/) 6. [Structural Modeling of Cytokine-Receptor-JAK2 Signaling Complexes Using AlphaFold Multimer.](https://pubmed.ncbi.nlm.nih.gov/37694948/) 7. [HDOCK-Multimer: improved structure prediction of large protein complexes by integrating docking and combinatorial assembly.](https://pubmed.ncbi.nlm.nih.gov/42875608/) 8. [Improving protein and protein interactions using pseudo-dimers derived from monomeric proteins.](https://pubmed.ncbi.nlm.nih.gov/42225645/) 9. [Protein Complex Structure Prediction With AlphaFold-Enhanced HDOCK in CAPRI Rounds 47-55.](https://pubmed.ncbi.nlm.nih.gov/40167205/) 10. [Computational methods for modeling protein-protein interactions in the AI era: Current status and future directions.](https://pubmed.ncbi.nlm.nih.gov/40398752/) 11. [Performance of pyDock in 8th CAPRI: Energy-Based Scoring Applied to Docking and AlphaFold Models.](https://pubmed.ncbi.nlm.nih.gov/40851418/) 12. [Enhancing alphafold-multimer-based protein complex structure prediction with MULTICOM in CASP15.](https://pubmed.ncbi.nlm.nih.gov/37949999/) 13. [Pupylation-Based Proximity Labeling Unravels a Comprehensive Protein and Phosphoprotein Interactome of the Arabidopsis TOR Complex.](https://pubmed.ncbi.nlm.nih.gov/40126378/) 14. [EquiRank: Improved protein-protein interface quality estimation using protein language-model-informed equivariant graph neural networks.](https://pubmed.ncbi.nlm.nih.gov/39850657/) 15. [Physical-aware model accuracy estimation for protein complex using deep learning method.](https://pubmed.ncbi.nlm.nih.gov/39916698/) 16. [Modeling Protein-Protein and Protein-Ligand Interactions by the ClusPro Team in CASP16.](https://pubmed.ncbi.nlm.nih.gov/41115690/) 17. [Modeling Protein-Protein Complexes by Combining pyDock and AlphaFold.](https://pubmed.ncbi.nlm.nih.gov/41174317/) 18. [Flexible protein-protein docking with a multitrack iterative transformer.](https://pubmed.ncbi.nlm.nih.gov/38148272/) 19. [The Construction of a Molecular Model for the Ternary Protein Complex of Intrinsic Coagulation Pathway Factors Provides Novel Insights for the Pathogenesis of Cross-Reactive Material Positive Coagulation Factor Mutations.](https://pubmed.ncbi.nlm.nih.gov/40508000/)