# AlphaFold-Multimer for Antibody-Antigen Complex Prediction By the end of this article you will be able to prepare a paired antibody-antigen FASTA input, run AlphaFold-Multimer either through ColabFold or a local installation, interpret the confidence metrics in the output (ipTM, pTM, PAE, and pLDDT), apply sampling strategies that measurably improve success rates, and recognize when a prediction is too unreliable to use for downstream design decisions. To follow the walkthrough you need either a free AlphaFold Server session in a browser, a Google account for ColabFold, or a Linux workstation with at least one modern GPU (a 16 GB A100-class card handles most antibody-antigen jobs, a 40 GB A100 handles larger antigens comfortably), the ColabFold package or the AlphaFold repository, and the mature heavy chain, light chain, and antigen sequences in FASTA format. No prior docking experience is required, but basic command-line familiarity helps for the local route. ## What AlphaFold-Multimer Does, and Why Antibody-Antigen Complexes Are Hard AlphaFold2 changed monomer structure prediction by learning to convert sequence alignments plus templates into accurate 3D coordinates. AlphaFold-Multimer (AF-Multimer) extends that architecture to complexes: you supply the sequences of every chain in the assembly, and the network folds them together, reasoning about inter-chain contacts directly rather than docking pre-folded partners [1]. For most protein-protein complexes, this works remarkably well. In a benchmark of 152 diverse heterodimers, AlphaFold produced a near-native top-ranked model for 43% of targets, which far exceeded the 9% success rate of unbound protein-protein docking on the same set [1]. That same study, however, reported that AlphaFold modeling of antibody-antigen complexes in the benchmark set was unsuccessful, and a follow-up benchmark using AlphaFold-Multimer against recently released antibody-antigen structures confirmed a low success rate of 11% [1]. T cell receptor-antigen complexes performed even worse [1]. Two structural features explain this weakness, and both matter for how you should run and interpret predictions. First, co-evolutionary signal is missing. AF-Multimer relies heavily on multiple sequence alignments (MSAs, tables of homologous sequences that reveal which residue positions co-vary across evolution) to find interfaces. Antibody and antigen chains come from different evolutionary lineages, so no [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) will ever show correlated mutations across the interface. The interface signal must come from structure learning alone, which is a much weaker cue. Second, the antibody paratope (the six complementarity-determining regions, or CDRs, that contact the antigen) is hypervariable. For a new antibody sequence, especially one with an unusual CDR length or geometry, the network has little training-like information to fall back on and can produce a confidently folded but structurally rare CDR conformation. A systematic comparison of six antibody-antigen prediction methods, including AF-Multimer, RoseTTAFold, ClusPro in antibody mode, SnugDock, and AbAdapt, found that AF-Multimer outperformed the alternatives, but that its lower-quality models showed a specific bias: their tertiary motifs (TERMs, short recurrent super-secondary structure fragments) matched fewer structures in the non-antibody portion of the [Protein Data Bank](/blog/guides/protein-data-bank) than better models did [3]. In practical terms, when an AF-Multimer antibody model is wrong, it is often wrong in a distinctive, detectable way: the CDR loops adopt PDB-unusual shapes. The absolute numbers improved substantially as the underlying models matured. A large evaluation on 427 nonredundant antibody-antigen complex structures found that the latest AlphaFold 2 version raised near-native modeling success to over 30%, versus roughly 20% for a previous version, and that increased sampling pushed success to approximately 50% [4]. Those figures frame the realistic expectation: AF-Multimer gives you a usable answer in roughly one out of three to one out of two cases, not nine out of ten, and the usable fraction depends heavily on how many models you generate and how carefully you rank them. ## Choosing a Route: AlphaFold Server, ColabFold, or a Local Install There are three practical ways to run a multimer prediction, and the right choice depends on how much control you need. The **AlphaFold Server** is the free web interface built around AlphaFold 3, the successor model with a diffusion-based architecture that handles proteins, [nucleic acids](/blog/guides/nucleic-acids), small molecules, ions, and modifications in one framework [5]. You paste sequences into a browser form, the server generates the MSAs and predictions for you, and no installation is required. The server enforces limits on the number of chains and total sequence length, so a full heterodimeric antibody (heavy and light chains) plus a large antigen can approach those limits. For a first look at a complex, or when you have no GPU access, start here. The trade-off is that you get a fixed number of predictions with no control over seeds, recycles, templates, or MSA parameters. **ColabFold** sits in the middle. It wraps AlphaFold-Multimer in a package that runs on free or paid Google Colab GPUs or on your own machine, exposes key sampling controls such as the number of seeds and recycles, and handles database downloads and MSA generation through remote MMseqs2 servers. Most step-by-step tutorials for complex prediction use ColabFold because it is reproducible and scriptable without a multi-day database setup. A **local AlphaFold-Multimer installation** from the official repository gives full control: custom MSAs, custom templates, more predictions per model, and batching over many candidates. The setup cost is real. You download several hundred gigabytes of genetic databases, and you need a GPU with enough memory for your complex size. Choose this route when you need high throughput (for example, screening many antibody candidates against one antigen, as in a nanobody discovery pipeline [6]) or when your network blocks the ColabFold remote MSA servers. For a single antibody-antigen question, the AlphaFold Server gives you a fast first answer, ColabFold gives you the sampling controls that measurably raise success rates [4], and a local install makes sense once you scale up. ## Step 1: Prepare the Sequences Correctly ![Four ordered steps for building a clean heavy chain, light chain and antigen FASTA file for AlphaFold-Multimer.](/article-images/illustrations/alphafold-multimer-antibody-antigen-prediction-fasta-prep.webp) AF-Multimer treats each FASTA entry as a separate chain, in the order given. For a conventional IgG-type antibody you must supply the mature heavy chain and the mature light chain as two entries, followed by the antigen. Sequence quality issues cause more silent failures than any software flag, so work through this checklist. Use mature sequences, not full preprotein sequences. Signal peptides, C-terminal peptides that get cleaved during secretion, and unstructured tags that are absent from the biological complex all confuse the model and waste compute. For most antibodies, this means the variable and constant regions as expressed in the antibody protein, without the leader peptide. Trim large multidomain antigens to the relevant region when you have evidence for one. If an antigen is 1,200 residues and only a 250-residue ectodomain is accessible to antibody binding, modeling the whole thing lowers interface quality and increases memory use. If you have no prior epitope information, model the full antigen and treat the prediction as exploratory. Chain identity matters for full antibodies. The heavy and light chains must be distinguishable, and the pairing between them is a known weak point for co-folding models. Name the chains explicitly in the FASTA headers so downstream analysis tools can tell them apart. A correctly formatted input for a Fab fragment in complex with an antigen looks like this: ``` >heavy_chain|antibody EVQLVESGGGLVQPGGSLRLSCAASGFTFSSYAMSWVRQAPGKGLEWVSAISGSGGSTYYADSVKGRFTISRDNSKNTLYLQMNSLRAEDTAVYYCARRGYGSSYWYFDVWGAGTTVTVSSASTKGPSVFPLAPSSKSTSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKVDKKVEPKSC >light_chain|antibody DIQMTQSPSSLSASVGDRVTITCRASQGIRNYLAWYQQKPGKAPKLLIYAASTLQSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQRYNRAPYTFGQGTKVEIKRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKVQWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGEC >antigen ECDVSSQQFTAIKSDAGAWVSFLKVEDSEVSVLLLRDQEQSGKVQVIVGADSAFDALKEVAHLTEVNSVTFHLQKAKGKVQVLQGADVFDTLTKVGHQAEVVSPALQAKEDLVVQLVVDWQEAIAAVGSGTVAAGSSSSAATLVASGALHFAAVQ ``` The pipe-separated annotations after each name are for your own bookkeeping. The parser only needs unique headers and valid amino acid characters. Common preparation errors include lowercase sequence letters mixed with residue numbers, whitespace inside the sequence lines, and duplicated FASTA headers when you paste two constructs with the same name. If you already have a crystal structure of the antibody bound to a different antigen, or a predicted monomer structure you trust, keep it. Several advanced pipelines use unbound structures as docking inputs instead of raw sequences, which changes the strengths and failure modes of the method [2]. ## Step 2: Run the Prediction With ColabFold ColabFold is the fastest scripted route for a single complex. Install it, then run the batch interface from a terminal or a notebook cell. ``` pip install "colabfold[alphafold-minus-jax]" colabfold_batch antibody_antigen.fasta ./afm_out \ --model-type alphafold2_multimer_v3 \ --num-seeds 5 \ --num-recycle 3 \ --amber ``` Each flag does specific work. `-model-type alphafold2_multimer_v3` selects the final AF-Multimer parameter set, which is the version associated with the improved antibody-antigen success rates reported in the large-scale evaluation [4]. `-num-seeds 5` runs five independent stochastic sampling trajectories per model, which is the single most effective lever against the problem that a different random seed can yield a correct model where the default run fails [4]. `-num-recycle 3` gives the network three extra refinement passes. The `-amber` flag adds Amber relaxation to clean up minor geometry violations in the final models. Expected output in `./afm_out` looks like this: ``` antibody_antigen_unrelaxed_rank_001_alphafold2_multimer_v3_model_1_seed_0000.pdb antibody_antigen_unrelaxed_rank_002_alphafold2_multimer_v3_model_1_seed_0001.pdb ... antibody_antigen_relaxed_rank_001_alphafold2_multimer_v3_model_1_seed_0000.pdb antibody_antigen_scores_rank_001_alphafold2_multimer_v3_model_1_seed_0000.json antibody_antigen_predicted_aligned_error_v1...json msas/ ``` The `scores_rank_*.json` files contain the numbers you need for ranking. Use Python to pull them out: ``` import json, glob for f in sorted(glob.glob("./afm_out/*scores*.json")): d = json.load(open(f)) print(f.split("/")[-1][:20], "iptm=%.3f" % d["iptm"], "ptm=%.3f" % d["ptm"], "mean_plddt=%.1f" % (sum(d["plddts"])/len(d["plddts"]))) ``` Expected output looks like: ``` antibody_antigen_unr iptm=0.842 ptm=0.781 mean_plddt=91.3 antibody_antigen_unr iptm=0.796 ptm=0.744 mean_plddt=88.7 antibody_antigen_unr iptm=0.412 ptm=0.519 mean_plddt=83.2 ``` A spread like this, where some seeds land near an ipTM of 0.8 and others collapse below 0.5, is normal for antibody-antigen targets and is exactly why sampling matters. Treat the highest-ipTM models as your candidate set, not a single ranked_0 file. ## Step 3: Run the Prediction With Local AlphaFold-Multimer For batch runs or custom inputs, use the official repository. After the databases are downloaded and environment variables are set, a multimer run looks like this: ``` python3 /opt/alphafold/run_alphafold.py \ --fasta_paths=antibody_antigen.fasta \ --output_dir=/data/afm_runs \ --model_preset=multimer \ --num_multimer_predictions_per_model=5 \ --db_preset=full_dbs \ --data_dir=/data/alphafold_databases \ --max_template_date=2024-01-01 \ --use_gpu_relax=true ``` The parameters that matter for antibody-antigen work are the following. `-model_preset=multimer` is mandatory for a multi-chain FASTA, since the monomer preset will reject or misbehave on multi-chain input. `-num_multimer_predictions_per_model=5` controls how many stochastic predictions each of the five neural network parameter sets produces, giving 25 models total, and increasing this value is the local equivalent of the ColabFold seed trick. `-max_template_date` cuts off template search, which prevents leakage of structures released after your knowledge cutoff into the prediction and, for antibody targets, limits over-reliance on templates of the same antibody bound to other antigens. `-db_preset=full_dbs` uses the largest sequence databases and takes considerably longer than `reduced_dbs`, but richer MSAs for the antigen side can help. Expected output follows the same structure as ColabFold: `ranked_0.pdb` through `ranked_4.pdb` per model parameter set, per-chain pLDDT in B-factor columns, PAE matrices as image and pickle files, and a JSON of scores. Models are ranked by a confidence score that combines the interface and overall metrics, so ranked_0 is usually, but not always, the model you want. ## Step 4: Read the Outputs Like a Skeptic ![Decision path from ipTM and PAE checks to a usable interface, more sampling, or a switch to AlphaFold 3 and docking.](/article-images/illustrations/alphafold-multimer-antibody-antigen-prediction-iptm-triage.webp) AF-Multimer reports four confidence numbers, and each answers a different question. Getting fluent in these is the core skill of reliable antibody-antigen prediction, because the large evaluation on 427 antibody-antigen complexes showed that confidence metrics can meaningfully predict which models are near-native [4]. | Metric | What it measures | Where to find it | Practical interpretation for antibody-antigen | |--|--|--|--| | pLDDT | Per-residue local confidence, 0 to 100 | B-factor column in the PDB | High pLDDT on framework regions is routine. Check pLDDT on the CDR loops themselves, since a confidently wrong CDR conformation is the classic failure mode [3] | | PAE | Predicted error in Angstroms for every residue pair | PAE matrix JSON or PNG | Inter-chain PAE blocks tell you how confidently the antibody is placed relative to the antigen. Dark (low) cross-chain blocks mean the interface position is well determined | | ipTM | Interface accuracy estimate derived from PAE and pTM | scores JSON | The primary model-selection metric for complexes. Values near or above roughly 0.8 usually indicate a trustworthy interface, values well below roughly 0.6 rarely correspond to a correct antibody-antigen pose [4] | | pTM | Global fold accuracy across all chains | scores JSON | Secondary check. High pTM with low ipTM means the individual chains look right but the binding pose does not | The workflow for a new prediction is simple. Sort models by ipTM. Look at the PAE matrix of the top few. If the cross-chain block is uniformly dark and the ipTM clears your confidence bar, the interface position is probably usable. If ipTM values scatter across the 0.4 to 0.7 range with no clear winner, you are in the common regime for antibody-antigen targets and you need either more sampling or an orthogonal method. When an experimental structure of your complex exists, or a close homolog does, quantify accuracy directly. DockQ is the standard continuous score for complex model quality, computed from interface contact similarity, ligand root-mean-square deviation, and interface radius-of-gyration deviation against the reference. You can compute it with the standalone DockQ script: ``` DockQ.py model_ranked_0.pdb reference_complex.pdb --short ``` This is also how the benchmark literature defines success, and the CAPRI categories (incorrect, acceptable, medium, high quality) are the vocabulary used across the field [1][7]. ## Step 5: Validate Beyond the Scores Confidence metrics are necessary but not sufficient for antibody-antigen work. Three additional checks catch failures that scores alone miss. Check the paratope for structural sanity. Superimpose your model's variable domain on a known [antibody structure](/blog/guides/antibody-structure) (any mature Fab from the PDB) and inspect the six CDR loops. CDRs H1, H2, L1, L2, and L3 adopt canonical conformations in most antibodies, so a loop that sticks out of the canonical set, combined with poor pLDDT, is a red flag. The TERM-level analysis mentioned earlier gives this intuition a formal basis: failed models tend to contain local motifs that are rare in the structural database [3]. Check the epitope for biological plausibility. A predicted epitope buried inside a domain core, or spanning a region known from prior experiments to be unexposed, indicates a wrong pose even when ipTM looks acceptable. When the antigen is well studied, cross-reference the predicted contacting residues against published mutation and mapping data. Check that the antibody is correctly assembled. For full-length antibodies, confirm that the heavy and light chains pair through their constant domains in the standard way and that the interface geometry between variable domains is reasonable. The CAPRI assessment community has repeatedly highlighted that AlphaFold-derived confidence numbers can mislead during model selection, particularly for immune-recognition targets, and that manual structural review of top models remains necessary [8]. ## Tips for Better Antibody-Antigen Predictions These tactics are ordered roughly by impact per unit of effort. Increase sampling first. The evaluation on 427 complexes found that increased AlphaFold sampling brought antibody-antigen success from the low-thirties to approximately 50% [4]. Practically, that means five seeds per model type as a floor, and 20 or more models when the target matters. Because a different random seed can rescue a failed run, never judge a target on one prediction. Optimize the MSA input. Top-performing groups at CASP16, the most recent community-wide structure prediction assessment, attributed much of their improvement over default predictions to deliberate MSA optimization, refined modeling constructs such as using partial rather than full sequences, and massive model sampling with careful selection [9]. For antibody-antigen targets, that translates into curating the antigen MSA to remove redundant or fragmented sequences and keeping the antibody chain MSAs clean so the heavy and light chains pair correctly. Use templates deliberately. For antibodies, framework templates are almost always beneficial because the fold is conserved. For the antigen, a good template stabilizes the fold the antibody must bind. The risk is epitope conformational bias from a template of the same antigen in a different context, so inspect the interface rather than accepting it blindly. Consider docking inputs when sequences alone fail. Structure-based docking methods take unbound monomer structures and predict the complex without MSAs, which sidesteps the co-evolution problem entirely [2]. A hybrid pipeline called AlphaRED combines AF-derived structures and confidence measures with physics-based replica exchange docking, and on a subset of antibody-antigen targets where AF-Multimer succeeds in only about 20% of cases, AlphaRED reaches a 43% success rate [7]. The key insight from that work is that pLDDT can be repurposed as a flexibility estimate, guiding where conformational change during binding needs to be sampled [7]. This is the recommended rescue path when a target fails repeated AF-Multimer runs. For single-domain antibodies and nanobodies, the situation is more favorable. Nanobody-antigen complexes omit the light chain entirely, removing the pairing problem, and AF-Multimer has been used prospectively as the engine of a virtual screen that identified nanobody binders to MRGPRX2, a G protein-coupled receptor, with binding confirmed in functional cellular assays [6]. That study used AF-Multimer output features to discriminate interacting from non-interacting nanobody-antigen pairs before experimental testing, a pattern worth copying if your goal is candidate prioritization rather than a single pose. Benchmark your own pipeline. The ABAG-docking benchmark provides a nonredundant antibody-antigen dataset categorized by docking difficulty and interface properties, with periodic updates, and it includes AF-Multimer among the tested methods [10]. Running your protocol on a handful of benchmark cases with known structures calibrates how your specific settings perform before you trust them on an unknown target. ## Common Mistakes and Limitations The following are the errors seen most often in real antibody-antigen prediction work, each paired with its fix. Trusting a single ranked_0 model. AF-Multimer's stochastic sampling means the top-ranked model and the best model frequently differ, and on antibody-antigen targets the success gap between one model and many models is large [4]. Always generate multiple seeds and inspect at least the top three. Treating ipTM as ground truth. A high ipTM reflects internal consistency, not biological correctness, and the model can be consistently wrong. The CAPRI community documented cases where accurate AlphaFold-derived models were poorly ranked by confidence metrics [8], and dedicated interface-quality estimators such as EquiRank exist because the self-assessment module of AF-Multimer, when repurposed for complex scoring, leaves room for improvement [11]. Use ipTM to filter, then validate the interface biologically. Inputting wrong antibody sequences. Submitting full heavy chain including [signal peptide](/knowledge/molecular-biology/signal-peptide), swapping heavy and light chain order without checking which is which, or providing a scFv when the experimental construct was a Fab all change the prediction. Verify chain assignment with a quick BLAST or alignment against the antibody's known germline. Assuming conformational change is handled. AF-Multimer co-folds from sequence, so it can miss binding-induced rearrangements that a docking method sampling from apo structures would catch. Reviews of the AI-era docking field identify protein flexibility as a persistent unsolved difficulty [12], and hybrid pipelines that combine AF-based sampling with physics-based sampling were built precisely because AF alone fails in cases with significant conformational change [7]. If your antigen undergoes a known rearrangement on binding, plan for a refinement or docking step. Ignoring the alternatives when AF-Multimer fails twice. Running the same settings with a third seed is not a plan. AlphaFold 3 substantially improves antibody-antigen accuracy over AF-Multimer v2.3 [5] and is available through the AlphaFold Server, and benchmarking work on antibody-antigen docking found AlphaFold 3 to be the most accurate method tested [13]. If you must model from sequences and AF-Multimer output is consistently low-confidence, the honest next step is the newer model, not more seeds. Over-interpreting side-chain and loop details in low-confidence models. A model with ipTM near 0.5 may still get the epitope's general surface roughly right while every CDR loop detail is wrong. Downstream tasks should match model quality: a rough epitope patch supports mutagenesis planning, an atomic interface does not support binding-affinity claims. Forgetting that predictions and reality scale differently across target classes. The 43% near-native success figure for general heterodimers does not transfer to antibody-antigen targets, where the corresponding figure was 11% in early benchmarking [1] and improved to roughly 30 to 50% with newer versions and heavier sampling [4]. Report and plan against the antibody-antigen numbers. ## When AF-Multimer Is Not Enough: The Current Tool Landscape Antibody-antigen complex prediction is a moving target, and AF-Multimer is now one engine among several. AlphaFold 3 is the most direct upgrade. Its unified diffusion-based architecture reports substantially higher antibody-antigen prediction accuracy than AlphaFold-Multimer v2.3 [5], and independent benchmarking against apo structures found it achieving a top-5 success rate of 77.98% on general complexes, versus 12.84% for the best traditional docking method HDOCK, and it remained the most accurate method for antibody-antigen docking in that evaluation [13]. For a new project starting today, the AlphaFold Server running AlphaFold 3 is the sensible default first pass, with AF-Multimer retained for its sampling controls and compatibility with existing pipelines. Docking-based and hybrid methods remain valuable, particularly for difficult targets. In recent CAPRI assessment rounds, strong teams combined physics-based sampling through ClusPro FFT docking and molecular dynamics with AlphaFold-based sampling and refinement, producing high-accuracy models for cases where AlphaFold alone failed [14]. The HDOCK team integrated AlphaFold into their template-based and ab initio hybrid strategy from Round 53 onward and improved accuracy further by refining their sampling strategy [15]. The AlphaRED pipeline demonstrated 43% success on antibody-antigen targets where AF-Multimer achieves 20% [7]. AF2Dock adapts AF-Multimer into a generative docking model that takes structures rather than MSAs, and while its success rates on non-holo inputs trail co-folding AF-Multimer and AlphaFold 3, it produces orthogonal predictions and solves targets where co-folding fails, including outperforming other docking methods on nanobody complexes [2]. The research frontier addresses the root cause, which is training data scarcity for complexes. A pre-training method called split and merge proxy constructs pseudo-dimers by splitting abundant monomeric proteins into two subunits and training models to merge them back, and models pre-trained this way surpass both AF-Multimer and AlphaFold 3 on several CASP15 dimer targets [16]. This line of work suggests antibody-antigen accuracy will keep improving without waiting for exponentially more experimental complex structures. A practical tiered protocol emerges from all of this. Start with the AlphaFold Server for a fast AlphaFold 3 answer. Run AF-Multimer with heavy sampling through ColabFold for comparison and seed diversity. If the two disagree or both show low confidence, move to hybrid approaches such as AlphaRED or a CAPRI-style docking-plus-refinement workflow, and benchmark your protocol on known antibody-antigen structures before committing to a pose. ## Frequently Asked Questions ### Can AlphaFold-Multimer predict antibody-antigen complexes reliably? Only part of the time. Early benchmarking found 11% success [1], newer versions and increased sampling raise success to roughly 30 to 50% [4], and AF-Multimer still outperformed other sequence-based antibody-antigen methods in a six-method comparison [3]. Treat every prediction as a hypothesis to validate, not an answer. ### Should I use the AlphaFold Server or run AlphaFold-Multimer locally? Use the AlphaFold Server for a fast first answer with AlphaFold 3 and no installation, and run ColabFold or local AlphaFold-Multimer when you need control over seeds, sampling counts, templates, and MSAs. Sampling control matters most for antibody-antigen targets because success rises substantially with more sampled models [4]. ### What ipTM score means my antibody-antigen model is trustworthy? Use ipTM as your primary filter, with values near or above roughly 0.8 usually indicating a reliable interface and values well below roughly 0.6 rarely corresponding to correct poses, then confirm with the PAE matrix and biological checks. Confidence metrics can mislead, so never rely on ipTM alone [8]. ### How do I format a FASTA file for a full antibody plus an antigen? Supply three entries in one FASTA file: the mature heavy chain, the mature light chain, and the antigen, each with a unique header. Remove signal peptides and non-native tags, and trim the antigen to the relevant domain when you have evidence for the binding region. ### Why does AlphaFold-Multimer fail on antibody-antigen targets? The interface lacks co-evolutionary signal because antibody and antigen come from different evolutionary lineages, so MSAs cannot teach the model where the interface is [2], and hypervariable CDR loops in novel sequences can fold into PDB-rare conformations in failed models [3]. ### Should I use AlphaFold 3 instead of AlphaFold-Multimer? For a new project, try both. AlphaFold 3 shows substantially higher antibody-antigen accuracy than AF-Multimer v2.3 [5] and was the most accurate method in independent antibody-antigen docking benchmarks [13], while AF-Multimer offers finer sampling controls that remain valuable for hard targets [4]. ### How many models should I generate for one antibody-antigen prediction? At least five seeds per model as a floor, and 20 or more for important targets, because increased sampling raised antibody-antigen success to approximately 50% in a large evaluation [4] and top CASP16 groups relied on massive model sampling with careful selection [9]. ### Can AlphaFold-Multimer be used for antibody discovery, not just pose prediction? Yes. A prospective virtual screen used AF-Multimer outputs to discriminate interacting from non-interacting nanobody-antigen pairs and identified nanobodies that bound MRGPRX2 with high affinity, confirmed in functional assays, establishing AF-Multimer as a viable engine for computational discovery pipelines [6]. ## Sources 1. [Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants.](https://pubmed.ncbi.nlm.nih.gov/35900023/) 2. [Adapting Co-Folding Models for Structure-Based Protein-Protein Docking Through Flow Matching.](https://pubmed.ncbi.nlm.nih.gov/41377499/) 3. [A comparison of antibody-antigen complex sequence-to-structure prediction methods and their systematic biases.](https://pubmed.ncbi.nlm.nih.gov/39167052/) 4. [Evaluation of AlphaFold antibody-antigen modeling with implications for improving predictive accuracy.](https://pubmed.ncbi.nlm.nih.gov/38073135/) 5. [Accurate structure prediction of biomolecular interactions with AlphaFold 3.](https://pubmed.ncbi.nlm.nih.gov/38718835/) 6. [In silico discovery of nanobody binders to a G-protein coupled receptor using AlphaFold-Multimer.](https://pubmed.ncbi.nlm.nih.gov/42026072/) 7. [Reliable protein-protein docking with AlphaFold, Rosetta, and replica exchange.](https://pubmed.ncbi.nlm.nih.gov/40424178/) 8. [AlphaFold and Docking Approaches for Antibody-Antigen and Other Targets: Insights From CAPRI Rounds 47-55.](https://pubmed.ncbi.nlm.nih.gov/39831331/) 9. [Assessment of Protein Complex Predictions in CASP16: Are We Making Progress?](https://pubmed.ncbi.nlm.nih.gov/41170922/) 10. [ABAG-docking benchmark: a non-redundant structure benchmark dataset for antibody-antigen computational docking.](https://pubmed.ncbi.nlm.nih.gov/38385879/) 11. [EquiRank: Improved protein-protein interface quality estimation using protein language-model-informed equivariant graph neural networks.](https://pubmed.ncbi.nlm.nih.gov/39850657/) 12. [Computational methods for modeling protein-protein interactions in the AI era: Current status and future directions.](https://pubmed.ncbi.nlm.nih.gov/40398752/) 13. [Revisiting Protein-Protein Docking: A Systematic Evaluation Framework.](https://pubmed.ncbi.nlm.nih.gov/40966105/) 14. [Modeling Protein-Protein and Protein-Ligand Interactions by the ClusPro Team in CASP16.](https://pubmed.ncbi.nlm.nih.gov/41115690/) 15. [Protein Complex Structure Prediction With AlphaFold-Enhanced HDOCK in CAPRI Rounds 47-55.](https://pubmed.ncbi.nlm.nih.gov/40167205/) 16. [Improving protein and protein interactions using pseudo-dimers derived from monomeric proteins.](https://pubmed.ncbi.nlm.nih.gov/42225645/)