# AlphaFold Input Requirements: Step-by-Step Guide By the end of this guide you will be able to prepare a correct AlphaFold input for a single protein or a protein complex, decide between a web-based AlphaFold server and a local installation, generate or sanity-check a [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) (MSA), submit a prediction, and verify that the output structure is worth trusting. You will need a protein [amino acid sequence](/blog/guides/amino-acid-sequence) in FASTA format, access to either an AlphaFold server or a machine with a modern GPU, and about an hour for your first run. This guide works as a hands-on alphafold tutorial. Every command shown has been written so you can copy it, run it, and see the same kind of output on your own terminal. ## AlphaFold Input Requirements at a Glance AlphaFold is a deep learning system that converts a protein amino acid sequence, plus evolutionary context, into atomic 3D coordinates. The input side of that pipeline is where most runs fail, so it pays to be precise about what the model actually consumes. | Input component | Required? | Typical source | What it affects | |--|--|--|--| | Amino acid sequence (FASTA) | Always | UniProt, NCBI RefSeq, or your own construct | Defines every residue the model builds | | Multiple sequence alignment (MSA) | Required for full accuracy | UniRef, MGnify, BFD, or MMseqs2 web search | Encodes co-evolutionary signal used by the model | | Structural templates | Optional but recommended | PDB via HHsearch or MMseqs2 | Anchors fold choice when homologs are sparse | | Chain definitions and stoichiometry | Required for complexes | Your construct design or a complex partner list | Drives AlphaFold-Multimer interface prediction | | Ligands, [nucleic acids](/blog/guides/nucleic-acids), modifications | Optional (AlphaFold 3 only) | Manual input on the AlphaFold server | Models protein-DNA, protein-RNA, and chemical contexts | The single most consequential input is the MSA. A mechanistic analysis of AlphaFold 3's internal representations found that removing the MSA collapses the model's latent geometry entirely, regardless of how familiar the sequence is, and that a minimal set of highly divergent homologs is sufficient to anchor the representation and activate the model's structural priors [1]. In plain terms, the model needs evolutionary context, and the context needs to be phylogenetically diverse rather than merely deep. ## What AlphaFold Accepts as Input ### The amino acid sequence AlphaFold accepts a standard FASTA file with the canonical one-letter amino acid codes. The 20 standard residues are expected. The ambiguity code X is accepted as a placeholder for unknown residues, though any position modeled as X will carry low confidence. Non-standard residues, numbering artifacts, and anything that is not a plain [protein sequence](/blog/guides/protein-sequence) must be removed or handled explicitly before submission. The canonical sequence matters. If your experimental construct contains an N-terminal His-tag, a linker, or a truncation, you must decide whether to predict the native protein or the construct. Predicting the native UniProt sequence gives you the cleanest comparison against the AlphaFold [Protein Structure](/blog/guides/protein-structure) Database, which contains over 200 million predicted structures as of May 2023 [2]. ### The multiple sequence alignment The MSA is not an optional garnish. It is the substrate from which AlphaFold extracts co-evolutionary signal. Practical evidence backs this up: in CASP15 assembly prediction, a system (MULTICOM) that sampled diverse MSAs and templates, including structure-based alignments from Foldseek, raised the average TM-score of first predictions to roughly 0.76 compared with about 0.72 for standard AlphaFold-Multimer, and best-of-top-5 performance from about 0.74 to about 0.80 [3]. Better MSA diversity, not just more sequences, translated directly into better models. The same principle holds at the level of the model's internals. The representational coherence of AlphaFold 3 requires phylogenetic diversity rather than alignment depth, so a handful of well-spread, highly divergent homologs can outperform a deep but narrow alignment [1]. ### Structural templates Templates are experimentally determined structures from the [Protein Data Bank](/blog/guides/protein-data-bank) that align to your query sequence. They are optional, they are usually ignored when the MSA is strong, and they rescue predictions when the MSA is thin. A common configuration issue is an overly restrictive template date cutoff that silently excludes useful PDB entries. ### Multichain and complex inputs AlphaFold-Multimer takes a FASTA file with one entry per chain and predicts the assembled complex. The MSA strategy also matters here, since co-folding models depend on MSAs for their interface predictions, which is precisely why MSA-free docking approaches have been developed for hard targets such as antibody-antigen pairs where co-folding underperforms but produces complementary alternative predictions [4]. ### AlphaFold 3 additions AlphaFold 3 extends the input schema beyond proteins to nucleic acids, chemical modifications, and common ligands, specified on the public AlphaFold server through a structured job builder rather than raw FASTA alone. Mechanistic work shows the network converts this sequence and evolutionary data into structural coordinates through internal pair representations that become linearly decodable for complex biophysical features [1], which is a useful reminder that everything you feed in is compressed into that latent geometry. Garbage in propagates. ## Choosing Between an AlphaFold Server and a Local Installation ![Decision path asking whether you need custom MSAs, batch runs or large complexes, leading to a local install or the AlphaFold server.](/article-images/illustrations/alphafold-input-requirements-server-vs-local.webp) Your input preparation workflow depends heavily on where you run the model. The public AlphaFold server (alphafoldserver.com) is the official DeepMind-hosted AlphaFold 3 service. You paste or upload sequences, add nucleic acid or ligand components interactively, and the server handles MSA generation and template search internally. It enforces a per-job token limit that counts all chains, ligands, and modifications combined, so very large complexes must be budgeted carefully. This is the correct choice for most biologists making a handful of predictions. Unified web portals reduce the friction further. Foldify is a free web application that integrates AlphaFold 3, AlphaFold 2, ColabFold, OmegaFold, and ESMFold behind a single graphical interface, accepts plain protein sequences as input, runs multiple models on the same sequence, and displays results in the Mol* viewer without requiring a login [5]. This is useful when you want to compare predictions from an MSA-dependent model (AlphaFold 2, AlphaFold 3) against sequence-only language models (OmegaFold, ESMFold) on the same input. Local installations make sense when you need batch processing, custom MSAs, reproducible settings, or proteins too large for server limits. The tradeoff is real. The full AlphaFold 2 database set is on the order of two terabytes, and a consumer GPU with 8 to 12 GB of memory limits you to modest sequence lengths without mitigation. For MSA-sensitive problems, note that co-folding models remain limited by MSA availability for challenging targets such as antibody-antigen complexes, and specialized adaptations restructure the model's input pathway to work around this [4]. If your target is an antibody or a rapidly evolving protein family, plan for weak or empty MSAs from the start. ## Step-by-Step Guide to Preparing AlphaFold Inputs ### Step 1: Obtain the canonical sequence Fetch the UniProt canonical sequence directly rather than copying from a webpage or manuscript. Copy-paste from documents introduces smart quotes, en-dashes, and hidden Unicode that break parsers. ```bash curl -s "https://rest.uniprot.org/uniprotkb/P04637.fasta" -o p53.fasta head -2 p53.fasta ``` Expected output: ``` >sp|P04637|P53_HUMAN Cellular tumor antigen p53 OS=Homo sapiens OX=9606 GN=TP53 PE=1 SV=4 MEEPQSDPSVEPPLSQETFSDLWKLLPENNVLSPLPS... ``` The header line contains the accession, protein name, organism, and gene. Keep this file untouched as your reference sequence. ### Step 2: Check whether a structure already exists Before spending compute on a prediction, check the AlphaFold Protein Structure Database. For well-studied organisms the answer is usually yes. The AlphaFoldDB Structure Extractor is a web server and API that accepts diverse identifiers, including NCBI Taxonomy IDs, RefSeq accessions, old and new locus tags, UniProt accessions, and AlphaFold accessions, and returns structures in PDB, mmCIF, bCIF, or PAE JSON formats with an accompanying ID mapping file for bulk queries up to thousands of accessions [6]. For a single known accession, direct download works: ```bash curl -s "https://alphafold.ebi.ac.uk/files/AF-P04637-F1-model_v4.pdb" -o AF-P04637-F1.pdb head -1 AF-P04637-F1.pdb ``` Expected output is a REMARK line followed by ATOM records: ``` REMARK 1 ALPHAFLOLD RESTRICTED ... ATOM 1 N MET A 1 -3.524 8.309 30.557 1.00 91.24 N ``` The number in the last numeric column (91.24 here) is the per-residue pLDDT score stored in the B-factor field, which Step 8 explains how to interpret. If you instead want structurally similar proteins across the database, AlphaFind accepts any valid UniProt ID, PDB ID, or gene symbol and returns the most structurally similar chains from AlphaFold DB with similarity metrics and superposition views [7]. ### Step 3: Format and validate the FASTA file AlphaFold is unforgiving about sequence content. Run a quick validation: ```python # validate_fasta.py seq = "".join(line.strip() for line in open("p53.fasta") if not line.startswith(">")) allowed = set("ACDEFGHIKLMNPQRSTVWYX") bad = set(seq.upper()) - allowed print(f"length: {len(seq)} invalid characters: {sorted(bad)}") ``` Expected output: ``` length: 393 invalid characters: [] ``` If the invalid character list is nonempty, strip gaps (`-`), asterisks (`*`), whitespace, digits, and lowercase ambiguity before submission. Replace unknown residues with X. Never submit a [nucleotide sequence](/knowledge/molecular-biology/nucleotide-sequence) to a protein predictor. For multimer predictions, ColabFold expects one FASTA entry per chain with a header describing the species: ``` >TP53|Homo sapiens MEEPQSDPSVEPPLSQETFSDLWKLLPENNVLSPLPS... >MDM2|Homo sapiens MCNTNMSVPTDKVKFVVYFLKMKHKY... ``` ### Step 4: Plan the MSA If you use the AlphaFold server or ColabFold's default mode, MSA generation is automatic. The AlphaFold server searches standard sequence databases and builds MSAs for you. ColabFold queries MMseqs2 against UniRef and environmental databases over the network. If you run locally or need a custom MSA, two principles apply. First, diversity beats depth. A minimal set of highly divergent homologs anchors the model's internal representation, and extremely deep but narrow alignments add little [1]. Second, sampling several distinct MSAs and templates improves results measurably. The MULTICOM system, which sampled diverse MSAs and templates using both sequence and Foldseek structure-based alignments, outperformed standard AlphaFold-Multimer in CASP15 assembly prediction [3]. Check your MSA depth after generation: ```bash grep -c '^>' results/p53.a3m ``` Expected output: ``` 5129 ``` A count in the hundreds to low thousands of nonredundant sequences is healthy for most proteins. If you see fewer than about 20 sequences, expect a poor model and consider either widening the database search or switching to a sequence-only model such as OmegaFold or ESMFold, which are available through portals like Foldify [5]. ### Step 5: Handle templates deliberately ColabFold enables templates with a flag: ```bash colabfold_batch --templates --num-recycle 3 p53.fasta results/ ``` For a local AlphaFold 2 installation, the template cutoff date is an explicit argument. Set it to a date you can justify in your methods section: ```bash python3 run_alphafold.py \ --fasta_paths=p53.fasta \ --output_dir=outputs \ --data_dir=/data/alphafold \ --db_preset=full_dbs \ --model_preset=monomer \ --max_template_date=2025-01-01 \ --use_gpu_relax=true ``` Key local flags: - `-model_preset=monomer` for single chains, `multimer` for complexes. - `-db_preset=full_dbs` uses the complete UniRef90, MGnify, and BFD databases, roughly two terabytes. `reduced_dbs` trades some sensitivity for speed. - `-max_template_date` filters PDB templates by deposition date. Setting it before your test-set cutoff matters for benchmarks, and setting it too early silently discards helpful templates. - `-use_gpu_relax=true` runs the amber relax step on GPU instead of CPU. ### Step 6: Build complex (multimer) inputs For a two-chain prediction, put both chains in one FASTA file and specify the multimer model: ```bash colabfold_batch --model-type alphafold2_multimer_v3 complex.fasta results/ ``` Expected output during the run: ``` Loading MSA... Running model_1_pred_0 model_1_pred_0: pLDDT=88.4 ptm=0.71 iptm=0.64 ``` The `iptm` (interface pTM) score is the number to watch for complexes. High pTM with low ipTM means each chain folded well but the interface is probably wrong. Specify stoichiometry explicitly, for example two chains of A and one of B as three FASTA entries, because the model will not infer oligomer state from sequence alone. ### Step 7: Submit on the AlphaFold server On the AlphaFold server, the input builder walks through entities. For a protein-nucleic acid complex you add each protein chain by sequence, each nucleic acid chain by sequence, and any ligand by its Chemical Component Dictionary three-letter code. Modifications and ions are selected from the supported list. The server generates MSAs automatically and returns a ranked set of up to five models per job. Practical server tips: - Paste sequences without numbering, gaps, or trailing stop characters. - Combine interacting partners in one job rather than predicting monomers separately, because joint prediction captures interfaces. - Budget tokens. Every residue, nucleotide, and ligand atom-count entity consumes the shared limit, so a job with several long chains plus many ligands can hit the ceiling quickly. ### Step 8: Validate the output Three files matter in every result directory: the coordinate file (mmCIF natively for AlphaFold, PDB in ColabFold), the confidence summary, and the predicted aligned error (PAE) JSON that encodes pairwise domain uncertainty [6]. Compute the mean pLDDT from a PDB file directly: ```bash awk '$1=="ATOM" && $3=="CA" {sum+=$11; n+=1} END {printf "mean pLDDT: %.1f\n", sum/n}' results/*.pdb ``` Expected output: ``` mean pLDDT: 91.2 ``` Interpret pLDDT bands the same way the AlphaFold developers color their structures. Above 90, backbone accuracy is high and side chains are usually well placed. Between 70 and 90, the fold is generally right and details are uncertain. Between 50 and 70, treat the region as a low-resolution sketch. Below 50, the model itself is declaring the region unstructured. One structural survey of AlphaFold models reported that low-confidence regions correlate with intrinsically disordered regions, which is exactly the biology you want the model to flag [8]. For anything with more than one domain, read the PAE plot before the structure. Blocks of low PAE within a domain with high PAE between domains mean each domain is reliable and the relative domain orientation is a guess. ## Common Errors and Their Fixes ![Flowchart branching from a FASTA validation check into fixes for invalid characters, empty MSAs and rejected multimer headers.](/article-images/illustrations/alphafold-input-requirements-fasta-triage.webp) **1. Out-of-memory crash during forward pass.** A representative JAX message looks like `RESOURCE_EXHAUSTED: Out of memory while flattening allocation.` This means the model, MSA, or sequence length exceeded GPU memory. Fixes in order of preference: reduce `-max-msa` (for example `-max-msa 256:512` in ColabFold) to cap MSA depth, predict domains separately and assemble them, run fewer model parameterizations in one pass, or move to a cloud notebook or a larger GPU. **2. Invalid characters rejected at parse time.** You will see an error stating that the sequence contains characters outside the allowed set, typically after pasting from a document. The fix is the validation script in Step 3. Retype the sequence from the raw UniProt FASTA rather than cleaning a contaminated one by hand. **3. Empty or near-empty MSA.** The log reports that no homologs were found and the run falls back to a single-sequence or shallow input. AlphaFold 3's internal representations collapse without evolutionary context [1], so do not expect a trustworthy model here. Widen the search databases, lower the hit threshold, add divergent orthologs manually to the A3M, or switch to a sequence-only model such as OmegaFold or ESMFold for a rough fold-level estimate [5]. **4. Multimer FASTA rejected.** With ColabFold, chain headers must follow the pipe-delimited convention (`>name|species`). A plain `>chainA` header can cause the paired MSA stage to misassign species labels and degrade interface predictions. Rewrite headers and rerun. **5. Templates silently absent.** No error appears, but models are worse than expected. Check that `-max_template_date` is not excluding recent PDB entries, and confirm the template database path actually contains the distilled PDB files. In ColabFold, verify you passed `-templates`. **6. Chain-count limits in output formats.** Legacy PDB format caps a structure at 62 chains because of single-character chain identifiers. Large complexes should be read from mmCIF output, which AlphaFold writes natively. ## Common Mistakes and Limitations These are the mistakes that show up repeatedly in real analyses of AlphaFold inputs. **Submitting a construct instead of the native sequence.** His-tags, cleavage sites, and linkers in the input produce predicted structures for a protein that does not exist biologically. Either predict the native sequence or justify the engineered construct in your methods. **Optimizing MSA depth instead of MSA diversity.** Deep alignments drawn from a narrow clade add little. The representational analysis of AlphaFold 3 showed that a small set of highly divergent homologs is sufficient to anchor the latent space, and that diversity, not depth, is what the model requires [1]. When curating a custom MSA, prioritize covering distinct phylogenetic branches. **Reading pLDDT as a global structure quality score.** pLDDT is per residue. A model with mean pLDDT of 85 can contain a 120-residue region at pLDDT 40 that is simply disordered. Confidence correlates with intrinsic disorder across AlphaFold model collections [8], so low scores in termini and loops are frequently correct biology, not model failure. **Ignoring the PAE for multidomain and multichain predictions.** The single most common misinterpretation is presenting a multidomain model as a rigid object when the PAE block structure says domain orientations are unresolved. **Expecting MSA-dependent performance from MSA-poor targets without a plan.** Antibodies, orphan proteins, and fast-evolving families often have weak MSAs. Co-folding models remain limited by MSA reliance on such targets, and complementary approaches that bypass MSAs exist for docking problems where co-folding fails [4]. Decide before the run which regime your target is in. **Treating the prediction as an experimental structure.** An AlphaFold model is a static, averaged prediction of a conformational ensemble under no ligand, membrane, or partner context. Downstream uses that accept this, such as embedding extraction for stability prediction, which has been shown to predict experimentally measured stability changes accurately [9], work with the model rather than against it. Functional annotation pipelines similarly treat models as features rather than truth, for example ProtPen, which takes a FASTA file with UniProt identifiers and combines eggNOG-mapper sequence annotation with Foldseek searches against AlphaFold-predicted structures, reaching over 90 percent annotation accuracy on a curated Pseudomonas aeruginosa benchmark [10]. **Skipping the database check.** For human proteins, model organisms, and most prokaryotic proteomes, a structure already exists in AlphaFold DB. Bulk retrieval tools map RefSeq accessions, locus tags, and taxonomy IDs onto AlphaFold entries and export coordinates plus PAE files [6], and structure-similarity search engines let you find related folds without running any prediction at all [7]. ## Beyond the Single Prediction: What the Input Enables Well-formed inputs feed a growing ecosystem. PANDA-3D predicts gene ontology terms from AlphaFold models using geometric graph neural networks over the predicted coordinates combined with language model sequence embeddings, and outperforms comparable methods trained only on experimental structures [2]. Structure-aware protein language models fuse AlphaFold-derived structural embeddings with sequence embeddings to improve classification and clustering tasks beyond sequence-only baselines [11]. Pipelines such as SNPeffect consume variant call data alongside AlphaFold DB models to assess missense variant effects on stability and aggregation across the human proteome [12]. The common thread is that every one of these tools starts from the same humble inputs described in this guide: a clean FASTA sequence and, where applicable, a diverse MSA. Getting those right multiplies the value of everything downstream. ## Frequently Asked Questions ### What is the minimum input AlphaFold needs? A single amino acid sequence in FASTA format is the minimum. Accuracy improves substantially when an MSA and templates accompany it, but the sequence alone is enough for the model to attempt a prediction. ### Do I always need a multiple sequence alignment? For highest accuracy, yes, and the model's internal coherence depends on evolutionary context. On the AlphaFold server the MSA is generated for you automatically, and in ColabFold it is fetched over the web, so in practice you rarely build one by hand for standard runs. ### How long a sequence can AlphaFold handle? The public AlphaFold server enforces a per-job token limit covering all chains, ligands, and modifications combined, so split very large assemblies into separate jobs. Locally, the practical limit is GPU memory, and very long proteins usually require domain splitting or reduced MSA settings. ### Which AlphaFold version should I use? Use AlphaFold 3 when your question involves complexes with nucleic acids, ligands, or modifications. Use AlphaFold 2 or ColabFold when you need custom MSAs, batch processing, or fine control over recycling and seeds. Web portals such as Foldify run both families on the same sequence input, which makes direct comparison straightforward [5]. ### Can AlphaFold predict protein-DNA or protein-RNA complexes? AlphaFold 3 handles protein-nucleic acid complexes natively through the server interface. AlphaFold 2 and AlphaFold-Multimer model protein chains only. ### What file formats come out of a prediction? AlphaFold writes mmCIF coordinates natively, ColabFold additionally writes PDB files, and both provide PAE data as a JSON file [6]. Confidence values are stored in the B-factor column of the coordinate files. ### Why is my predicted structure low confidence? Low confidence usually traces back to a sparse or undiverse MSA, a genuinely disordered region, absence of useful templates, or an interface the model cannot resolve. Check MSA composition first, then look at whether low pLDDT tracks expected disorder. ### Where can I get already-predicted structures instead of running AlphaFold? The AlphaFold Protein Structure Database holds predictions for UniProt entries across many organisms, and bulk tools map RefSeq, locus tag, and taxonomy identifiers onto those entries for download in PDB, mmCIF, bCIF, or PAE JSON formats [6]. For finding structurally similar proteins across the database, AlphaFind searches by UniProt ID, PDB ID, or gene symbol [7]. ## Sources 1. [AlphaInterp: Mechanistic Interpretability of AlphaFold 3 Reveals How Evolutionary Information Shapes Protein Structure Prediction.](https://pubmed.ncbi.nlm.nih.gov/42079181/) 2. [PANDA-3D: protein function prediction based on AlphaFold models](https://pubmed.ncbi.nlm.nih.gov/39108640/) 3. [Enhancing alphafold-multimer-based protein complex structure prediction with MULTICOM in CASP15](https://pubmed.ncbi.nlm.nih.gov/37949999/) 4. [Adapting Co-Folding Models for Structure-Based Protein-Protein Docking Through Flow Matching.](https://pubmed.ncbi.nlm.nih.gov/41377499/) 5. [Foldify: Web Application for Protein Structure Prediction.](https://pubmed.ncbi.nlm.nih.gov/42359896/) 6. [AlphaFold Database Structure Extractor: a web server and API to download AlphaFold structures using common protein accessions.](https://pubmed.ncbi.nlm.nih.gov/41291409/) 7. [AlphaFind: discover structure similarity across the proteome in AlphaFold DB](https://pubmed.ncbi.nlm.nih.gov/38747341/) 8. [Real-time structure search and structure classification for AlphaFold protein models](https://pubmed.ncbi.nlm.nih.gov/35383281/) 9. [Applications of AlphaFold beyond Protein Structure Prediction](https://doi.org/10.1101/2021.11.03.467194) 10. [ProtPen Combines Sequence- and Structure-based Approaches to Facilitate Protein Function Predictions on a Proteome-wide Scale.](https://pubmed.ncbi.nlm.nih.gov/42695892/) 11. [S-PLM: Structure-Aware Protein Language Model via Contrastive Learning Between Sequence and Structure.](https://pubmed.ncbi.nlm.nih.gov/39665266/) 12. [SNPeffect 5.0: large-scale structural phenotyping of protein coding variants extracted from next-generation sequencing data using AlphaFold models.](https://pubmed.ncbi.nlm.nih.gov/37464277/)