ESMFold

Unlike AlphaFold2 or RoseTTAFold, ESMFold predicts atomic-resolution 3D structures directly from a single amino acid sequence without requiring Multiple Sequence Alignments (MSAs) or templates. The model internalizes evolutionary and biophysical constraints during pre-training on hundreds of millions of protein sequences, allowing it to bypass the computationally expensive MSA generation step entirely.

• Reduces inference time from minutes/hours to seconds per protein
• Enables metagenomic-scale structure prediction on millions of sequences
• Eliminates dependency on homologous sequence databases like UniRef or BFD

ESMFold attaches a lightweight folding trunk and structure module to the frozen representations of the ESM-2 protein language model. The transformer's internal attention maps capture residue-residue interactions that directly correlate with 3D spatial proximity, effectively learning a statistical potential of protein folding.

• Leverages 650M+ parameter ESM-2 language model backbone
• Folding head trained end-to-end on PDB structures with geometric losses
• Internal representations encode secondary structure, contact maps, and solvent accessibility

ESMFold outputs full backbone coordinates (N, Cα, C) and Cβ positions for all residues, achieving sub-3Å RMSD on folded domains. The model predicts pLDDT confidence scores per residue and Predicted Aligned Error (PAE) matrices for assessing domain orientation reliability.

• Outputs compatible with standard PDB format for downstream analysis
• Confidence metrics enable automated filtering of low-quality regions
• Captures intrinsically disordered regions (IDRs) with low pLDDT, accurately reflecting structural flexibility

The speed of ESMFold enables structural annotation of entire metagenomic databases that were previously computationally intractable. Meta AI demonstrated this by folding over 600 million metagenomic protein sequences in just two weeks, revealing thousands of novel folds absent from the PDB.

• Unlocks dark matter of protein universe from environmental sequencing
• Identifies novel structural families for enzyme engineering and drug discovery
• Enables large-scale evolutionary analysis of fold emergence across biomes

ESMFold provides per-residue pLDDT and pairwise PAE metrics that correlate strongly with experimental validation accuracy. These confidence scores allow researchers to identify reliable domains versus disordered regions without requiring experimental structures.

• pLDDT > 90: High confidence, suitable for drug docking and mutagenesis studies
• pLDDT 70-90: Good backbone prediction, useful for fold classification
• pLDDT < 50: Likely disordered or low-confidence region, interpret with caution
• PAE matrices reveal domain boundaries and relative orientation uncertainty

A computational method that aligns three or more biological sequences to identify regions of evolutionary conservation and homology.

• Role in Structure Prediction: MSAs reveal co-evolving residues that are spatially proximal in the 3D structure, providing critical distance constraints.
• ESMFold's Innovation: Bypasses the need for explicit MSA input by using a large language model to internalize evolutionary patterns during pre-training.
• Bottleneck: Searching and constructing deep MSAs against large sequence databases is the primary speed bottleneck for traditional methods like AlphaFold2.

A collection of structurally distinct states representing the intrinsic dynamic flexibility of a protein, moving beyond a single static prediction.

• Limitation of ESMFold: Like AlphaFold2, ESMFold is primarily trained to predict a single, static ground-state structure.
• Emerging Solutions: Denoising Diffusion Probabilistic Models (DDPMs) are being applied to generate diverse structural ensembles by iteratively denoising random atomic coordinates.
• Biological Relevance: Capturing multiple conformations is critical for understanding allostery, enzyme catalysis, and cryptic binding pocket formation.

A per-residue confidence metric output by structure prediction models that estimates local accuracy on a scale from 0 to 100.

• Interpretation: Scores > 90 indicate very high confidence; scores < 50 suggest disorder or low reliability.
• ESMFold Output: ESMFold provides pLDDT scores, allowing researchers to distinguish reliably folded domains from Intrinsically Disordered Regions (IDRs).
• Practical Use: Critical for filtering predictions in large-scale metagenomic studies to focus analysis on high-confidence structural elements.