Thursday, November 19, 2026
HomeData ScienceProtein Structure Prediction – AlphaFold2, ESMFold, ProteinMPNN and Structural Biology ML

Protein Structure Prediction – AlphaFold2, ESMFold, ProteinMPNN and Structural Biology ML

Table of Content

📋 KEY INSIGHTS

  • Protein structure prediction — determining the 3D arrangement of a protein’s atoms from its amino acid sequence — was one of the grand challenges of computational biology for 50 years. The challenge arises from Levinthal’s paradox: a 100-amino-acid protein has 10^47 possible conformations if each backbone dihedral angle can take just three values, yet proteins fold reproducibly to their native structure in milliseconds. AlphaFold2 (Jumper et al., Nature 2021) solved this challenge: trained on the PDB (Protein Data Bank), it predicts protein structures from sequence with accuracy competitive with experimental X-ray crystallography on the CASP14 benchmark. AlphaFold2’s key innovations were the Evoformer attention module (processing multiple sequence alignments and pairwise residue representations jointly), the structure module (directly predicting 3D coordinates from the Evoformer output using invariant point attention), and iterative recycling (feeding the predicted structure back into the Evoformer for refinement).
  • The Evoformer (Evolutionary former) is AlphaFold2’s core architecture: a Transformer that processes two representations simultaneously — the multiple sequence alignment (MSA) representation (a 2D matrix of aligned homologous sequences providing evolutionary information) and the pair representation (a 2D pairwise residue matrix encoding predicted inter-residue distances and orientations). The MSA representation captures which residue positions co-evolve across species (indicating spatial proximity in the folded structure — a hallmark of structural contacts). The pair representation is updated by triangular attention (a custom attention variant that enforces the triangle inequality of 3D distances: if residue A is close to B and B is close to C, A and C should be close). These two representations are updated jointly through 48 Evoformer blocks before being passed to the structure module for 3D coordinate prediction.
  • ESMFold (Lin et al., Science 2023) achieves AlphaFold2-quality structure prediction without multiple sequence alignment (MSA) — using only the query sequence fed through a large protein language model (ESM-2 with 15 billion parameters). The key insight: ESM-2 is trained as a masked language model on 250 million protein sequences from UniProt, learning rich evolutionary and structural representations from sequence statistics without explicit alignment. These representations capture residue co-evolution implicitly through the language model’s attention patterns, so MSA computation (the most expensive preprocessing step in AlphaFold2, requiring sequence homology search against large databases) is unnecessary. ESMFold achieves 10× faster prediction than AlphaFold2 at comparable accuracy on well-represented protein families, enabling proteome-scale structure prediction.
  • Protein language models (pLMs) — BERT-style models trained on protein sequences as “text” — have emerged as universal encoders for protein biology, learning representations that capture structure, function, evolutionary conservation, and mutational fitness landscapes without requiring labelled data. ESM-2 is the leading pLM: its representations achieve state-of-the-art on protein property prediction (thermostability, binding affinity, catalytic activity), protein structure prediction (via ESMFold), and protein variant effect prediction (predicting whether a mutation is beneficial or deleterious). The connection to our foundation models guide is direct: ESM-2 is a domain-specific foundation model for protein biology — analogous to GPT pre-training on text, ESM-2 pre-trains on protein sequences and provides task-agnostic representations fine-tuned for downstream tasks, following the same pre-train/fine-tune paradigm described there for time series, vision, and genomics foundation models.
  • Inverse protein folding — designing amino acid sequences that fold into a specified 3D structure — is the complement of structure prediction and the key computational tool for protein engineering and drug design. ProteinMPNN (Dauparas et al., Science 2022) uses a message-passing GNN over the protein structure graph (residues as nodes, spatial contacts as edges with 3D coordinate features) trained with a masked sequence prediction objective: given the backbone structure (N, Cα, C, O atom coordinates), predict the amino acid identity at each masked position. ProteinMPNN achieves 52% sequence recovery on CAPI test structures (vs. 32% for Rosetta FastRelax), and experimentally validated sequences designed by ProteinMPNN fold to their target structures at high rates. The connection to our GNN guide is direct: ProteinMPNN’s structure-conditioned sequence design is a graph-to-sequence generation problem — the protein structure graph is processed by message-passing GNN layers, and the autoregressive sequence decoder generates amino acid identities conditioned on the structural context, the same framework as conditional graph generation described in that guide.
  • AlphaFold3 (Abramson et al., Nature 2024) extends AlphaFold2 to predict the structures of protein complexes with DNA, RNA, small molecules (ligands), and post-translational modifications — covering the full biomolecular interaction landscape. The key architectural change: replace the Evoformer + structure module with a Pairformer + diffusion module, where the diffusion module generates atomic coordinates by denoising (using the same DDPM framework from our diffusion models guide, applied to 3D coordinates rather than pixel grids). AlphaFold3 achieves state-of-the-art on protein-ligand docking (predicting how small molecule drugs bind to their target proteins) — a critical step in drug discovery — with accuracy competitive with physics-based docking tools (Vina, Glide) while running 100-1000× faster.

Proteins are the molecular machines of life: enzymes that catalyse metabolic reactions, structural proteins that give cells their shape, receptors that transduce signals across membranes, and antibodies that recognise pathogens. Understanding how protein sequence determines protein structure, and how structure determines function, is fundamental to biology and medicine. For 50 years, the protein structure prediction problem resisted computational solution — until AlphaFold2 achieved near-experimental accuracy in 2021, triggering a revolution in structural biology and drug discovery. This guide covers the full structural biology ML stack: AlphaFold2’s Evoformer architecture and structure module, ESMFold and protein language models, inverse folding with ProteinMPNN, AlphaFold3 for biomolecular complexes, protein fitness landscape prediction, and the connection to drug discovery pipelines.

AlphaFold2 Architecture, Evoformer and ESMFold

AlphaFold2’s pipeline begins with sequence homology search: given a query protein sequence, search databases (UniRef90, BFD, Mgnify) for homologous sequences using jackhmmer and HHblits to build a multiple sequence alignment (MSA). The MSA captures evolutionary co-variation: pairs of residues that have co-evolved across millions of years of evolution are likely in spatial contact in the folded structure, providing a rich signal for contact map prediction. The MSA representation (shape: n_seq × n_residue × d_model) is processed jointly with the pair representation (shape: n_residue × n_residue × d_model) through 48 Evoformer blocks. Each Evoformer block applies: (1) MSA row-wise attention (each sequence row attends to others, biased by the pair representation); (2) MSA column-wise attention (each residue position attends to its alignments across sequences); (3) pair triangular self-attention (the triangle multiplicative update and attention enforce geometric consistency in the pairwise distance map); (4) outer product mean (updating the pair representation from the MSA representation). The structure module takes the per-residue single representation (the final row of the MSA representation) and the pair representation as input, and iteratively refines the backbone frames (N, Cα, C atom positions as rigid bodies in SE(3)) and torsion angles (side-chain conformations) using Invariant Point Attention (IPA) — a form of attention that is invariant to global rotations and translations of the protein structure, ensuring that the predicted structure does not depend on its global orientation in space. The connection to our geometric deep learning guide is direct: IPA is an SE(3)-equivariant attention mechanism — the same class of equivariant neural networks described in that guide for molecular force fields and point cloud processing — applied to protein backbone frames as the equivariant objects. The AlphaFold Protein Structure Database (AlphaFold DB) contains predicted structures for 200 million proteins from UniProt, covering essentially all known protein sequences — a resource that has transformed structural biology by providing structural context for proteins that have never been crystallised experimentally. ESMFold achieves AlphaFold2-competitive accuracy on common proteins without MSA computation by substituting the evolutionary information from MSA with the statistical information learned by ESM-2 from 250 million protein sequences. The ESMFold architecture: ESM-2 (650M or 15B parameters) generates per-residue and pairwise representations from the input sequence, which are then processed by a lightweight folding trunk (8 structure module blocks) to generate 3D coordinates. ESM-2’s training on the evolutionary information implicit in sequence databases allows it to capture residue co-evolution without explicit alignment. The connection to our self-supervised learning guide is direct: ESM-2 is trained with masked language modelling on protein sequences — the same BERT-style masked prediction objective described there for self-supervised representation learning — applied to amino acid sequences rather than natural language tokens.

ModelArchitectureMSA Required?ScopeKey Benchmark
AlphaFold2Evoformer + structure moduleYesSingle protein chainsCASP14: TM-score 0.92
ESMFoldESM-2 + folding trunkNoSingle chains (fast)TM-score 0.90 on CAMEO
RoseTTAFold3-track network (1D/2D/3D)YesSingle chains + complexesCASP14 runner-up
AlphaFold3Pairformer + diffusionYesProtein-DNA/RNA/ligand complexesCASP15 protein-ligand SOTA
ProteinMPNNMessage-passing GNNNo (structure input)Inverse folding (sequence design)52% sequence recovery

Inverse Folding, AlphaFold3 and Fitness Landscape Modelling

ProteinMPNN’s success at inverse folding has enabled a new paradigm for protein engineering: computational design of protein sequences with desired properties, validated by experimental testing of a small set of computationally predicted top candidates. The workflow: (1) specify a target backbone structure (from X-ray crystallography, cryo-EM, or AlphaFold2 prediction); (2) use ProteinMPNN to generate diverse sequences predicted to fold to that backbone; (3) rank sequences by predicted stability (AlphaFold2 pLDDT score, ESM-2 log-likelihood) and desired functional properties; (4) synthesise and experimentally test the top 10-100 candidates. This computational design cycle has been used to design thermostable enzymes, de novo protein-protein interfaces for therapeutic biologics, and symmetric protein assemblies (cages, filaments) for vaccine delivery applications. RFdiffusion (Watson et al., Nature 2023) uses the DDPM diffusion framework to generate novel protein backbones satisfying functional constraints: given a binding site (the pocket shape and chemical environment required to bind a target molecule), RFdiffusion generates a protein backbone that presents the binding site in the correct geometry, which ProteinMPNN then threads with a sequence. This backbone diffusion + sequence design pipeline has designed protein binders for clinical targets including IL-7Rα and influenza haemagglutinin with nanomolar binding affinity. The connection to our diffusion models guide is direct: RFdiffusion applies DDPM to 3D protein backbone coordinates (Cα positions and orientations as rigid bodies) rather than pixel grids, using the same forward noising (adding Gaussian noise to coordinates) and reverse denoising (a structure module backbone predicts the clean structure from the noisy input) framework — with the key modification that the denoising is performed in SE(3) (the rotation-translation group) rather than Euclidean space. Protein fitness landscape modelling predicts the effect of amino acid mutations on protein function (binding affinity, thermostability, enzyme activity) from sequence alone, without requiring structure or MSA. ESM-2 log-likelihood ratios between the mutant and wild-type sequences achieve strong zero-shot fitness prediction (ΔΔG for thermostability, ΔΔG_binding for protein-protein interactions) — the implicit structure and evolutionary information in ESM-2’s parameters provides a fitness landscape proxy. ProteinGym (Notin et al., NeurIPS 2023) is the standard benchmark: 217 deep mutational scanning datasets covering point mutations across diverse protein families, with ESM-1v, EVmutation, and GEMME providing strong zero-shot baselines. The connection to our drug discovery guide is direct: fitness landscape models are used for antibody maturation (improving binding affinity of therapeutic antibodies by predicting beneficial mutations), enzyme engineering (improving thermostability and activity of industrial enzymes), and antigen drift modelling (predicting which viral mutations escape immune recognition for vaccine design).

✦ SUMMARIZE THIS ARTICLE WITH AI

ML for drug discovery (AlphaFold3 for protein-ligand docking; antibody engineering with fitness models) is in our Drug Discovery guide. Geometric deep learning (IPA as SE(3)-equivariant attention; protein structure as a geometric object) is in our Geometric Deep Learning guide. Graph Neural Networks (ProteinMPNN message passing for inverse folding) are in our GNN guide. Diffusion models (RFdiffusion for backbone generation; AlphaFold3 diffusion module) are in our Diffusion Models guide. Self-supervised learning (ESM-2 masked language modelling on protein sequences) is in our Self-Supervised Learning guide. Foundation models (ESM-2 as a protein domain foundation model) are in our Foundation Models guide. ML for materials science (crystal structure prediction sharing methodology with protein structure prediction) is in our Materials Science guide.

Leave feedback about this

  • Rating

Durgesh Kekare
Durgesh Kekarehttps://www.dataexpertise.in
Durgesh Kekare is a data science educator and founder of DataExpertise.in. With expertise in Python, machine learning, and analytics, he helps 10,000+ learners break into data careers.

Latest Posts

List of Categories