Multi-chain structure prediction and molecular docking — predicting protein-ligand, protein-protein, and protein-nucleic acid complex structures with AI diffusion models
AlphaFold 3 (Abramson et al., Nature 2024) represents a paradigm shift from AlphaFold 2: instead of predicting only protein structures, it predicts the structures of arbitrary biomolecular complexes — protein-protein, protein-small molecule, protein-DNA/RNA, protein-ion — all within a unified architecture. The input pipeline handles heterogeneous molecular entities through a shared tokenization and embedding framework.
AlphaFold 3 key architectural changes vs. AlphaFold 2:
1. Unified tokenization: • Protein chains: each residue = one token (amino acid identity + backbone + sidechain) • Small molecules (ligands): each heavy atom = one token (element type + formal charge + chirality) • DNA/RNA: each nucleotide = one token • Ions: each ion = one token (element + charge) • All tokens share a single embedding space — enabling cross-molecular attention
2. Input features (per residue/atom): Protein residues: • One-hot amino acid identity (21 categories) • Relative sequence position encoding (Rotary embeddings, RoPE) • MSA features: conservation, sequence weight, deletion fractions for each alignment column • Template features: Cα distance maps, backbone dihedral angles from PDB templates Small molecule atoms: • SMILES-derived: element (one-hot), aromaticity, charge, chirality, bond types • RDKit 3D conformer as initial coordinate guess (or random noise for diffusion start) • Relative positions within molecule encoded in pair representation
3. MSA construction for multi-chain: • Unpaired MSA: each chain searched independently (UniRef90/Mgnify/JGI) • Paired MSA: chains searched together to identify sequences from same organism/complex • Pairing ratio (paired:unpaired) tuned: ~50:50 for protein-protein; 100% unpaired for protein-ligand • No MSA for small molecules (they use only SMILES + conformer coordinates)
4. Template search: • MMSeqs2 search against PDB70 → identifies similar structures • Up to 4 templates used per chain • Template features: distogram (Cα distances), torsion angle pairs, sequence similarity mask • Templates provide evolutionary structural constraints that guide diffusion
Key advantage over classical docking (AutoDock, Glide): • Classical docking requires a known receptor structure (apo or holo) • AF3 simultaneously predicts receptor structure AND docked pose • Especially powerful for novel targets with no experimental structures • Challenge: AF3 generates ensemble of conformations — selecting correct one for VS requires careful scoring
The Evoformer++ is the central processing engine of AlphaFold 3, transforming raw sequence and co-evolutionary information into a rich geometric representation of inter-residue relationships. Through iterative message-passing across a two-dimensional pairwise representation matrix, it implicitly learns the physical forces — hydrogen bonding, hydrophobic packing, electrostatics, steric exclusion — that determine protein structure, without ever encoding them explicitly.
The Evoformer++ processes two representations in parallel:
1. Single representation (s_i): per-token embedding (dim 384) • Captures individual residue/atom properties • Updated by row/column attention over pair representation • Used to gate pair updates and condition diffusion module
2. Pair representation (z_ij): N×N token-pair embedding (dim 128) • Captures directional relationship between every pair of tokens i and j • z_ij ≠ z_ji: asymmetric (contact propensity and distance distribution, not symmetric) • Updated by triangle multiplicative updates and triangle self-attention
Triangle multiplicative update (symmetric): • z_ij ← LayerNorm(z_ij + Σ_k a(z_ik) ⊙ b(z_kj)) where a(), b() = learned gated linear projections • Interpretation: for each residue pair (i,j), broadcast information through all intermediate residues k • Equivalent to learning implicit triangle inequalities in distance space • Closing (outgoing): i→k, j→k → update i→j • Opening (incoming): k→i, k→j → update i→j
Triangle self-attention: • Attention over rows (all pairs sharing residue i) or columns (all pairs sharing residue j) • Allows long-range information propagation without increasing receptive field linearly
Outer product mean update: • z_ij ← z_ij + outer_product_mean(s_i, s_j) • Fuses per-token information into pair representation at each block
Algorithmic difference from AF2: • AF3 adds cross-chain pair representation: pairs between chain A and chain B residues • Enables direct geometric reasoning about protein-protein interface • New: small molecule atom pairs included in pair matrix • New: DNA/RNA nucleotide pairs included • Memory challenge: N² pairs; AF3 handles ~3000-token sequences practically
AlphaFold 3 replaces the deterministic Structure Module of AF2 (which used invariant point attention to iteratively refine backbone frames) with a conditional denoising diffusion probabilistic model (DDPM) over all-atom 3D coordinates. This enables the model to generate diverse structural ensembles, handle ligand conformational sampling natively, and produce physically realistic all-atom geometries — a fundamental change that generalizes structure prediction to arbitrary molecular inputs.
DDPM score-matching framework applied to 3D coordinates:
Forward process (training time only): q(x_t | x_{t-1}) = N(x_t; √(1-β_t)x_{t-1}, β_t·I) • Gradually adds Gaussian noise to true atomic coordinates x_0 • After T=1000 steps: x_T ≈ N(0, I) (pure noise) • Variance schedule β_t: cosine schedule for biomolecular coordinates
Reverse process (inference): p_θ(x_{t-1} | x_t, c) = N(x_{t-1}; μ_θ(x_t, t, c), σ_t²·I) • Neural network learns to denoise x_t → x_{t-1} • c = conditioning signal (Evoformer++ single and pair representations) • At inference: start from x_T ~ N(0,I), denoise 200 steps → atomic coordinates
Conditioning mechanism: • Each denoising step: Evoformer++ representations queried via cross-attention • Time embedding t injected via AdaLN (Adaptive Layer Normalization) • Chain identity, bond connectivity, chirality constraints injected as hard constraints
Frames and equivariance: • Outputs are SE(3)-equivariant: rotating or translating input → corresponding rotation of output • Achieved via invariant features + application to local backbone frames • Ligand atoms: no backbone frame → directly diffused in global frame, constrained by bond geometry
Post-diffusion all-atom geometry optimization: • Bond lengths enforced by geometry relaxation (OpenMM Amber ff14SB + GAFF2 for ligand) • Clash removal: VDW clashes scored and filtered • Complete hydrogens added by RDKit + protonation states assigned by propka3
Performance on CASP15 + PoseBusters: • Protein monomer: GDT_TS 93.1 (comparable to AF2) • Protein-protein (dimer): DockQ 0.81 (vs. 0.72 for AF2-Multimer) • Protein-ligand (PoseBusters): 76% within 2Å RMSD (vs. 52% for RF-Diffusion+DiffDock) • Covalent ligands: 65% within 2Å (previously unsolvable with classical docking)
AlphaFold 3 does not just produce a structure — it produces a confidence map that tells you exactly which regions to trust for drug design decisions. The pLDDT score per residue is so well-calibrated that it has become a standard research tool: loops with pLDDT < 50 are intrinsically disordered, surfaces with pLDDT > 90 are structurally defined, and interfaces with ipTM > 0.75 are predicted with crystallographic accuracy. Understanding these scores is essential for translating computational predictions into experimental campaigns.
AlphaFold 3 outputs four complementary confidence metrics:
1. pLDDT (predicted Local Distance Difference Test): • Per-residue/per-atom confidence, 0–100 scale • Predicted by a separate "confidence head" (linear layer on single representation) • lDDT-Cα measures fraction of Cα-Cα distances within 0.5/1/2/4 Å of true structure • pLDDT interpretation: >90: high confidence — reliable for drug design, rigid in crystal structure 70–90: good — generally reliable backbone, some sidechain flexibility 50–70: low — likely flexible region, do not use for precise docking <50: very low — intrinsically disordered region (IDR), will not crystallize in this conformation • Color-coded in AlphaFold DB visualizer (dark blue/light blue/yellow/orange)
2. pTM (predicted Template Modeling score): • Single number (0–1) for whole chain or complex • Predicted TM-score: TM-score >0.5 → same fold; >0.7 → very similar; >0.8 → essentially identical • pTM >0.5: fold is correctly predicted; pTM >0.8: high confidence for experimental use • Chain-level: each chain gets individual pTM
3. ipTM (interface predicted TM-score): • Measures accuracy of predicted inter-chain contacts specifically • Critical for docking: ipTM > 0.75 → interface correctly predicted • Formula: TM-score restricted to inter-chain residue pairs within 24Å • Complex ranking: 0.8 × ipTM + 0.2 × pTM (down-weights individual chain accuracy vs. interface) • ipTM < 0.5: complex likely does not exist or is incorrect — do not use for docking
4. PAE (Predicted Aligned Error): • N×N matrix: PAE[i,j] = expected position error of residue j when residue i is aligned • Low PAE between two domains → rigid body relationship → useful for multi-domain modeling • Low PAE across chain boundary → confident interface prediction • High PAE within chain → intrinsically flexible linker between domains • Used for domain boundary detection and multi-body modeling
Practical workflow for drug design: 1. Screen pTM/ipTM of predicted complex → reject if ipTM <0.6 2. Map pLDDT onto binding site → regions <70 should be excluded from pharmacophore 3. PAE matrix → identify rigid binding pocket vs. flexible loops 4. Generate 5-seed ensemble → cluster by RMSD → select representative from largest cluster 5. Only use AF3-predicted pocket with crystallographic ligand for virtual screening if RMSD <2Å vs. known structure
AlphaFold 3 is not merely an academic tool — it is actively driving drug discovery campaigns across pharma. The ability to predict protein-ligand complexes for any target, including novel GPCRs, transcription factors, and undruggable proteins, has opened an entirely new era of structure-based drug design. When combined with FEP+ free energy perturbation calculations, AF3-predicted structures can serve as the starting point for lead optimization campaigns that previously required years of crystallography.
Complete AF3-driven drug discovery workflow:
Step 1 — Target identification and AF3 structure prediction: • Input: target protein sequence (from genome/transcriptome of disease tissue) • AF3 predicts: monomeric structure + all known binding partners (PPI) • Assess druggability: FTMap pocket analysis on AF3 structure → hotspot residues + pocket volume • Compare AF3 vs. ESMFold/RoseTTAFold predictions: consensus = high confidence
Step 2 — AF3-predicted binding site characterization: • Site Finder (MOE) / fpocket / DoGSiteScorer on AF3 structure • DrugScore/SiteHound maps druggable hotspot atoms • If no known ligand: probe-based consensus binding site prediction • Template-guided docking: if homologous structure has ligand → transplant to AF3 model
Step 3 — Virtual screening on AF3 structure: • Protein preparation: Maestro Protein Prep Wizard (H-bond optimization, protonation, minimization) • Grid generation: binding site from AF3 pocket analysis • Docking: Glide SP (10k compounds) → Glide XP (1k) → MMGBSA rescoring (100) • Challenge: AF3 structure may be apo-form; induced fit docking (IFD) used for flexible binding site
Step 4 — Fragment growing and scaffold merging: • AF3 structure + fragment screen (DSF or SPR): identifies fragment binding poses • AF3 re-predicted with fragment → confirms binding mode • Fragment growing: add vectors from fragment to fill predicted hydrophobic sub-pockets • Scaffold merging: combine two fragments using AF3-guided linker design
Step 5 — FEP+ lead optimization from AF3 starting point: • Set up FEP+ perturbation map: 20 compounds around lead scaffold • Minimize AF3 structure with ligand in OPLS4 force field → production MD • FEP ΔΔG calculated for each analog: accuracy ±0.8 kcal/mol (vs. ±1.5 for docking) • Synthesis prioritization: select analogs with FEP-predicted ΔΔG improvement >1 kcal/mol
Case studies of AF3-enabled drug discovery: • KRAS G12D (formerly undruggable): AF3 predicted cryptic pocket → covalent inhibitor designed • WDR domain proteins: AF3 homo-dimer interfaces → first WDR PPI inhibitors identified • Viral polymerases (novel pathogens): AF3 predicted structure within days of outbreak → emergency antivirals • TCR-pMHC complexes: AF3 ipTM enables T-cell receptor engineering for cancer immunotherapy • Antibody-antigen: AF3 models antibody:epitope → guides antibody engineering campaigns
The most transformative application of AlphaFold 3 is in tackling transcription factors — a class of 1,600 proteins responsible for driving cancer gene expression programs that were historically considered undruggable because they lack enzymatic cavities. Using AF3 to predict transcription factor-DNA-coactivator ternary complexes, researchers at multiple companies have now identified cryptic pockets that only appear in the DNA-bound state — expanding the druggable genome by an estimated 15–20%.