Starting from function, not sequence — ProteinMPNN sequence design, RFdiffusion backbone hallucination, and ProteinGym functional fitness landscapes for engineering novel proteins
RFdiffusion (Watson et al., Nature 2023) represents the most powerful tool for de novo protein design ever created. It uses a denoising diffusion model trained on the entire PDB to generate protein backbone structures — Cα positions and backbone orientations — satisfying any set of geometric constraints you specify. Whether you want a binder for a specific viral surface, an enzyme with a catalytic triad in a precise geometry, or a membrane pore with a 10Å diameter, RFdiffusion can generate topologically novel backbones that natural evolution never explored.
RFdiffusion builds on RoseTTAFold's (Baek et al. 2021) SE(3)-equivariant transformer architecture, repurposed as a denoising diffusion model:
Network: SE(3)-equivariant transformer (same architecture as RoseTTAFold) • Inputs: noisy backbone coordinates + time embedding + constraint features • Outputs: denoised backbone coordinates (Cα, N, C, O per residue) at step t-1 • Training: forward diffusion adds Gaussian noise to PDB coordinates → network learns to reverse
Forward process (training): • x_0 = true protein backbone coordinates • x_t = x_0 + N(0, σ_t²·I) with σ_t following VP-SDE schedule • Training: minimize ||x_̂_0 - x_0||² where x_̂_0 is network prediction from noisy x_t
Inference (generating new backbones): • Start: x_T ~ N(0, I) (pure Gaussian noise, size L×3 for L residues) • Iterative denoising: x_T → x_{T-1} → ... → x_0 (200 steps) • Result: physically realistic protein backbone geometry (correct bond lengths, angles)
Constraint injection types:
1. Partial diffusion (motif scaffolding): • Input: fixed motif coordinates (e.g., catalytic triad, epitope residues) + noisy scaffold • Only scaffold residues noised; motif residues held fixed • Network fills in scaffold around the fixed motif • Applications: enzyme design, binder design to specific epitope, PPI stabilization
2. Symmetric design: • Symmetry constraints injected as equivariance requirement • Generates Cn (cyclic), Dn (dihedral), icosahedral assemblies • Used for: vaccine nanoparticles, antigen-displaying scaffolds • David Baker group: 60-subunit icosahedral nanoparticle with RSV F protein displayed
3. Pore/channel design: • Specific geometric constraint: central axis void with minimum 8Å radius • Generates transmembrane beta-barrel or helix bundles • Applications: synthetic ion channels, drug delivery pores
4. Binder design: • Target protein surface provided as fixed atomic coordinates • Hotspot residues specified: designed protein must contact these positions • RFdiffusion generates binder backbones that cradle the hotspot surface
Post-processing: • Rosetta fast relax: energy minimization of generated backbone • Filter: Rosetta total_score per residue <-2.5 (stable packing) • AF2 pLDDT filter: >70 overall (confident structure) • RMSD filter: AF2 predicted structure vs. RFdiffusion target <2Å
ProteinMPNN (Dauparas et al., Science 2022) solves the "inverse folding" problem: given a 3D protein backbone, what amino acid sequence will fold to adopt that structure? Using a message-passing graph neural network that encodes the geometric relationships between residues, ProteinMPNN achieves 52% native sequence recovery on held-out proteins — a dramatic improvement over Rosetta's 32%. More importantly, its designed sequences actually fold to the target structure with much higher reliability than classical approaches.
ProteinMPNN models the conditional probability P(sequence | backbone structure):
Graph construction: • Node i = residue i: features include backbone angles (φ,ψ,ω) + virtual Cβ position • Edge (i,j): activated if residue j is among K=48 nearest neighbors of residue i • Edge features: inter-residue distance + direction vector (unit vector Cα_i → Cα_j) • Sequence masking: target residues masked; context residues unmasked
Message passing encoder (3 layers): • Each round: aggregate messages from all neighbors → update node hidden state • h_i^(l+1) = MLP(h_i^(l), Σ_j message(h_j^(l), e_ij)) • Encoder is bidirectional: all residues can communicate with all others
Autoregressive decoder (3 layers): • Generates sequence one residue at a time (random order) • When decoding residue i: sees backbone of ALL residues but sequence of only already-decoded residues • Captures context: hydrophobic core residue is adjacent to 8 aliphatic residues → tends toward Leu/Val/Ile • log P(aa_i | backbone, already_decoded) → sample aa_i from categorical distribution
Sampling temperature: • T=0.1 (greedy-like): high native recovery (~52%); less diversity; best for conservative redesigns • T=0.3: balanced; ~42% recovery; recommended for most applications • T=0.5: diverse; ~35% recovery; good for exploring sequence space • T=1.0: nearly random; ~25% recovery; for generating diverse libraries
Fixed residues: • Catalytic residues: forced to specific amino acid (e.g., His57, Asp102, Ser195 for serine protease) • Disulfide positions: Cys forced, partner identified by distance • Binding hotspot: positions contacting known ligand fixed to native • Rest designed freely: scaffold surface + hydrophobic core
Chain-specific design: • Multi-chain: chains A+B designed simultaneously; chain C ligand fixed • Symmetry: chains constrained to same sequence (homo-oligomers) • Tied positions: two residues forced to same amino acid (symmetric interfaces)
Hallmark result (David Baker lab validation): • Designed de novo mini-proteins (40–60 aa) expressed in E.coli → Tm >95°C • Natural Evolution took billions of years; ProteinMPNN produces comparable sequences in minutes
The central validation challenge in computational protein design is the "two-way problem": you start from a target structure, design a sequence with ProteinMPNN, but the sequence might fold to a completely different conformation in reality. Self-consistency testing with AlphaFold2 provides a stringent computational filter before expensive experimental validation: if AF2 independently predicts the same structure from the designed sequence (with no template information), the design has passed the most sensitive folding test available.
Self-consistency filtering pipeline:
Step 1 — ProteinMPNN generates N sequences for each backbone: • Typically: N=8–20 sequences per backbone; T=0.1 or T=0.3 • Sequences collected into FASTA file; labeled with backbone ID and sequence index
Step 2 — AlphaFold2 prediction WITHOUT templates: • Critical: template database search disabled (--max_template_date 1900 or template mode off) • Without templates, AF2 must rely entirely on MSA evolutionary information AND its internal physical model • For de novo-designed sequences: no MSA matches (novel sequence) → AF2 uses only pairwise coevolution • This is the strongest test: AF2 "knows" nothing about our design process
Step 3 — Evaluation metrics:
a) Cα RMSD (vs. RFdiffusion target): • Align AF2 predicted backbone to RFdiffusion target • RMSD calculated over all Cα atoms • Pass criteria: RMSD <1.5Å for most designs; <1.0Å for high-precision functional sites • RMSD 1–2Å: similar fold but some drift in loops → usually acceptable if active site unchanged • RMSD >3Å: sequence does not encode target fold → reject
b) AF2 pLDDT: • Per-residue confidence: mean over all residues >80 → confident prediction • Low pLDDT in designed loop: may be flexible → problematic if loop is functional • pLDDT in functional site: must be >90
c) Sccos (side-chain consistency score): • For each residue: are the designed side chains compatible with AF2-predicted backbone? • packing score and burial consistent with hydrophobic/hydrophilic requirement
d) TM-score: • TM-score >0.8 (vs. target): essentially same fold; best filter
Step 4 — Selection and diversity: • From 10,000 sequences: typically 200–1,000 pass RMSD + pLDDT filters • Cluster by sequence identity (MMSeqs2 at 30% identity): select representative from each cluster • Ensure diversity: don't select 200 nearly identical sequences — cover sequence space broadly • Final selection: 48–96 designs (typical 96-well plate format) for experimental validation
Interpretation nuances: • High RMSD + high pLDDT: AF2 confident but predicts different fold → sequence recapitulates a natural fold similar to an existing protein → useful if that fold is acceptable; reject if need precise geometry • Low RMSD + low pLDDT: AF2 predicts correct fold but not confident → may still work experimentally; uncertain • Ideal: RMSD <1.0Å + pLDDT >90 → strongest prediction of experimental success
Before committing to expensive expression and purification of 200 computational designs, deep mutational scanning (DMS) fitness models allow us to computationally predict which designs are tolerant to mutations (indicating a stable, well-folded core) and which are fragile. ProteinGym aggregates 250+ DMS datasets from experimental studies, training zero-shot fitness predictors that can evaluate single or combinatorial mutations in seconds — reducing experimental burden by 10-fold.
Computational fitness prediction (zero-shot models):
1. ESM-1v (Meier et al. 2021): • Masked language model (protein BERT) trained on UniRef90 (98M sequences) • Fitness prediction: P(wildtype) / P(mutant) → log likelihood ratio • No fine-tuning required: "zero-shot" for any protein • Accuracy: Spearman ρ ≈ 0.44 across 41 DMS datasets • Residue masked → probability of each amino acid at that position • Mutant effect = log P(wt) - log P(mut) at masked position
2. EVE (Evolutionary Variational autoEncoder, Frazer et al. 2021): • VAE trained on MSA of homologous sequences for each protein individually • Latent space encodes evolutionary constraints → fitness proxy • ELBO (evidence lower bound) for mutant sequence → lower = deleterious • More accurate than ESM-1v (Spearman 0.52) but requires homologs (not applicable to de novo proteins)
3. ProteinGym benchmark: • 250 DMS experiments across diverse protein families, therapeutic targets, viral proteins • Evaluates: single mutant effects, multiple-mutant combinations • Leaderboard: ESM-1v, EVE, TranceptEVE, Tranception, PoET, ProteinMPNN score • Best performer (2024): Tranception + EVE ensemble (Spearman 0.55)
Experimental DMS protocol: 1. Library construction: • MAGE (multiplex automated genome engineering) or DMS library synthesis • Saturating mutagenesis: every position mutated to all 19 other amino acids • Library size: L × 19 variants (for 100 aa protein: 1,900 variants) • Oligo pooling: >99% coverage by deep DNA sequencing verification
2. Functional selection: • Binding: yeast surface display → FACS sorting with fluorescent antigen → sequencing • Enzyme: growth complementation assay → surviving cells have active enzyme • Stability: protease treatment → stable variants survive → deep sequencing input/output
3. Data analysis: • Enrichment ratio: (count_post / count_pre) × normalization • Fitness score: log2(enrichment) → >0 = beneficial; <0 = deleterious; ≈0 = neutral • Correlation with computational predictions: validates/benchmarks zero-shot models
Application to designed proteins: • Designed sequences lack natural MSA → EVE inapplicable for de novo proteins • ESM-1v applicable: uses trained model's priors, does not need homologs • ProteinMPNN log-probability: directly measures "how likely is this sequence given this backbone" • Low ProteinMPNN log-prob at position → designed amino acid unusual → may be destabilizing • Positions with log-prob <-2: flag for experimental alanine scan → confirm necessity
The final and most sobering step of inverse protein design is experimental validation. Despite enormous computational advances, only 10–40% of designs typically express as soluble, folded proteins — a success rate 5–10× higher than classical Rosetta methods, but still requiring hundreds of designs to yield a handful of successes. Each experimental success is a triumph: a completely novel protein sequence with no natural counterpart, designed to a target geometry and achieving the predicted binding or catalytic function.
Experimental validation pipeline for designed proteins:
1. Gene synthesis and expression: • Codon optimization for E. coli (IDT, GenScript): ensures >65% codon adaptation index (CAI) • Expression vector: pET-28a (His6 tag, T7 promoter) or pGEX-4T-1 (GST fusion) • BL21(DE3) expression: IPTG at OD600=0.6; 18°C overnight for soluble expression • SDS-PAGE: check expression level (>1mg/L is good); compare soluble vs. total fraction • Ni-IMAC purification: 5-column volumes binding → 250mM imidazole elution • Size-exclusion chromatography (SEC): monomeric peak → correctly folded; aggregated → discard
2. Biophysical characterization:
a) Circular Dichroism (CD) spectroscopy: • α-helical protein: minima at 208nm and 222nm • β-sheet: minimum at 216nm, maximum at 195nm • Random coil: minimum at 200nm (misfolded) • Designed helix bundle: verify αCD matches computational prediction • Thermal denaturation: heat 20→95°C → Tm (midpoint); Tm >60°C = thermostable
b) SEC-MALS (Multi-Angle Light Scattering): • Absolute molecular weight determination (independent of column calibration) • Confirms oligomeric state: monomer vs. dimer vs. higher order • Designed 1:1 heterodimer: MALS should show MW = MW_A + MW_B
c) NMR (for small proteins <20 kDa): • 1H-15N HSQC: one peak per NH in folded protein → dispersion confirms folded core • Random coil proteins: peaks clustered at 8.2 ppm narrow region
3. Binding/functional assays:
SPR (Surface Plasmon Resonance) for binders: • Immobilize antigen/target on chip (CM5, amine coupling) • Inject designed protein at 5–7 concentrations • Measure on-rate (ka), off-rate (kd) → Kd = kd/ka • Designed binder: Kd target <100nM for therapeutic use • Best of 2023: de novo binder to RSV F protein, Kd=3.2nM, Tm=78°C, crystal structure confirmed 1.4Å RMSD
Enzyme activity: • Michaelis-Menten kinetics: Vmax, Km, kcat • Designed Kemp eliminase: >1,000-fold above background (Baker lab, Science 2010 classical) • Modern: designed luciferase (Baker lab 2023) — reaction rate 10× natural luciferase
4. Crystal structure determination: • Successful designs submitted for X-ray crystallography (APS/Diamond/ESRF) • Solve structure: molecular replacement with RFdiffusion target as search model • Experimental RMSD vs. designed structure: <1.5Å = design prediction confirmed • Reveals unforeseen contacts, solvent channels, ligand-binding mode
Practical success rates from >20 published papers: • de novo binding proteins: 1 in 10 expressed and folded designs achieves target Kd • Enzyme active site design: 1 in 50–100 shows detectable activity • Dramatic improvement vs. Rosetta-only era: then 1 in 1000–10000
In 2023, David Baker's laboratory published the design of a de novo protein that binds the botulinum neurotoxin receptor-binding domain with Kd=4.6 nM — with no evolutionary relationship to natural inhibitors. This demonstrates that inverse protein design can now generate therapeutic protein leads entirely from first principles. The protein was expressed in E.coli, correctly folded (Tm=73°C), confirmed by X-ray crystallography (0.9Å RMSD vs. design), and showed protective efficacy in mouse botulism models.