Computational design of novel mini protein binders using the RFdiffusion → ProteinMPNN → AlphaFold pipeline
Design de novo mini protein binders (~70 residues) targeting the PD-1 binding surface of PD-L1, using modern generative protein design tools. PD-L1 is a key immune checkpoint target in oncology — blocking its interaction with PD-1 restores T cell anti-tumor activity. This project demonstrates end-to-end computational protein design: from target structure analysis through backbone generation, sequence design, and in silico validation.
Computational protein engineering draws from three complementary methodological domains. Each captures different information about proteins — evolutionary history, physical constraints, or generative capacity — and is suited to different design questions. Knowing which tool to reach for, and why, matters more than knowing how to run any single tool.
Creating new proteins that don't exist in nature. Includes diffusion models (RFdiffusion — generate novel backbones from noise), inverse folding (ProteinMPNN — design sequences for a given backbone), and language model-guided generation. These methods explore regions of sequence and structure space that evolution never visited. This project sits primarily here, using RFdiffusion for backbone generation and ProteinMPNN for sequence design, with structure-based validation (ESMFold) and sequence-based analysis (per-position entropy).
The strategic question is always: what do I know, and what do I need? The starting point determines the method.
Every experimental screening campaign faces the same constraint: a finite budget buys a finite number of variants, and each well, droplet, or display cycle costs real money. The cost per variant is aggressively minimized during assay development, but there is always a ceiling. The strategic question has remained the same for 30+ years: given the available spots, how do I pack them with the best guesses for the most optimal candidates, so I'm not spending good money on dead wells?
In the absence of a good prediction, you go with randomness — error-prone PCR, NNK saturation, random mutagenesis — and let the screen do the work. This is a defensible strategy when you truly have no prior information, but it means the vast majority of your screening capacity is consumed by variants that were never going to work. The question has always been: can we do better than random? That is where AI-based approaches now sit. The tools are new — RFdiffusion, ProteinMPNN, protein language models — but the framework they serve is not. They are the current generation's answer to the same question that drove rational design, consensus sequence analysis, and structure-guided mutagenesis before them. What changes is the quality of the prediction. What stays the same is the goal: enrich the library for viable candidates so that experimental throughput is spent on sequences that have a reasonable chance of working.
This project's three-phase structure reflects that logic. Each phase filters candidates more aggressively, at increasing cost per candidate. The goal of upstream computation is to reduce the number of candidates that enter expensive downstream experimental work.
This project covers Phase 1 (RFdiffusion → ProteinMPNN → ESMFold) with elements of Phase 2 (structural validation, per-position entropy analysis, developability considerations). Phase 3 — gene synthesis, display-based screening, biophysical characterization — is where computational designs meet experimental reality. Every candidate killed computationally in Phases 1–2 is a candidate you don't synthesize, express, and screen.
Click each stage to see details about the tools, inputs, outputs, and decision points.
Merge all metrics (backbone geometry, ProteinMPNN scores, ESMFold pLDDT, RMSD to design) into a unified ranking. Apply dual threshold: pLDDT > 80 (confident fold) AND RMSD < 2.0 Å (matches designed backbone). 4 of 10 candidates passed both structural filters. These are not confirmed binders — they are the subset whose designed sequences are predicted to adopt folds close to the intended backbones.
Interactive view of PD-L1 target (PDB 4ZQK) with the PD-1 binding interface highlighted. Rotate by dragging, zoom with scroll. Toggle views to explore the binding surface your designed binders target.
Designs passing both structural filters (pLDDT > 80, RMSD < 2.0 Å) are highlighted. These are not confirmed binders — they are the subset whose designed sequences are predicted to adopt folds close to the intended backbones. ProteinMPNN score reflects inverse-folding confidence; higher means the model is more certain the sequence matches the backbone.
| Design | pLDDT | RMSD (Å) | MPNN score | Status |
|---|
The quadrant plot below separates in silico pass candidates (high confidence, low deviation) from failures. Designs in the upper-left quadrant pass both structural filters.
Does higher ProteinMPNN confidence predict better structural recapitulation? In this small sample, the correlation is weak — high MPNN scores don't guarantee low RMSD, reinforcing that structure-prediction validation is essential and MPNN score alone is insufficient for filtering.
PD-L1 was chosen because high-resolution crystal structures exist (PDB 4ZQK, 2.45 Å), the PD-1/PD-L1 binding interface is extensively characterized in the literature, the Baker lab has published de novo binder designs against this exact target using the same tools, and it carries immediate therapeutic relevance — enabling direct benchmarking against known results.
RFdiffusion was not constrained to any particular topology. Given a 70-residue length and the flat PD-L1 binding surface, the model independently converged on three-helix bundles — consistent with published results showing that helical scaffolds are thermodynamically favorable at this size and present good complementary surfaces for flat protein-protein interfaces.
The RFdiffusion → ProteinMPNN → AlphaFold pipeline optimizes for two objectives: folding and binding geometry. It says nothing about expression yield, aggregation propensity, immunogenicity, or manufacturing feasibility. In a real drug development campaign, in silico pass candidates would enter a developability assessment before any experimental work — scoring for surface hydrophobicity, charge distribution, deamidation motifs, and unpaired cysteines. The computational pipeline constrains the search space; it does not replace experimental validation.
One of the most practically valuable outputs of ProteinMPNN is not the top-scoring sequence — it's the per-position probability distribution across all 20 amino acids. When you sample 8 sequences from a single backbone at low temperature (T=0.1), positions where the model is confident produce the same amino acid across all samples (low Shannon entropy). Positions where the model is uncertain sample freely across multiple identities (high entropy). This distinction maps directly onto a fundamental choice in experimental protein engineering: what to fix and what to randomize.
In a conventional directed evolution campaign starting from scratch — using NNK codons, error-prone PCR, or random mutagenesis — every position is treated as equally uncertain. For a 70-residue mini binder, this produces a theoretical sequence space of 2070 variants. In practice, the vast majority of this space is dead: sequences that don't fold, don't express, or don't engage the target. The chance of sampling a functional variant is vanishingly small, and most screening throughput is wasted on sequences of no interest.
An AI-informed library compresses this space by fixing the positions ProteinMPNN is confident about — the residues where backbone geometry, packing contacts, and interface interactions strongly constrain amino acid identity — and applying combinatorial diversity only at the tolerant positions. The result is a library enriched for sequences that can actually fold and bind, while remaining large enough for robust screening.
AI-informed vs random library — screening funnel
Per-position entropy analysis across the 4 in silico pass designs in this project (computed from 8 ProteinMPNN samples per backbone at T=0.1) shows that 16–34 positions are effectively fixed, while 36–54 positions are tolerant — compressing theoretical diversity from 2070 to 2036–2054 depending on the scaffold. This represents a reduction of 16–34 orders of magnitude in sequence space, with the remaining diversity concentrated in positions the model identifies as computationally permissive.
In practice, this framework integrates AI-guided design with existing high-throughput experimental infrastructure. Fixed positions inform a defined sequence core; tolerant positions are degenerated using NNK codons or tailored degenerate codons (e.g. NDT, VHG) that enrich for the specific subset of amino acids ProteinMPNN identifies as compatible at each site. The resulting gene library feeds directly into display technologies — phage, yeast, or mRNA display — for ultra-high-throughput screening at 106–108 variants, followed by plate-based and biophysical screens at progressively higher resolution. The key advantage over naive randomization is not just a smaller library — it is a library where a much higher fraction of members are variants of interest rather than dead sequences sampling irrelevant space.
This approach also connects naturally to developability. Positions identified as tolerant by ProteinMPNN are often solvent-exposed, meaning substitutions there are less likely to disrupt folding but may influence expression, aggregation, or immunogenicity. Rational substitutions at tolerant sites — replacing methionines to reduce oxidation liability, or introducing charged residues to improve solubility — can be incorporated into library design without sacrificing the AI-informed scaffold constraints. The result is a campaign that is simultaneously AI-guided and experimentally tractable.
This analysis evaluates monomer fold recapitulation only — whether a designed sequence is predicted to adopt its intended backbone structure. It does not evaluate binding affinity, interface stability, complex formation with PD-L1, expression yield, solubility, aggregation propensity, specificity, or immunogenicity. ESMFold was used for binder-only structure prediction, not for binder–target complex modeling. Passing the in silico structural filters (pLDDT > 80, RMSD < 2.0 Å) means the sequence is predicted to fold as designed — it does not mean the resulting protein will bind PD-L1 or function as a therapeutic.
Experimental validation would require gene synthesis, recombinant expression, binding assays (SPR, BLI, or ELISA against PD-L1), and biophysical characterization (SEC, DLS, thermal stability). In published pipelines using similar tools, 10–30% of computationally passing designs show measurable binding in the lab. The 4 candidates identified here represent a starting point for experimental testing, not an endpoint.
Molecular dynamics validation of top candidates using OpenMM (fold stability at 300K, interface contact persistence). Disulfide bridge engineering as a discriminating test — if a marginally stable design unravels even with a covalent crosslink, the fold was never a real energy minimum. VHH scaffold-constrained design against the same target as a second project demonstrating range across design paradigms.