De Novo Protein Binder Design Pipeline

Overview

This pipeline designs novel protein binders for a target protein structure through six phases:

1. Target Preparation -- Clean and validate the input PDB
2. Epitope Selection -- Identify binding sites via PeSTo + DBSCAN clustering
3. Binder Length Optimization -- Calculate optimal binder size from epitope geometry
4. Backbone Generation -- Generate candidate scaffolds with RFdiffusion
5. Sequence Design -- Design sequences via inverse folding (ProteinMPNN / ESM-IF1)
6. Validation -- Multi-metric screening with early termination
7. Round Advancement -- Iterate or finalize based on results

Interaction Model

Phases 0--1.5 (preparation): Run autonomously -- clean the target, select epitopes, determine binder length. No user approval needed.

Before Phase 2 (backbone generation): Present a design plan to the user summarizing: target info, selected epitope and hotspot residues, proposed binder length, number of backbones, and any concerns (large target, polar epitope, etc.). Wait for user approval before proceeding.

Phases 2--5 (generation + validation): Run autonomously after approval. Report results at natural checkpoints (after each round).

Design Principles

• Fixed tool pipeline -- The tools and their order are prescribed (see each phase). Parameters and scale are flexible based on context.
• Early termination -- Run cheap checks (solubility) before expensive ones (structure prediction) to save compute.
• Iterative refinement -- Start with high-diversity sequence sampling and progressively lower temperature for convergence.

Phase 0: Target Preparation

Input: Raw PDB file (from database or user upload).

Steps

1. Parse PDB for metadata -- Extract experimental method, resolution, chain composition, sequences.
2. Clean structure -- Remove waters/heterogens, restore missing heavy atoms, add hydrogens at physiological pH, standardize non-canonical residues, detect disulfide bonds.
3. Select chains of interest -- Extract relevant chain(s). Can be user-specified or automatic.
4. Quality checks (warnings, not blockers):
   • Targets >2000 residues are computationally expensive downstream
   • Multi-chain targets add complexity to epitope selection

Output: Cleaned PDB with metadata (sequences, chain IDs, residue counts).

Phase 1: Epitope Selection

Purpose: Identify and rank candidate binding sites on the target surface.

Step 1: Binding Interface Prediction (PeSTo)

Run PeSTo (Protein Structure Transformer, model i_v4_1) to get per-residue protein-protein binding probabilities (0.0--1.0). Classify residues above a threshold (recommended: 0.3) as binding candidates.

Step 2: Glycosylation Filtering (Optional)

Run glycosylation prediction ensemble (EMNgly, LMNglyPred, ISOGlyP) to identify residues at risk of post-translational glycosylation. Penalize high-risk sites during scoring since glycans physically occlude the surface.

Step 3: DBSCAN Spatial Clustering

Cluster binding residues by CA coordinates using DBSCAN to group spatially proximal residues into discrete epitope regions. Discard unclustered noise residues.

Step 4: Hotspot Selection

Rank residues within each cluster by composite score that weighs:

• PeSTo binding probability (primary signal)
• Hydrophobic/aromatic character (favorable for binding interfaces)
• Glycosylation risk (penalty)

Select the best cluster and pick the top-scoring residues (recommended: 3--6) as hotspot residues for RFdiffusion.

Scoring weights and parameter defaults: references/parameters.md -- load when adjusting epitope selection behavior.

Phase 1.5: Binder Length Optimization

Runs only if the user did not specify a custom binder length.

Approach

1. Compute pairwise CA distances between all epitope residues.
2. Take the maximum distance as the epitope span.
3. Choose a binder length proportional to the span -- larger epitopes need longer binders to cover the interface. General guidelines:

Binder length should stay within 50--200 residues. When in doubt, err toward longer binders.

Phase 2: Backbone Generation (RFdiffusion)

Purpose: Generate diverse backbone structures complementary to the target epitope.

Inputs

• Cleaned target PDB from Phase 0
• Hotspot residues from Phase 1 (e.g., ["A42", "A45", "A78"])
• Binder length range from Phase 1.5 (e.g., "70-110")
• Number of designs (scale based on target difficulty -- typically 50--200 backbones)

Execution

• Each call produces backbone PDB files with backbone atoms (N, CA, C, O) for both binder (chain X) and target chains.
• Parallelization: Each RFdiffusion call should generate at most 10 designs. For larger counts, split into parallel calls (e.g., 100 designs = 10 parallel calls x 10 designs each). This is significantly faster than a single call with num_designs=100.
• Optional beta model variant available (use_beta_model, default: off).

Output

Independent backbone scaffolds, registered for downstream sequence design.

Design tips for hotspot selection, target truncation, and scale: references/design-tips.md -- load when planning a design campaign or troubleshooting RFdiffusion.

Phase 3: Sequence Design

Purpose: Design amino acid sequences predicted to fold into each backbone.

Temperature Schedule

Use exponential decay across sequence rounds to shift from exploration to convergence. Recommended starting point:

Adjust the decay rate and number of rounds based on how quickly valid designs emerge. If round 1 already produces many valid designs, fewer rounds are needed.

Tool Selection

ProteinMPNN (default): Graph neural network for inverse folding.

• model_type: "soluble"
• chains_to_design: "X" (binder only; target fixed)
• omit_AAs: "X" (exclude non-standard)
• Confidence: native 0--1 range (higher = better)

ESM-IF1 (alternative): Structure-conditioned language model.

• mode: "sample", multichain_backbone: True
• Confidence normalization: 1 / (1 + exp(-(log_likelihood + 2.0) * 1.5))

Sequence Filtering

1. Over-generate candidates (e.g., 10x the desired count), then keep only the top sequences by confidence score.
2. Deduplicate -- skip any sequence already generated in the current or prior rounds.

Full parameter tables: references/parameters.md

Phase 4: Validation

Purpose: Multi-metric validation with early termination for failing designs.

Validation Pipeline (ordered by cost)

Sequence --> Solubility --> Boltz-2 --> Metrics --> US-align --> DockQ --> Pass/Fail

Step 1: Solubility Screening

Fine-tuned ESM-2 classifier predicts E. coli expression solubility. Binary decision:

• Soluble: proceed to structure prediction
• Insoluble: early termination -- skip all remaining steps

Step 2: Structure Prediction (Boltz-2)

Predict binder-target complex structure from sequences.

• Binder chain X: single-sequence mode (no MSA)
• Target chains: MSAs auto-generated by Boltz-2

Step 3: Extract Confidence Metrics

From Boltz-2 output:

• pLDDT: Per-residue confidence averaged across full structure (0--100)
• iPTM: Interface predicted TM-score across chain interfaces (0--1)
• ipSAE: avg(d0res_min, d0res_max) for binder-target pairs only (0--1). Skipped if no interface detected.

Step 4: Structural Alignment (US-align)

Compare predicted binder (chain X) vs RFdiffusion backbone (chain X) in monomer mode.

• Output: TM-score (0--1) -- whether the sequence folds into the intended shape

Step 5: Interface Quality (DockQ)

Evaluate predicted complex vs RFdiffusion design. Binder-target pairs only (e.g., [["X", "A"]]).

• Output: DockQ score (0--1) plus iRMSD, LRMSD, fnat components

Step 6: Pass/Fail Decision

A binder passes only if ALL metrics meet their thresholds (AND logic). Check all metrics and report all failures, not just the first.

Recommended thresholds (balanced):

Adjust thresholds based on campaign goals -- use lenient thresholds for difficult targets to avoid rejecting everything, or strict thresholds when high confidence is required.

Lenient and strict threshold variants: references/parameters.md Detailed metric definitions and interpretation: references/validation-metrics.md -- load when interpreting results or adjusting thresholds.

Phase 5: Round Advancement

Purpose: After validating all designs in a round, decide whether to continue or finalize.

Decision Guidelines

After each validation round, decide the next action:

1. Finalize -- If enough valid designs have been found (target met), or if the user is satisfied with results so far.
2. Advance sequence round -- If the current backbone produced some promising results, try new sequences at a lower temperature to refine around successful folds. Same backbone, narrower sampling.
3. Advance backbone -- If the current backbone's sequence rounds are exhausted or producing diminishing returns, move to the next backbone and reset the temperature schedule.
4. Finalize with partial results -- If all backbones and rounds are exhausted, report whatever valid designs were found.

Use the valid design rate and per-metric failure patterns to guide the decision. If most failures are on the same metric (e.g., DockQ), the issue is likely the backbone geometry rather than sequence design -- advancing to a new backbone is more productive than more sequence rounds.

Reporting

• Round report: After each round -- iterations tested, valid found, failures by metric, temperature used, duplicates skipped.
• Final report: At completion -- comprehensive summary of entire design campaign.

Reference Files