Tooluniverse Protein Therapeutic Design

Therapeutic Protein Designer

AI-guided de novo protein design using RFdiffusion backbone generation, ProteinMPNN sequence optimization, and structure validation for therapeutic protein development.

KEY PRINCIPLES: 1. Structure-first - Generate backbone geometry before sequence 2. Target-guided - Design binders with target structure in mind 3. Iterative validation - Predict structure to validate designs 4. Developability-aware - Consider aggregation, immunogenicity, expression 5. Evidence-graded - Grade designs by confidence metrics 6. Actionable output - Provide sequences ready for experimental testing 7. English-first queries - Always use English terms in tool calls

Therapeutic protein design starts with the target interaction. What binding surface do you need to cover? A small pocket = nanobody or peptide. A large flat surface = designed protein. Stability, immunogenicity, and manufacturability constrain the design space.

LOOK UP, DON'T GUESS

When uncertain about any scientific fact, SEARCH databases first rather than reasoning from memory. A database-verified answer is always more reliable than a guess.

---

COMPUTE, DON'T DESCRIBE

When analysis requires computation (statistics, data processing, scoring, enrichment), write and run Python code via Bash. Don't describe what you would do — execute it and report actual results. Use ToolUniverse tools to retrieve data, then Python (pandas, scipy, statsmodels, matplotlib) to analyze it.

When to Use

Apply when user asks to:

• Design a protein binder, therapeutic protein, or scaffold
• Optimize a protein sequence for function
• Design a de novo enzyme
• Generate protein variants for target binding

---

Workflow Overview

Phase 1: Target Characterization
  Get structure (PDB, EMDB cryo-EM, AlphaFold), identify binding epitope

Phase 2: Backbone Generation (RFdiffusion)
  Define constraints, generate >= 5 backbones, filter by geometry

Phase 3: Sequence Design (ProteinMPNN)
  Design >= 8 sequences per backbone, sample with temperature control

Phase 4: Structure Validation (ESMFold/AlphaFold2)
  Predict structure, compare to backbone, assess pLDDT/pTM

Phase 5: Developability Assessment
  Aggregation, pI, expression prediction

Phase 6: Report Synthesis
  Ranked candidates, FASTA, experimental recommendations

---

Critical Requirements

Report-First Approach (MANDATORY)

1. Create [TARGET]_protein_design_report.md first with section headers 2. Progressively update as designs are generated 3. Output [TARGET]_designed_sequences.fasta and [TARGET]_top_candidates.csv

Design Documentation (MANDATORY)

Every design MUST include: Sequence, Length, Target, Method, and Quality Metrics (pLDDT, pTM, MPNN score, binding prediction).

---

NVIDIA NIM Tools

Common Parameter Mistakes

NVIDIA NIM Requirements

• API Key: NVIDIA_API_KEY environment variable required
• Rate limits: 40 RPM (1.5 second minimum between calls)
• AlphaFold2 may return 202 (polling required); RFdiffusion and ESMFold are synchronous

---

Supporting Tools

---

Evidence Grading

---

Completeness Checklist

• [ ] Target structure obtained (PDB or predicted)
• [ ] Binding epitope identified
• [ ] >= 5 backbones generated, top 3-5 selected
• [ ] >= 8 sequences per backbone, MPNN scores reported
• [ ] All sequences validated (ESMFold), pLDDT/pTM reported, >= 3 passing
• [ ] Developability assessed (aggregation, pI, expression)
• [ ] Ranked candidate list, FASTA file, experimental recommendations

---

Reference Files

• DESIGN_PROCEDURES.md - Phase-by-phase code examples, sampling parameters, fallback chains
• TOOLS_REFERENCE.md - Complete tool documentation with code examples
• EXAMPLES.md - Sample design workflows and outputs
• CHECKLIST.md - Detailed phase checklists and quality metrics
• design_templates.md - Report templates and output format examples

Therapeutic Protein Designer Checklist

Pre-delivery verification checklist for protein design reports.

Report Quality Checklist

Structure & Format

• [ ] Report file created: [TARGET]_protein_design_report.md
• [ ] All 8 main sections present
• [ ] Executive summary completed (not [Designing...])
• [ ] Data sources section populated

Phase 1: Target Characterization

• [ ] Target protein identified with UniProt ID
• [ ] Target structure obtained (PDB ID or predicted)
• [ ] Structure source documented (PDB/AlphaFold/NvidiaNIM)
• [ ] Binding epitope defined (residue range)
• [ ] Existing binders/therapeutics noted
• [ ] Target function described

Phase 2: Backbone Generation

• [ ] NVIDIA_API_KEY availability confirmed
• [ ] Design mode specified (unconditional/binder/motif)
• [ ] Diffusion steps documented
• [ ] ≥5 backbones generated
• [ ] Backbone topologies described
• [ ] Quality assessment for each backbone
• [ ] Top 3-5 selected with rationale

Phase 3: Sequence Design

• [ ] ProteinMPNN parameters documented
• [ ] Temperature setting stated
• [ ] ≥8 sequences per backbone
• [ ] Total sequence count reported
• [ ] MPNN scores for all sequences
• [ ] Top 10 sequences ranked
• [ ] Full sequences provided (FASTA format)

Phase 4: Structure Validation

• [ ] All sequences validated by ESMFold/AlphaFold
• [ ] pLDDT reported for each design
• [ ] pTM reported for each design
• [ ] RMSD to designed backbone (if applicable)
• [ ] Pass/fail criteria applied (pLDDT >70, pTM >0.7)
• [ ] Regional confidence breakdown for top designs
• [ ] ≥3 designs pass validation

Phase 5: Developability Assessment

• [ ] Aggregation propensity scored
• [ ] Isoelectric point calculated
• [ ] Cysteine count and pairing noted
• [ ] Expression prediction (E. coli/mammalian)
• [ ] Hydrophobic patch analysis
• [ ] Overall developability rating

Phase 6: Final Deliverables

• [ ] Ranked candidate list with all metrics
• [ ] FASTA file with designed sequences
• [ ] CSV with candidate metrics
• [ ] Evidence tier assigned to each design

---

Citation Requirements

Every Design Must Include

• [ ] Design method (RFdiffusion → ProteinMPNN → ESMFold)
• [ ] Tool names in backticks
• [ ] All numerical metrics

Format Examples

*Design Method: RFdiffusion backbone → ProteinMPNN sequence → ESMFold validation*
*Source: NVIDIA NIM via `NvidiaNIM_rfdiffusion` (50 diffusion steps)*
*Source: NVIDIA NIM via `NvidiaNIM_proteinmpnn` (T=0.1, 8 sequences)*
*Source: NVIDIA NIM via `NvidiaNIM_esmfold` (pLDDT: 88.5, pTM: 0.85)*

---

Evidence Grading

All Designs Must Have

• [ ] Evidence tier assigned (★★★ to ☆☆☆)
• [ ] All quality metrics documented
• [ ] Developability assessment complete

Tier Definitions

---

Quantified Minimums

---

Design Quality Metrics

Structure Prediction Quality

ProteinMPNN Score

Developability

---

Output Files

Required

• [ ] [TARGET]_protein_design_report.md - Main report

Required Data Files

• [ ] [TARGET]_designed_sequences.fasta - All sequences
• [ ] [TARGET]_top_candidates.csv - Ranked candidates

FASTA Format

>Design_001 | pLDDT=88.5 | pTM=0.85 | MPNN=-1.89 | Tier=T1
MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSH
GSAQVKGHGKKVADALTNAVAHVDDMPNALSALSDLHAHKL
>Design_002 | pLDDT=82.3 | pTM=0.79 | MPNN=-1.95 | Tier=T2
...

CSV Column Requirements

Rank,Design_ID,Sequence,Length,Backbone,MPNN_Score,pLDDT,pTM,Aggregation,pI,Tier

---

Design-Specific Checks

For Binder Design

• [ ] Target structure used in design
• [ ] Epitope/hotspot residues specified
• [ ] Interface quality assessed
• [ ] Binding mode predicted

For Scaffold Design

• [ ] Topology specified (alpha/beta/mixed)
• [ ] Size constraints met
• [ ] Functional sites preserved

For Enzyme Design

• [ ] Active site geometry maintained
• [ ] Catalytic residues in place
• [ ] Substrate access verified

---

Experimental Recommendations

Must Include

• [ ] Expression system recommendation
• [ ] Purification strategy
• [ ] Characterization assays
• [ ] Expected challenges

Expression Systems

---

Final Review

Before Delivery

• [ ] No [Designing...] placeholders remaining
• [ ] All tables properly formatted
• [ ] Executive summary synthesizes key findings
• [ ] Top 3 candidates clearly identified
• [ ] Sequences ready for ordering
• [ ] Experimental plan outlined

Common Issues to Avoid

• [ ] Missing pLDDT/pTM values
• [ ] Sequences without MPNN scores
• [ ] Incomplete developability assessment
• [ ] Missing FASTA file
• [ ] Unclear ranking criteria
• [ ] No experimental recommendations

---

NVIDIA NIM Tool Checklist

Pre-Run Verification

• [ ] NVIDIA_API_KEY environment variable set
• [ ] Rate limit awareness (40 RPM)
• [ ] Async handling for AlphaFold2

Post-Run Validation

• [ ] All tool calls succeeded
• [ ] Reasonable output values
• [ ] No error responses in data
• [ ] Fallbacks used if needed

---

Iteration Notes

If initial designs don't meet criteria: 1. [ ] Adjust RFdiffusion steps (try 75-100) 2. [ ] Lower ProteinMPNN temperature (try 0.05) 3. [ ] Generate more backbones 4. [ ] Try different design mode 5. [ ] Document iterations in report

Therapeutic Protein Designer - Design Procedures

Detailed code examples and procedures for each design phase.

---

Phase 1: Target Characterization

def get_target_structure(tu, target_id):
    """Get target structure: PDB first, then EMDB cryo-EM, then AlphaFold."""
    # Try PDB (X-ray/NMR)
    pdb_results = tu.tools.PDBe_get_uniprot_mappings(uniprot_id=target_id)
    if pdb_results:
        best_pdb = sorted(pdb_results, key=lambda x: x['resolution'])[0]
        return {'source': 'PDB', 'pdb_id': best_pdb['pdb_id']}

    # Try EMDB (cryo-EM, good for membrane proteins)
    protein_info = tu.tools.UniProt_get_entry_by_accession(accession=target_id)
    emdb_results = tu.tools.EMDB_search_structures(query=protein_info['proteinDescription']['recommendedName']['fullName']['value'])
    if emdb_results:
        best_emdb = sorted(emdb_results, key=lambda x: x.get('resolution', 99))[0]
        emdb_details = tu.tools.EMDB_get_structure(entry_id=best_emdb['emdb_id'])
        if emdb_details.get('pdb_ids'):
            return {'source': 'EMDB cryo-EM', 'emdb_id': best_emdb['emdb_id'], 'pdb_id': emdb_details['pdb_ids'][0]}

    # Fallback: AlphaFold prediction
    sequence = tu.tools.UniProt_get_sequence_by_accession(accession=target_id)
    structure = tu.tools.NvidiaNIM_alphafold2(sequence=sequence['sequence'], algorithm="mmseqs2")
    return {'source': 'AlphaFold2 (predicted)', 'structure': structure}

---

Phase 2-3: Backbone + Sequence Design

# Phase 2: RFdiffusion backbone generation
backbones = tu.tools.NvidiaNIM_rfdiffusion(diffusion_steps=50)

# Phase 3: ProteinMPNN sequence design
sequences = tu.tools.NvidiaNIM_proteinmpnn(pdb_string=backbone_pdb, num_sequences=8, temperature=0.1)

Sampling Parameters

Design modes: Unconditional (de novo scaffold), Binder design (target-guided), Motif scaffolding (functional motif embedding).

---

Phase 4: Structure Validation

predicted = tu.tools.NvidiaNIM_esmfold(sequence=sequence)
plddt = extract_plddt(predicted)
ptm = extract_ptm(predicted)
passes = np.mean(plddt) > 70 and ptm > 0.7

---

Phase 5: Developability Assessment

---

Fallback Chains

Protein Design: Report Templates and Output Formats

---

Report Template

# Therapeutic Protein Design Report: [TARGET]

**Generated**: [Date] | **Query**: [Original query] | **Status**: In Progress

---

## Executive Summary
[Designing...]

---

## 1. Target Characterization
### 1.1 Target Information
| Property | Value |
|----------|-------|
| **Target** | [Name] |
| **UniProt** | [ID] |
| **Structure source** | [PDB/EMDB/AlphaFold] |
| **Binding epitope** | [Description] |
| **Known binders** | [List] |

### 1.2 Epitope Analysis
| Residue Range | Type | Surface Area | Druggability |
|---------------|------|--------------|--------------|

### 1.2b Cryo-EM Structures (if membrane protein)
| EMDB ID | Resolution | PDB Model | Conformation |
|---------|------------|-----------|--------------|

---

## 2. Backbone Generation
### 2.1 Design Parameters
| Parameter | Value |
|-----------|-------|
| **Method** | RFdiffusion via NVIDIA NIM |
| **Design mode** | [mode] |
| **Diffusion steps** | [N] |

### 2.2 Generated Backbones
| Backbone | Length | Topology | Quality |
|----------|--------|----------|---------|

---

## 3. Sequence Design
### 3.1 Design Parameters
| Parameter | Value |
|-----------|-------|
| **Method** | ProteinMPNN via NVIDIA NIM |
| **Temperature** | [value] |
| **Sequences per backbone** | [N] |

### 3.2 Designed Sequences (Top 10)
| Rank | Backbone | Sequence ID | Length | MPNN Score | pI |
|------|----------|-------------|--------|------------|-----|

---

## 4. Structure Validation
| Sequence | pLDDT | pTM | RMSD to Design | Status |
|----------|-------|-----|----------------|--------|

---

## 5. Developability Assessment
| Design | Aggregation | pI | Cysteines | Expression | Overall |
|--------|-------------|-----|-----------|------------|---------|

---

## 6. Final Candidates
### 6.1 Ranked List
### 6.2 Sequences for Testing

---

## 7. Experimental Recommendations

---

## 8. Data Sources

---

EMDB Structure Table (for membrane proteins)

When targeting membrane proteins, prioritize cryo-EM structures:

### Cryo-EM Structures (EMDB)

| EMDB ID | Resolution | PDB Model | Conformation |
|---------|------------|-----------|--------------|
| EMD-12345 | 2.8 A | 7ABC | Active state |
| EMD-23456 | 3.1 A | 8DEF | Inactive state |

**Note**: Cryo-EM structures capture physiologically relevant conformations.

---

Design Entry Format

### Design: Binder_001

**Sequence**: MVLSPADKTN...
**Length**: 85 amino acids
**Target**: PD-L1 (UniProt: Q9NZQ7)
**Method**: RFdiffusion -> ProteinMPNN -> ESMFold validation

**Quality Metrics**:
| Metric | Value | Interpretation |
|--------|-------|----------------|
| pLDDT | 88.5 | High confidence |
| pTM | 0.82 | Good fold |
| ProteinMPNN score | -2.3 | Favorable |
| Predicted binding | Strong | Based on interface pLDDT |

*Source: NVIDIA NIM*

Therapeutic Protein Designer - Examples

Example 1: De Novo Binder Design

User Query

"Design a small protein binder against PD-L1 for cancer immunotherapy"

Expected Output

# Therapeutic Protein Design Report: PD-L1 Binder

**Generated**: 2026-02-04 | **Target**: PD-L1 (Q9NZQ7)

---

## Executive Summary

Successfully designed 4 high-quality protein binders targeting the PD-1 binding interface of PD-L1. **Top candidate (Design_001)** is a compact 85-residue 3-helix bundle with:
- pLDDT: 88.5 (high confidence)
- pTM: 0.85 (excellent topology)
- Low aggregation propensity
- Favorable expression characteristics

**Recommended for experimental testing**: Design_001, Design_003, Design_005 (all Tier 1)

---

## 1. Target Characterization

### 1.1 Target Information

| Property | Value |
|----------|-------|
| **Target** | PD-L1 (Programmed death-ligand 1) |
| **UniProt** | Q9NZQ7 |
| **Structure** | PDB: 4ZQK (2.0 Å, X-ray) |
| **Binding site** | IgV domain, PD-1 interface |
| **Known therapeutics** | Atezolizumab, Durvalumab, Avelumab |

### 1.2 Target Structure Quality

| Metric | Value |
|--------|-------|
| Resolution | 2.0 Å |
| R-free | 0.228 |
| Completeness | 98.7% |
| Binding site pLDDT | N/A (experimental) |

### 1.3 Binding Epitope

**Selected interface**: PD-1 binding face (residues 54-68, 115-125)

| Hotspot Residue | Role | Conservation |
|-----------------|------|--------------|
| Y56 | Hydrophobic contact | 100% |
| D61 | Salt bridge | 98% |
| N63 | H-bond | 95% |
| V68 | Hydrophobic | 100% |
| R113 | Salt bridge | 92% |

*Source: PDB 4ZQK, ConSurf analysis*

---

## 2. Backbone Generation

### 2.1 Design Parameters

| Parameter | Value |
|-----------|-------|
| **Method** | RFdiffusion via NVIDIA NIM |
| **Design mode** | Unconditional (scaffolds for grafting) |
| **Diffusion steps** | 50 |
| **Target size** | 70-100 residues |
| **Number generated** | 10 backbones |

### 2.2 Generated Backbones

| Backbone | Length | Topology | SS Content | Quality |
|----------|--------|----------|------------|---------|
| BB_001 | 85 aa | 3-helix bundle | 72% α | ★★★ |
| BB_002 | 92 aa | β-sandwich | 45% β | ★★☆ |
| BB_003 | 78 aa | α/β mixed | 40% α, 25% β | ★★★ |
| BB_004 | 88 aa | 4-helix bundle | 78% α | ★★☆ |
| BB_005 | 95 aa | β-barrel | 55% β | ★★☆ |
| BB_006 | 82 aa | α/β roll | 35% α, 30% β | ★☆☆ |

### 2.3 Selected Backbones

**Selected for sequence design**: BB_001, BB_003, BB_004, BB_005

**Rationale**:
- BB_001: Compact, stable fold, ideal for binders
- BB_003: Versatile topology for interface optimization
- BB_004: Larger interface area possible
- BB_005: Different topology for diversity

*Source: NVIDIA NIM via `NvidiaNIM_rfdiffusion`*

---

## 3. Sequence Design

### 3.1 Design Parameters

| Parameter | Value |
|-----------|-------|
| **Method** | ProteinMPNN via NVIDIA NIM |
| **Temperature** | 0.1 (conservative) |
| **Sequences per backbone** | 8 |
| **Total sequences** | 32 |

### 3.2 All Designed Sequences (Ranked by MPNN Score)

| Rank | Backbone | ID | Length | MPNN Score |
|------|----------|-------|--------|------------|
| 1 | BB_001 | Design_001 | 85 | -1.89 |
| 2 | BB_003 | Design_003 | 78 | -1.92 |
| 3 | BB_001 | Design_002 | 85 | -2.01 |
| 4 | BB_004 | Design_008 | 88 | -2.05 |
| 5 | BB_005 | Design_005 | 95 | -2.08 |
| 6 | BB_003 | Design_004 | 78 | -2.12 |
| 7 | BB_004 | Design_009 | 88 | -2.15 |
| 8 | BB_001 | Design_006 | 85 | -2.18 |
| 9 | BB_005 | Design_007 | 95 | -2.21 |
| 10 | BB_003 | Design_010 | 78 | -2.25 |

### 3.3 Top Sequences

**Design_001** (Rank 1):

>Design_001 | BB_001 | MPNN=-1.89 MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSH GSAQVKGHGKKVADALTNAVAHVDDMPNALSALSDLHAHKL

**Design_003** (Rank 2):

>Design_003 | BB_003 | MPNN=-1.92 MKKYTCTVCGYIYNPEDGDPDDNGGGGGGGKVWKGGGGGGDPTSDELIK ALQEASGTEVELKTKGDKFKKVS

*Source: NVIDIA NIM via `NvidiaNIM_proteinmpnn`*

---

## 4. Structure Validation

### 4.1 ESMFold Validation Results

| Design | pLDDT | pTM | RMSD | Status |
|--------|-------|-----|------|--------|
| Design_001 | 88.5 | 0.85 | 1.2 Å | ✓ PASS |
| Design_003 | 84.2 | 0.81 | 1.4 Å | ✓ PASS |
| Design_002 | 82.1 | 0.78 | 1.6 Å | ✓ PASS |
| Design_008 | 79.5 | 0.75 | 1.9 Å | ✓ PASS |
| Design_005 | 86.3 | 0.83 | 1.3 Å | ✓ PASS |
| Design_004 | 75.8 | 0.72 | 2.1 Å | ✓ PASS |
| Design_009 | 68.2 | 0.64 | 2.8 Å | ✗ FAIL |
| Design_006 | 71.4 | 0.69 | 2.4 Å | ✓ PASS |

### 4.2 Top Design: Design_001 Regional Breakdown

| Region | Residues | pLDDT | Role |
|--------|----------|-------|------|
| Helix 1 | 1-28 | 92.3 | Core |
| Loop 1 | 29-35 | 78.4 | Connector |
| Helix 2 | 36-58 | 91.8 | Core |
| Loop 2 | 59-65 | 75.2 | Potential interface |
| Helix 3 | 66-85 | 90.1 | Core |

**Assessment**: Excellent fold quality with confident core and flexible loops suitable for interface formation.

*Source: NVIDIA NIM via `NvidiaNIM_esmfold`*

---

## 5. Developability Assessment

### 5.1 Developability Scores

| Design | Aggregation | pI | MW (kDa) | Cys | Expression | Overall |
|--------|-------------|-----|----------|-----|------------|---------|
| Design_001 | 0.32 | 6.2 | 9.5 | 0 | High | ★★★ |
| Design_003 | 0.38 | 5.8 | 8.7 | 0 | High | ★★★ |
| Design_005 | 0.41 | 6.8 | 10.6 | 0 | High | ★★★ |
| Design_002 | 0.45 | 7.1 | 9.5 | 0 | Medium | ★★☆ |
| Design_008 | 0.52 | 5.4 | 9.8 | 2 | Medium | ★★☆ |

### 5.2 Expression Predictions

| Design | E. coli | Mammalian | Recommended |
|--------|---------|-----------|-------------|
| Design_001 | High (soluble) | High | E. coli |
| Design_003 | High (soluble) | High | E. coli |
| Design_005 | Medium | High | E. coli |

*Source: Sequence analysis, aggregation prediction*

---

## 6. Final Candidates

### 6.1 Ranked Candidate List

| Rank | Design | pLDDT | pTM | MPNN | Aggreg. | Tier | Recommendation |
|------|--------|-------|-----|------|---------|------|----------------|
| 1 | **Design_001** | 88.5 | 0.85 | -1.89 | 0.32 | ★★★ | TOP CHOICE |
| 2 | **Design_005** | 86.3 | 0.83 | -2.08 | 0.41 | ★★★ | High priority |
| 3 | **Design_003** | 84.2 | 0.81 | -1.92 | 0.38 | ★★★ | High priority |
| 4 | Design_002 | 82.1 | 0.78 | -2.01 | 0.45 | ★★☆ | Backup |
| 5 | Design_008 | 79.5 | 0.75 | -2.05 | 0.52 | ★★☆ | Backup |

### 6.2 Sequences for Testing

>Design_001 | pLDDT=88.5 | pTM=0.85 | MPNN=-1.89 | Tier=T1 MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSH GSAQVKGHGKKVADALTNAVAHVDDMPNALSALSDLHAHKL

>Design_005 | pLDDT=86.3 | pTM=0.83 | MPNN=-2.08 | Tier=T1 MTTLAELKKLAELLPKLEELKKELLEALKKAELKKEVAELLKKLAKELLKE LAKELLKKLAKELLKKELKEALKKALKELLKEALKKAVEEL

>Design_003 | pLDDT=84.2 | pTM=0.81 | MPNN=-1.92 | Tier=T1 MKKYTCTVCGYIYNPEDGDPDDNGGGGGGGKVWKGGGGGGDPTSDELIK ALQEASGTEVELKTKGDKFKKVS

---

## 7. Experimental Recommendations

### 7.1 Expression and Purification

| Step | Protocol | Notes |
|------|----------|-------|
| **Expression** | E. coli BL21(DE3) | All designs cysteine-free |
| **Induction** | 0.5 mM IPTG, 18°C, 16h | Optimize for each |
| **Lysis** | Sonication in PBS | Add protease inhibitors |
| **Purification** | Ni-NTA (if His-tagged) | Or ion exchange |
| **Polishing** | Size exclusion | Superdex 75 |

### 7.2 Characterization Assays

| Assay | Purpose | Priority |
|-------|---------|----------|
| SEC-MALS | Oligomeric state | High |
| CD spectroscopy | Secondary structure | High |
| DSF/Thermal shift | Stability | High |
| BLI/SPR | PD-L1 binding | High |
| Crystallography | Structure validation | Medium |

### 7.3 Expected Challenges

1. **Loop flexibility** - May need optimization for binding
2. **Affinity maturation** - Initial designs may need improvement
3. **Specificity** - Test against related proteins (PD-L2)

---

## 8. Data Sources

| Tool | Purpose | Query |
|------|---------|-------|
| PDB | Target structure | 4ZQK |
| UniProt | Target sequence | Q9NZQ7 |
| NvidiaNIM_rfdiffusion | Backbone generation | 50 steps |
| NvidiaNIM_proteinmpnn | Sequence design | T=0.1 |
| NvidiaNIM_esmfold | Validation | All designs |

---

Example 2: Enzyme Scaffold Design

User Query

"Design a stable protein scaffold that could host a catalytic triad"

Expected Output (Key Sections)

## Executive Summary

Designed 5 stable α/β hydrolase-like scaffolds capable of hosting a Ser-His-Asp catalytic triad. **Top candidate (Scaffold_003)** features:
- Optimal geometry for triad placement
- High stability (pLDDT: 91.2)
- TIM-barrel-like fold for substrate access

---

## 2. Backbone Generation

### Design Constraints

| Constraint | Value | Rationale |
|------------|-------|-----------|
| Topology | α/β hydrolase | Hosts catalytic triad |
| Size | 200-250 aa | Sufficient for active site |
| Cavity | Central | Substrate access |

### Generated Scaffolds

| Scaffold | Length | Topology | Cavity Volume | Quality |
|----------|--------|----------|---------------|---------|
| Scaffold_001 | 215 aa | 8-strand TIM barrel | 580 ų | ★★★ |
| Scaffold_003 | 228 aa | α/β hydrolase | 620 ų | ★★★ |
| Scaffold_005 | 240 aa | Rossmann-like | 450 ų | ★★☆ |

---

## 4. Active Site Design

### Catalytic Triad Placement

| Scaffold | Ser | His | Asp | Geometry |
|----------|-----|-----|-----|----------|
| Scaffold_003 | 152 | 237 | 209 | Optimal (3.1Å, 2.8Å) |
| Scaffold_001 | 145 | 220 | 192 | Good (3.3Å, 2.9Å) |

### Triad Quality Metrics

| Metric | Ideal | Scaffold_003 |
|--------|-------|--------------|
| Ser-His distance | 3.0 Å | 3.1 Å |
| His-Asp distance | 2.8 Å | 2.8 Å |
| Ser Oγ accessibility | Exposed | Exposed |

---

Example 3: Miniprotein Design

User Query

"Design a minimal stable fold under 50 amino acids"

Expected Output (Key Sections)

## Executive Summary

Designed ultra-stable miniproteins (35-50 aa) using constrained RFdiffusion. **Top design (Mini_002)** is a 42-residue designed fold with:
- pLDDT: 94.1 (exceptional)
- Three disulfide bonds for stability
- Predicted Tm > 90°C

---

## 2. Design Constraints

| Constraint | Value | Rationale |
|------------|-------|-----------|
| Length | 35-50 aa | Miniprotein range |
| Disulfides | 2-3 | Stability enhancement |
| Topology | Constrained | Ensure compact fold |

---

## 4. Validation Results

| Design | Length | Disulfides | pLDDT | pTM | Predicted Tm |
|--------|--------|------------|-------|-----|--------------|
| Mini_002 | 42 | 3 (C8-C38, C15-C29, C22-C45) | 94.1 | 0.92 | >90°C |
| Mini_005 | 38 | 2 (C5-C35, C12-C28) | 91.8 | 0.89 | ~85°C |
| Mini_001 | 45 | 3 | 89.2 | 0.86 | ~80°C |

---

## 6. Final Candidate

>Mini_002 | 42aa | 3SS | pLDDT=94.1 | Tier=T1 RCPEFVGCCNPACGPKYCRSCVRSGCVGCCPEFVGPCNPACKCR

**Notes**: 
- Express in oxidizing environment (periplasm or in vitro)
- Verify disulfide pairing by mass spec
- Test thermal stability by CD

Therapeutic Protein Designer - Tool Reference

Core NVIDIA NIM Tools

RFdiffusion - Backbone Generation

Example - Generate backbones:

# Generate de novo backbones
result = tu.tools.NvidiaNIM_rfdiffusion(
    diffusion_steps=50
)
# Returns: {"structure": "<PDB content>", "sequence": "GGG..."}

Parameters:

Notes:

• More steps (75-100) = higher quality but slower
• Output is backbone-only (Gly residues)
• Use with ProteinMPNN for sequence design

---

ProteinMPNN - Sequence Design

Example - Design sequences:

# Design sequences for backbone
result = tu.tools.NvidiaNIM_proteinmpnn(
    pdb_string=backbone_pdb_content,
    num_sequences=8,
    temperature=0.1
)
# Returns: {"sequences": ["MVLS...", "MKKT...", ...], "scores": [-1.89, -2.01, ...]}

Parameters:

Temperature Guide:

---

ESMFold - Fast Structure Validation

Example - Validate design:

# Predict structure for designed sequence
result = tu.tools.NvidiaNIM_esmfold(sequence=designed_sequence)
# Returns: {"structure": "<PDB content>", "plddt": [...], "ptm": 0.85}

Parameters:

Output Interpretation:

---

AlphaFold2 - High-Accuracy Validation

Example - High-accuracy prediction:

# High-accuracy structure prediction
result = tu.tools.NvidiaNIM_alphafold2(
    sequence=designed_sequence,
    algorithm="mmseqs2",
    relax_prediction=False
)
# Returns: {"structure": "<PDB content>", "plddt": [...]}

When to use instead of ESMFold:

• Final validation of top candidates
• Sequences >1024 aa
• When highest accuracy needed

---

ESM2 - Sequence Embeddings

Example - Get embeddings:

# Get sequence embeddings for similarity analysis
result = tu.tools.NvidiaNIM_esm2_650m(
    sequences=[seq1, seq2, seq3],
    format="npz"
)
# Returns: Binary NPZ file with embeddings

Use cases:

• Compare designed sequences to natural proteins
• Cluster designs by similarity
• Quality assessment

---

Supporting Tools

Target Structure Retrieval

Example - Get target structure:

# Try PDB first
pdb_hits = tu.tools.PDBe_get_uniprot_mappings(uniprot_id="Q9NZQ7")
if pdb_hits:
    structure = tu.tools.PDB_get_structure(pdb_id=pdb_hits[0]['pdb_id'])
else:
    # Fallback to AlphaFold
    structure = tu.tools.alphafold_get_prediction(accession="Q9NZQ7")

EMDB Cryo-EM Structures (NEW)

When to use EMDB:

• Membrane protein targets (GPCRs, ion channels)
• Large macromolecular complexes
• Targets where conformational states matter
• When X-ray structures unavailable

Example - Get cryo-EM structure for membrane target:

# Search EMDB for membrane receptor
emdb_hits = tu.tools.EMDB_search_structures(query="EGFR membrane receptor")

if emdb_hits:
    # Get details including associated PDB models
    best_entry = emdb_hits[0]  # Often sorted by resolution
    details = tu.tools.EMDB_get_structure(entry_id=best_entry['emdb_id'])
    
    # Get the atomic model (PDB) for design
    if details.get('pdb_ids'):
        structure = tu.tools.PDB_get_structure(pdb_id=details['pdb_ids'][0])
        print(f"Got structure from cryo-EM: {details['pdb_ids'][0]}")
        print(f"Resolution: {best_entry.get('resolution', 'N/A')} Å")

Output Quality Assessment:

Sequence Analysis

---

Workflow Code Examples

Example 1: Complete Design Pipeline

def design_protein_binder(tu, target_uniprot):
    """Complete binder design pipeline."""
    
    # Phase 1: Get target
    target_seq = tu.tools.UniProt_get_sequence_by_accession(
        accession=target_uniprot
    )
    target_structure = tu.tools.NvidiaNIM_alphafold2(
        sequence=target_seq['sequence'],
        algorithm="mmseqs2"
    )
    
    # Phase 2: Generate backbones
    backbones = []
    for i in range(5):
        bb = tu.tools.NvidiaNIM_rfdiffusion(diffusion_steps=50)
        backbones.append(bb)
    
    # Phase 3: Design sequences
    all_sequences = []
    for bb in backbones:
        seqs = tu.tools.NvidiaNIM_proteinmpnn(
            pdb_string=bb['structure'],
            num_sequences=8,
            temperature=0.1
        )
        all_sequences.extend(zip(seqs['sequences'], seqs['scores']))
    
    # Phase 4: Validate
    validated = []
    for seq, mpnn_score in all_sequences:
        pred = tu.tools.NvidiaNIM_esmfold(sequence=seq)
        plddt = np.mean(pred['plddt'])
        ptm = pred['ptm']
        
        if plddt > 70 and ptm > 0.7:
            validated.append({
                'sequence': seq,
                'mpnn_score': mpnn_score,
                'plddt': plddt,
                'ptm': ptm
            })
    
    # Rank by quality
    return sorted(validated, 
                  key=lambda x: (x['plddt'] + x['ptm']*100 - x['mpnn_score']),
                  reverse=True)

Example 2: Iterative Refinement

def iterative_design(tu, initial_backbone, target_plddt=85):
    """Iteratively improve design quality."""
    
    best_design = None
    best_plddt = 0
    
    for iteration in range(3):
        # Increase diffusion steps each iteration
        steps = 50 + iteration * 25
        
        # Generate backbone
        bb = tu.tools.NvidiaNIM_rfdiffusion(diffusion_steps=steps)
        
        # Design sequences with decreasing temperature
        temp = 0.1 / (iteration + 1)
        seqs = tu.tools.NvidiaNIM_proteinmpnn(
            pdb_string=bb['structure'],
            num_sequences=16,
            temperature=temp
        )
        
        # Validate all
        for seq, score in zip(seqs['sequences'], seqs['scores']):
            pred = tu.tools.NvidiaNIM_esmfold(sequence=seq)
            plddt = np.mean(pred['plddt'])
            
            if plddt > best_plddt:
                best_plddt = plddt
                best_design = {
                    'sequence': seq,
                    'structure': pred['structure'],
                    'plddt': plddt,
                    'iteration': iteration
                }
        
        # Early exit if target reached
        if best_plddt >= target_plddt:
            break
    
    return best_design

Example 3: Developability Screening

def assess_developability(sequence):
    """Assess developability of designed protein."""
    
    # Calculate properties
    from Bio.SeqUtils.ProtParam import ProteinAnalysis
    analysis = ProteinAnalysis(sequence)
    
    # Basic properties
    mw = analysis.molecular_weight()
    pi = analysis.isoelectric_point()
    gravy = analysis.gravy()
    
    # Count cysteines
    cys_count = sequence.count('C')
    
    # Aggregation propensity (simplified)
    hydrophobic = sum(1 for aa in sequence if aa in 'VILMFYW')
    agg_score = hydrophobic / len(sequence)
    
    # Score
    score = 0
    if 5 <= pi <= 9: score += 1
    if mw < 50000: score += 1
    if agg_score < 0.5: score += 1
    if cys_count == 0 or cys_count % 2 == 0: score += 1
    
    return {
        'molecular_weight': mw,
        'isoelectric_point': pi,
        'gravy': gravy,
        'cysteine_count': cys_count,
        'aggregation_score': agg_score,
        'developability_score': score,
        'tier': '★★★' if score >= 4 else '★★☆' if score >= 3 else '★☆☆'
    }

---

Fallback Chains

Backbone Generation

Sequence Design

Structure Validation

Target Structure

Cryo-EM (Membrane Targets) (NEW)

---

Common Parameter Mistakes

---

NVIDIA NIM Requirements

API Key: NVIDIA_API_KEY environment variable required

Rate limits: 40 RPM (1.5 second minimum between calls)

Check Availability

import os

nvidia_available = bool(os.environ.get("NVIDIA_API_KEY"))
if not nvidia_available:
    raise ValueError("NVIDIA_API_KEY required for protein design")

Async Operations

• AlphaFold2 may return 202 (polling required)
• RFdiffusion is typically synchronous
• ESMFold is synchronous

---

Quality Thresholds

Structure Prediction

ProteinMPNN Score

Design Tiers

---

Batch Processing Tips

Efficient Pipeline

def batch_validate(tu, sequences, batch_size=5):
    """Validate sequences in batches to manage rate limits."""
    import time
    
    results = []
    for i in range(0, len(sequences), batch_size):
        batch = sequences[i:i+batch_size]
        
        for seq in batch:
            result = tu.tools.NvidiaNIM_esmfold(sequence=seq)
            results.append(result)
            time.sleep(1.5)  # Rate limit
        
        # Longer pause between batches
        time.sleep(5)
    
    return results

Parallel Backbone Generation

# Generate diverse backbones
backbones = []
for _ in range(10):
    bb = tu.tools.NvidiaNIM_rfdiffusion(diffusion_steps=50)
    backbones.append(bb)
    time.sleep(1.5)  # Rate limit