Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →AI-powered ternary complex prediction for targeted protein degradation, modeling POI-degrader-E3 ligase assemblies to optimize PROTAC and molecular glue efficacy.
.claude/skills/tpd-ternary-complex-agent/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | 36% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 62% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 88% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 93% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 72% | 0% |
<!--
#
#
-->
The TPD Ternary Complex Agent specializes in predicting and modeling ternary complex formation for targeted protein degradation (TPD). It uses AlphaFold-Multimer, molecular dynamics, and deep learning to model Protein of Interest (POI)-degrader-E3 ligase assemblies, enabling rational optimization of PROTACs and molecular glues.
| E3 Ligase | Structure | Complex Quality | |-----------|-----------|-----------------| | CRBN-DDB1-CUL4A | High resolution | Excellent | | VHL-ELOB-ELOC-CUL2 | High resolution | Excellent | | MDM2 | Good | Good | | IAP (cIAP1/XIAP) | Moderate | Moderate | | DCAF15-DDB1 | Emerging | Developing | | KEAP1 | High resolution | Good |
User: "Model the ternary complex for this BRD4 PROTAC with VHL to understand the protein-protein interface."
Agent Action:
bashpython3 Skills/Drug_Discovery/TPD_Ternary_Complex_Agent/predict_ternary.py \ --poi_structure brd4_bd1.pdb \ --warhead_pose brd4_warhead_docked.sdf \ --e3_ligase VHL \ --e3_ligand vhl_ligand.sdf \ --protac_smiles "PROTAC_SMILES_STRING" \ --linker_conformations 100 \ --md_refinement true \ --output ternary_complex_results/
| Score Component | Weight | Interpretation | |-----------------|--------|----------------| | Interface Area | 20% | Larger = more stable | | Shape Complementarity | 25% | Better fit = stability | | Electrostatics | 20% | Charge matching | | Linker Strain | 15% | Lower = better geometry | | Complex Stability (ΔG) | 20% | Favorable energetics |
| Output | Description | Format | |--------|-------------|--------| | Ternary Structure | POI-PROTAC-E3 model | .pdb | | Confidence Scores | pLDDT, PAE | .json | | Interface Map | Contact residues | .csv | | Lysine Accessibility | Ubiquitination sites | .csv | | Cooperativity | α factor estimate | .json | | Optimization Suggestions | Design recommendations | .md | | MD Trajectory | Stability simulation | .xtc |
| Metric | Definition | Good Value | |--------|------------|------------| | Buried Surface Area | Contact area | >800 Ų | | Shape Complementarity | Sc score | >0.65 | | Gap Volume Index | Interface packing | <2.0 | | Hydrogen Bonds | Intermolecular H-bonds | >3 | | Salt Bridges | Charged interactions | >1 |
Structure Prediction:
Conformational Sampling:
Scoring Functions:
| α Factor | Interpretation | Mechanism | |----------|----------------|-----------| | α > 1 | Positive cooperativity | E3 binding enhances POI binding | | α = 1 | No cooperativity | Independent binding | | α < 1 | Negative cooperativity | E3 binding reduces POI binding |
| Requirement | Threshold | Rationale | |-------------|-----------|-----------| | Surface Accessibility | >30 Ų | E2 access | | Distance to E2~Ub | <15 Å | Transfer distance | | Lysine Environment | Favorable | Not buried | | Number of Sites | ≥1 | At least one Lys |
| E3 | Advantages | Considerations | |----|------------|----------------| | CRBN | Broad applicability, many ligands | Some immune targets | | VHL | High selectivity, well-validated | Limited tissue in some organs | | MDM2 | No CRBN competition | Fewer validated targets | | IAP | Cancer expression, dual mechanism | Complex biology |
| Method | Purpose | Confidence | |--------|---------|------------| | Crystal Structure | Ground truth | Highest | | Cryo-EM | Large complexes | High | | HDX-MS | Interface mapping | Moderate-High | | Crosslinking MS | Distance constraints | Moderate | | Mutagenesis | Interface validation | Functional |
| Structural Finding | Design Action | |--------------------|---------------| | Poor interface | Change E3 or target site | | Long distance | Longer linker | | Steric clash | Shorter linker or different exit vector | | No accessible Lys | Different binding mode | | High flexibility | Constrained linker |
| QC Metric | Threshold | Interpretation | |-----------|-----------|----------------| | pLDDT (interface) | >70 | Reliable prediction | | PAE (POI-E3) | <10 Å | Good relative positioning | | MD RMSD | <3 Å | Stable complex | | Clash Score | <50 | Good packing |
AI Group - Biomedical AI Platform
<!-- AUTHOR_SIGNATURE: 9a7f3c2e-MD-BABU-MIA-2026-MSSM-SECURE -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 21,226 | 18,320 | -14% | 1 | 1 | 0% | 3,555 | 5,034 | +42% | 0 | 0 | — |
case-09 | fail→pass | 12,485 | 5,989 | -52% | 1 | 1 | 0% | 2,115 | 2,882 | +36% | 0 | 0 | — |
case-10 | fail→pass | 12,230 | 8,223 | -33% | 1 | 1 | 0% | 1,932 | 3,133 | +62% | 0 | 0 | — |
case-11 | fail→pass | 7,559 | 4,083 | -46% | 1 | 1 | 0% | 1,344 | 2,531 | +88% | 0 | 0 | — |
case-12 | pass→pass | 9,392 | 4,238 | -55% | 1 | 1 | 0% | 1,697 | 2,557 | +51% | 0 | 0 | — |
case-07 | fail→pass | 10,235 | 9,813 | -4% | 1 | 1 | 0% | 1,856 | 3,582 | +93% | 0 | 0 | — |
case-08 | pass→pass | 6,289 | 3,154 | -50% | 1 | 1 | 0% | 1,043 | 2,410 | +131% | 0 | 0 | — |
case-02 | fail→pass | 24,434 | 3,865 | -84% | 1 | 1 | 0% | 1,579 | 2,711 | +72% | 0 | 0 | — |
case-03 | fail→pass | 13,641 | 3,758 | -72% | 1 | 1 | 0% | 1,032 | 2,276 | +121% | 0 | 0 | — |
case-04 | fail→pass | 16,771 | 8,795 | -48% | 1 | 1 | 0% | 2,792 | 3,256 | +17% | 0 | 0 | — |
case-05 | pass→pass | 5,653 | 3,707 | -34% | 1 | 1 | 0% | 964 | 2,442 | +153% | 0 | 0 | — |
case-06 | pass→pass | 8,117 | 11,930 | +47% | 1 | 1 | 0% | 1,663 | 3,867 | +133% | 0 | 0 | — |
case-13 | pass→pass | 10,995 | 6,733 | -39% | 1 | 1 | 0% | 1,922 | 3,133 | +63% | 0 | 0 | — |
case-14 | fail→pass | 19,175 | 8,479 | -56% | 1 | 1 | 0% | 1,188 | 2,987 | +151% | 0 | 0 | — |
case-15 | pass→pass | 13,390 | 11,565 | -14% | 1 | 1 | 0% | 2,074 | 3,583 | +73% | 0 | 0 | — |
case-16 | pass→pass | 13,732 | 5,768 | -58% | 1 | 1 | 0% | 2,359 | 2,832 | +20% | 0 | 0 | — |
case-22 | fail→fail | 11,562 | 19,323 | +67% | 1 | 1 | 0% | 2,302 | 6,114 | +166% | 0 | 0 | — |
case-17 | fail→pass | 10,454 | 3,084 | -70% | 1 | 1 | 0% | 1,686 | 2,344 | +39% | 0 | 0 | — |
case-18 | fail→pass | 25,783 | 2,479 | -90% | 1 | 1 | 0% | 1,365 | 2,239 | +64% | 0 | 0 | — |
case-19 | pass→pass | 8,200 | 1,657 | -80% | 1 | 1 | 0% | 1,216 | 2,115 | +74% | 0 | 0 | — |
case-20 | fail→fail | 20,399 | 29,334 | +44% | 1 | 1 | 0% | 3,809 | 8,029 | +111% | 0 | 0 | — |
case-21 | fail→fail | 14,513 | 20,093 | +38% | 1 | 1 | 0% | 3,052 | 5,683 | +86% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 18 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 7/27/2026 | +27% |
Other measured skills in the registry, with their headline benchmark lift.