protein-design-workflow

Name: protein-design-workflow
Author: adaptyvbio

43installs20forks

How to install protein-design-workflow?

adaptyvbio/protein-design-skills/protein-design-workflow

$npx -y skills add adaptyvbio/protein-design-skills --skill protein-design-workflow

Installs into the current project.

Summary

End-to-end guidance for protein design pipelines. Use this skill when: (1) Starting a new protein design project, (2) Need step-by-step workflow guidance, (3) Understanding the full design pipeline, (4) Planning compute resources and timelines, (5) Integrating multiple design too

Files · 1

View on GitHub

SKILL.md

1	# Protein Design Workflow Guide
2
3	## Standard binder design pipeline
4
5	### Overview
6	```
7	Target Preparation --> Backbone Generation --> Sequence Design
8	\| \| \|
9	v v v
10	(pdb skill) (rfdiffusion) (proteinmpnn)
11	\| \|
12	v v
13	Structure Validation --> Filtering
14	\| \|
15	v v
16	(alphafold/chai) (protein-qc)
17	```
18
19	## Phase 1: Target preparation
20
21	### 1.1 Obtain target structure
22	```bash
23	# Download from PDB
24	curl -o target.pdb "https://files.rcsb.org/download/XXXX.pdb"
25	```
26
27	### 1.2 Clean and prepare
28	```python
29	# Extract target chain
30	# Remove waters, ligands if needed
31	# Trim to binding region + 10A buffer
32	```
33
34	### 1.3 Select hotspots
35	- Choose 3-6 exposed residues
36	- Prefer charged/aromatic (K, R, E, D, W, Y, F)
37	- Check surface accessibility
38	- Verify residue numbering
39
40	Output: `target_prepared.pdb`, hotspot list
41
42	## Phase 2: Backbone generation
43
44	### Option A: RFdiffusion (diverse exploration)
45	```bash
46	# RFdiffusion runs from the official repo, not biomodals
47	python run_inference.py \
48	inference.input_pdb=target_prepared.pdb \
49	contigmap.contigs=[A1-150/0 70-100] \
50	ppi.hotspot_res=[A45,A67,A89] \
51	inference.num_designs=500
52	```
53
54	### Option B: BindCraft (end-to-end)
55	```bash
56	modal run modal_bindcraft.py \
57	--input-pdb target_prepared.pdb \
58	--target-hotspot-residues "45,67,89" \
59	--number-of-final-designs 100
60	```
61
62	Output: 100-500 backbone PDBs
63
64	## Phase 3: Sequence design
65
66	### For RFdiffusion backbones
67	```bash
68	for backbone in backbones/*.pdb; do
69	modal run modal_ligandmpnn.py \
70	--input-pdb "$backbone" \
71	--params-str "--number_of_batches 8 --temperature 0.1"
72	done
73	```
74
75	Output: 8 sequences per backbone (800-4000 total)
76
77	## Phase 4: Structure validation
78
79	### Predict complexes
80	```bash
81	# Prepare FASTA with binder + target
82	# binder:target format for multimer
83
84	modal run modal_alphafold.py \
85	--input-fasta all_sequences.fasta \
86	--out-dir predictions/
87	```
88
89	Output: AF2 predictions with pLDDT, ipTM, PAE
90
91	## Phase 5: Filtering and selection
92
93	### Apply standard thresholds
94	```python
95	import pandas as pd
96
97	# Load metrics
98	designs = pd.read_csv('all_metrics.csv')
99
100	# Filter
101	filtered = designs[
102	(designs['pLDDT'] > 0.85) &
103	(designs['ipTM'] > 0.50) &
104	(designs['PAE_interface'] < 10) &
105	(designs['scRMSD'] < 2.0) &
106	(designs['esm2_pll'] > 0.0)
107	]
108
109	# Rank by composite score
110	filtered['score'] = (
111	0.3 * filtered['pLDDT'] +
112	0.3 * filtered['ipTM'] +
113	0.2 * (1 - filtered['PAE_interface'] / 20) +
114	0.2 * filtered['esm2_pll']
115	)
116
117	top_designs = filtered.nlargest(50, 'score')
118	```
119
120	Output: 50-200 filtered candidates
121
122	## Resource planning
123
124	### Compute requirements
125
126	\| Stage \| GPU \| Time (100 designs) \|
127	\|-------\|-----\|-------------------\|
128	\| RFdiffusion \| A10G \| 30 min \|
129	\| ProteinMPNN \| T4 \| 15 min \|
130	\| Chai / AlphaFold \| A100 \| 4-8 hours \|
131	\| Filtering \| CPU \| 15 min \|
132
133	### Total timeline
134	- Small campaign (100 designs): 8-12 hours
135	- Medium campaign (500 designs): 24-48 hours
136	- Large campaign (1000+ designs): 2-5 days
137
138	## Quality checkpoints
139
140	### After backbone generation
141	- [ ] Visual inspection of diverse backbones
142	- [ ] Secondary structure present
143	- [ ] No clashes with target
144
145	### After sequence design
146	- [ ] ESM2 PLL > 0.0 for most sequences
147	- [ ] No unwanted cysteines (unless intentional)
148	- [ ] Reasonable sequence diversity
149
150	### After validation
151	- [ ] pLDDT > 0.85
152	- [ ] ipTM > 0.50
153	- [ ] PAE_interface < 10
154	- [ ] Self-consistency RMSD < 2.0 A
155
156	### Final selection
157	- [ ] Diverse sequences (cluster if needed)
158	- [ ] Manufacturable (no problematic motifs)
159	- [ ] Reasonable molecular weight
160
161	## Common issues
162
163	\| Problem \| Solution \|
164	\|---------\|----------\|
165	\| Low ipTM \| Check hotspots, increase designs \|
166	\| Poor diversity \| Higher temperature, more backbones \|
167	\| High scRMSD \| Backbone may be unusual \|
168	\| Low pLDDT \| Check design quality \|
169
170	## Advanced workflows
171
172	### Multi-tool combination
173	1. RFdiffusion for initial backbones
174	2. Mosaic for gradient-based refinement
175	3. ProteinMPNN diversification
176	4. AF2 final validation
177
178	### Iterative refinement
179	1. Run initial campaign
180	2. Analyze failures
181	3. Adjust hotspots/parameters
182	4. Repeat with insights

People who liked this also viewed

6 picks

Type

protein-design-workflow

43installs20forks

View on GitHub

How to install protein-design-workflow?

adaptyvbio/protein-design-skills/protein-design-workflow

$npx -y skills add adaptyvbio/protein-design-skills --skill protein-design-workflow

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Protein Design Workflow Guide
2
3	## Standard binder design pipeline
4
5	### Overview
6	```
7	Target Preparation --> Backbone Generation --> Sequence Design
8	\| \| \|
9	v v v
10	(pdb skill) (rfdiffusion) (proteinmpnn)
11	\| \|
12	v v
13	Structure Validation --> Filtering
14	\| \|
15	v v
16	(alphafold/chai) (protein-qc)
17	```
18
19	## Phase 1: Target preparation
20
21	### 1.1 Obtain target structure
22	```bash
23	# Download from PDB
24	curl -o target.pdb "https://files.rcsb.org/download/XXXX.pdb"
25	```
26
27	### 1.2 Clean and prepare
28	```python
29	# Extract target chain
30	# Remove waters, ligands if needed
31	# Trim to binding region + 10A buffer
32	```
33
34	### 1.3 Select hotspots
35	- Choose 3-6 exposed residues
36	- Prefer charged/aromatic (K, R, E, D, W, Y, F)
37	- Check surface accessibility
38	- Verify residue numbering
39
40	Output: `target_prepared.pdb`, hotspot list
41
42	## Phase 2: Backbone generation
43
44	### Option A: RFdiffusion (diverse exploration)
45	```bash
46	# RFdiffusion runs from the official repo, not biomodals
47	python run_inference.py \
48	inference.input_pdb=target_prepared.pdb \
49	contigmap.contigs=[A1-150/0 70-100] \
50	ppi.hotspot_res=[A45,A67,A89] \
51	inference.num_designs=500
52	```
53
54	### Option B: BindCraft (end-to-end)
55	```bash
56	modal run modal_bindcraft.py \
57	--input-pdb target_prepared.pdb \
58	--target-hotspot-residues "45,67,89" \
59	--number-of-final-designs 100
60	```
61
62	Output: 100-500 backbone PDBs
63
64	## Phase 3: Sequence design
65
66	### For RFdiffusion backbones
67	```bash
68	for backbone in backbones/*.pdb; do
69	modal run modal_ligandmpnn.py \
70	--input-pdb "$backbone" \
71	--params-str "--number_of_batches 8 --temperature 0.1"
72	done
73	```
74
75	Output: 8 sequences per backbone (800-4000 total)
76
77	## Phase 4: Structure validation
78
79	### Predict complexes
80	```bash
81	# Prepare FASTA with binder + target
82	# binder:target format for multimer
83
84	modal run modal_alphafold.py \
85	--input-fasta all_sequences.fasta \
86	--out-dir predictions/
87	```
88
89	Output: AF2 predictions with pLDDT, ipTM, PAE
90
91	## Phase 5: Filtering and selection
92
93	### Apply standard thresholds
94	```python
95	import pandas as pd
96
97	# Load metrics
98	designs = pd.read_csv('all_metrics.csv')
99
100	# Filter
101	filtered = designs[
102	(designs['pLDDT'] > 0.85) &
103	(designs['ipTM'] > 0.50) &
104	(designs['PAE_interface'] < 10) &
105	(designs['scRMSD'] < 2.0) &
106	(designs['esm2_pll'] > 0.0)
107	]
108
109	# Rank by composite score
110	filtered['score'] = (
111	0.3 * filtered['pLDDT'] +
112	0.3 * filtered['ipTM'] +
113	0.2 * (1 - filtered['PAE_interface'] / 20) +
114	0.2 * filtered['esm2_pll']
115	)
116
117	top_designs = filtered.nlargest(50, 'score')
118	```
119
120	Output: 50-200 filtered candidates
121
122	## Resource planning
123
124	### Compute requirements
125
126	\| Stage \| GPU \| Time (100 designs) \|
127	\|-------\|-----\|-------------------\|
128	\| RFdiffusion \| A10G \| 30 min \|
129	\| ProteinMPNN \| T4 \| 15 min \|
130	\| Chai / AlphaFold \| A100 \| 4-8 hours \|
131	\| Filtering \| CPU \| 15 min \|
132
133	### Total timeline
134	- Small campaign (100 designs): 8-12 hours
135	- Medium campaign (500 designs): 24-48 hours
136	- Large campaign (1000+ designs): 2-5 days
137
138	## Quality checkpoints
139
140	### After backbone generation
141	- [ ] Visual inspection of diverse backbones
142	- [ ] Secondary structure present
143	- [ ] No clashes with target
144
145	### After sequence design
146	- [ ] ESM2 PLL > 0.0 for most sequences
147	- [ ] No unwanted cysteines (unless intentional)
148	- [ ] Reasonable sequence diversity
149
150	### After validation
151	- [ ] pLDDT > 0.85
152	- [ ] ipTM > 0.50
153	- [ ] PAE_interface < 10
154	- [ ] Self-consistency RMSD < 2.0 A
155
156	### Final selection
157	- [ ] Diverse sequences (cluster if needed)
158	- [ ] Manufacturable (no problematic motifs)
159	- [ ] Reasonable molecular weight
160
161	## Common issues
162
163	\| Problem \| Solution \|
164	\|---------\|----------\|
165	\| Low ipTM \| Check hotspots, increase designs \|
166	\| Poor diversity \| Higher temperature, more backbones \|
167	\| High scRMSD \| Backbone may be unusual \|
168	\| Low pLDDT \| Check design quality \|
169
170	## Advanced workflows
171
172	### Multi-tool combination
173	1. RFdiffusion for initial backbones
174	2. Mosaic for gradient-based refinement
175	3. ProteinMPNN diversification
176	4. AF2 final validation
177
178	### Iterative refinement
179	1. Run initial campaign
180	2. Analyze failures
181	3. Adjust hotspots/parameters
182	4. Repeat with insights

People who liked this also viewed

6 picks

Type