boltz

Name: boltz
Author: adaptyvbio

36installs20forks

How to install boltz?

adaptyvbio/protein-design-skills/boltz

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

Installs into the current project.

Summary

Structure prediction using Boltz-1/Boltz-2, an open biomolecular structure predictor. Use this skill when: (1) Predicting protein complex structures, (2) Validating designed binders, (3) Need open-source alternative to AF2, (4) Predicting protein-ligand complexes, (5) Using local

Files · 1

View on GitHub

SKILL.md

1	# Boltz Structure Prediction
2
3	## Prerequisites
4
5	\| Requirement \| Minimum \| Recommended \|
6	\|-------------\|---------\|-------------\|
7	\| Python \| 3.10+ \| 3.11 \|
8	\| CUDA \| 12.0+ \| 12.1+ \|
9	\| GPU VRAM \| 24GB \| 48GB (L40S) \|
10	\| RAM \| 32GB \| 64GB \|
11
12	## How to run
13
14	> First time? See [Getting started](../../docs/getting-started.md) to set up Modal and biomodals.
15
16	### Option 1: Modal
17	```bash
18	cd biomodals
19	modal run modal_boltz.py \
20	--input-faa complex.fasta \
21	--out-dir predictions/
22	```
23
24	GPU: L40S (48GB) \| Timeout: 1800s default
25
26	### Option 2: Local installation
27	```bash
28	pip install boltz
29
30	boltz predict \
31	--fasta complex.fasta \
32	--output predictions/
33	```
34
35	## Key parameters
36
37	\| Parameter \| Default \| Range \| Description \|
38	\|-----------\|---------\|-------\|-------------\|
39	\| `--recycling_steps` \| 3 \| 1-10 \| Recycling iterations \|
40	\| `--sampling_steps` \| 200 \| 50-500 \| Diffusion steps \|
41	\| `--use_msa_server` \| true \| bool \| Use MSA server \|
42
43	## FASTA Format
44
45	```
46	>protein_A
47	MKTAYIAKQRQISFVK...
48	>protein_B
49	MVLSPADKTNVKAAWG...
50	```
51
52	## Output format
53
54	```
55	predictions/
56	├── model_0.cif # Best model (CIF format)
57	├── confidence.json # pLDDT, pTM, ipTM
58	└── pae.npy # PAE matrix
59	```
60
61	Note: Boltz outputs CIF format. Convert to PDB if needed:
62	```python
63	from Bio.PDB import MMCIFParser, PDBIO
64	parser = MMCIFParser()
65	structure = parser.get_structure("model", "model_0.cif")
66	io = PDBIO()
67	io.set_structure(structure)
68	io.save("model_0.pdb")
69	```
70
71	## Comparison
72
73	\| Feature \| Boltz-1 \| Boltz-2 \| AF2-Multimer \|
74	\|---------\|---------\|---------\|--------------\|
75	\| MSA-free mode \| Yes \| Yes \| No \|
76	\| Diffusion \| Yes \| Yes \| No \|
77	\| Speed \| Fast \| Faster \| Slower \|
78	\| Open source \| Yes \| Yes \| Yes \|
79
80	## Sample output
81
82	### Successful run
83	```
84	$ boltz predict --fasta complex.fasta --output predictions/
85	[INFO] Loading Boltz-1 weights...
86	[INFO] Predicting structure...
87	[INFO] Saved model to predictions/model_0.cif
88
89	predictions/confidence.json:
90	{
91	"ptm": 0.78,
92	"iptm": 0.65,
93	"plddt": 0.81
94	}
95	```
96
97	What good output looks like:
98	- pTM: > 0.7 (confident global structure)
99	- ipTM: > 0.5 (confident interface)
100	- pLDDT: > 0.7 (confident per-residue)
101	- CIF file: ~100-500 KB for typical complex
102
103	## Decision tree
104
105	```
106	Should I use Boltz?
107	│
108	├─ What are you predicting?
109	│ ├─ Protein-protein complex → Boltz ✓ or Chai or ColabFold
110	│ ├─ Protein + ligand → Boltz ✓ or Chai
111	│ └─ Single protein → Use ESMFold (faster)
112	│
113	├─ Need MSA?
114	│ ├─ No / want speed → Boltz ✓
115	│ └─ Yes / maximum accuracy → ColabFold
116	│
117	└─ Why Boltz over Chai?
118	├─ Open weights preference → Boltz ✓
119	├─ Boltz-2 speed → Boltz ✓
120	└─ DNA/RNA support → Consider Chai
121	```
122
123	## Typical performance
124
125	\| Campaign Size \| Time (L40S) \| Cost (Modal) \| Notes \|
126	\|---------------\|-------------\|--------------\|-------\|
127	\| 100 complexes \| 30-45 min \| ~$8 \| Standard validation \|
128	\| 500 complexes \| 2-3h \| ~$35 \| Large campaign \|
129	\| 1000 complexes \| 4-6h \| ~$70 \| Comprehensive \|
130
131	Per-complex: ~15-30s for typical binder-target complex.
132
133	---
134
135	## Verify
136
137	```bash
138	find predictions -name "*.cif" \| wc -l # Should match input count
139	```
140
141	---
142
143	## Troubleshooting
144
145	Low confidence: Increase recycling_steps
146	OOM errors: Use MSA-free mode or A100-80GB
147	Slow prediction: Reduce sampling_steps
148
149	### Error interpretation
150
151	\| Error \| Cause \| Fix \|
152	\|-------\|-------\|-----\|
153	\| `RuntimeError: CUDA out of memory` \| Complex too large \| Use `--use_msa_server false` or larger GPU \|
154	\| `KeyError: 'iptm'` \| Single chain only \| Ensure FASTA has 2+ chains \|
155	\| `FileNotFoundError: weights` \| Missing model \| Run `boltz download` first \|
156	\| `ValueError: invalid residue` \| Non-standard AA \| Check for modified residues in sequence \|
157
158	### Boltz-1 vs Boltz-2
159
160	\| Aspect \| Boltz-1 \| Boltz-2 \|
161	\|--------\|---------\|---------\|
162	\| Speed \| Fast \| Faster \|
163	\| Accuracy \| Good \| Improved, notably antibody-antigen \|
164	\| Ligands \| Basic \| Better support \|
165	\| Affinity prediction \| No \| Yes (small-molecule binding) \|
166	\| Release \| 2024 \| 2025 \|
167
168	Boltz-2 is the current default. Boltz-1 is still used where a design pipeline
169	inverts the v1 model.
170
171	### Affinity prediction (Boltz-2)
172
173	Boltz-2 adds an affinity-prediction module that approaches free-energy-perturbation
174	accuracy at a fraction of the cost. It is trained on small-molecule binding data, so
175	use it for protein-ligand and small-molecule work. It does not predict
176	protein-protein binding affinity; for protein binders, rely on interface confid

People who liked this also viewed

6 picks

Type

boltz

36installs20forks

View on GitHub

How to install boltz?

adaptyvbio/protein-design-skills/boltz

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

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Boltz Structure Prediction
2
3	## Prerequisites
4
5	\| Requirement \| Minimum \| Recommended \|
6	\|-------------\|---------\|-------------\|
7	\| Python \| 3.10+ \| 3.11 \|
8	\| CUDA \| 12.0+ \| 12.1+ \|
9	\| GPU VRAM \| 24GB \| 48GB (L40S) \|
10	\| RAM \| 32GB \| 64GB \|
11
12	## How to run
13
14	> First time? See [Getting started](../../docs/getting-started.md) to set up Modal and biomodals.
15
16	### Option 1: Modal
17	```bash
18	cd biomodals
19	modal run modal_boltz.py \
20	--input-faa complex.fasta \
21	--out-dir predictions/
22	```
23
24	GPU: L40S (48GB) \| Timeout: 1800s default
25
26	### Option 2: Local installation
27	```bash
28	pip install boltz
29
30	boltz predict \
31	--fasta complex.fasta \
32	--output predictions/
33	```
34
35	## Key parameters
36
37	\| Parameter \| Default \| Range \| Description \|
38	\|-----------\|---------\|-------\|-------------\|
39	\| `--recycling_steps` \| 3 \| 1-10 \| Recycling iterations \|
40	\| `--sampling_steps` \| 200 \| 50-500 \| Diffusion steps \|
41	\| `--use_msa_server` \| true \| bool \| Use MSA server \|
42
43	## FASTA Format
44
45	```
46	>protein_A
47	MKTAYIAKQRQISFVK...
48	>protein_B
49	MVLSPADKTNVKAAWG...
50	```
51
52	## Output format
53
54	```
55	predictions/
56	├── model_0.cif # Best model (CIF format)
57	├── confidence.json # pLDDT, pTM, ipTM
58	└── pae.npy # PAE matrix
59	```
60
61	Note: Boltz outputs CIF format. Convert to PDB if needed:
62	```python
63	from Bio.PDB import MMCIFParser, PDBIO
64	parser = MMCIFParser()
65	structure = parser.get_structure("model", "model_0.cif")
66	io = PDBIO()
67	io.set_structure(structure)
68	io.save("model_0.pdb")
69	```
70
71	## Comparison
72
73	\| Feature \| Boltz-1 \| Boltz-2 \| AF2-Multimer \|
74	\|---------\|---------\|---------\|--------------\|
75	\| MSA-free mode \| Yes \| Yes \| No \|
76	\| Diffusion \| Yes \| Yes \| No \|
77	\| Speed \| Fast \| Faster \| Slower \|
78	\| Open source \| Yes \| Yes \| Yes \|
79
80	## Sample output
81
82	### Successful run
83	```
84	$ boltz predict --fasta complex.fasta --output predictions/
85	[INFO] Loading Boltz-1 weights...
86	[INFO] Predicting structure...
87	[INFO] Saved model to predictions/model_0.cif
88
89	predictions/confidence.json:
90	{
91	"ptm": 0.78,
92	"iptm": 0.65,
93	"plddt": 0.81
94	}
95	```
96
97	What good output looks like:
98	- pTM: > 0.7 (confident global structure)
99	- ipTM: > 0.5 (confident interface)
100	- pLDDT: > 0.7 (confident per-residue)
101	- CIF file: ~100-500 KB for typical complex
102
103	## Decision tree
104
105	```
106	Should I use Boltz?
107	│
108	├─ What are you predicting?
109	│ ├─ Protein-protein complex → Boltz ✓ or Chai or ColabFold
110	│ ├─ Protein + ligand → Boltz ✓ or Chai
111	│ └─ Single protein → Use ESMFold (faster)
112	│
113	├─ Need MSA?
114	│ ├─ No / want speed → Boltz ✓
115	│ └─ Yes / maximum accuracy → ColabFold
116	│
117	└─ Why Boltz over Chai?
118	├─ Open weights preference → Boltz ✓
119	├─ Boltz-2 speed → Boltz ✓
120	└─ DNA/RNA support → Consider Chai
121	```
122
123	## Typical performance
124
125	\| Campaign Size \| Time (L40S) \| Cost (Modal) \| Notes \|
126	\|---------------\|-------------\|--------------\|-------\|
127	\| 100 complexes \| 30-45 min \| ~$8 \| Standard validation \|
128	\| 500 complexes \| 2-3h \| ~$35 \| Large campaign \|
129	\| 1000 complexes \| 4-6h \| ~$70 \| Comprehensive \|
130
131	Per-complex: ~15-30s for typical binder-target complex.
132
133	---
134
135	## Verify
136
137	```bash
138	find predictions -name "*.cif" \| wc -l # Should match input count
139	```
140
141	---
142
143	## Troubleshooting
144
145	Low confidence: Increase recycling_steps
146	OOM errors: Use MSA-free mode or A100-80GB
147	Slow prediction: Reduce sampling_steps
148
149	### Error interpretation
150
151	\| Error \| Cause \| Fix \|
152	\|-------\|-------\|-----\|
153	\| `RuntimeError: CUDA out of memory` \| Complex too large \| Use `--use_msa_server false` or larger GPU \|
154	\| `KeyError: 'iptm'` \| Single chain only \| Ensure FASTA has 2+ chains \|
155	\| `FileNotFoundError: weights` \| Missing model \| Run `boltz download` first \|
156	\| `ValueError: invalid residue` \| Non-standard AA \| Check for modified residues in sequence \|
157
158	### Boltz-1 vs Boltz-2
159
160	\| Aspect \| Boltz-1 \| Boltz-2 \|
161	\|--------\|---------\|---------\|
162	\| Speed \| Fast \| Faster \|
163	\| Accuracy \| Good \| Improved, notably antibody-antigen \|
164	\| Ligands \| Basic \| Better support \|
165	\| Affinity prediction \| No \| Yes (small-molecule binding) \|
166	\| Release \| 2024 \| 2025 \|
167
168	Boltz-2 is the current default. Boltz-1 is still used where a design pipeline
169	inverts the v1 model.
170
171	### Affinity prediction (Boltz-2)
172
173	Boltz-2 adds an affinity-prediction module that approaches free-energy-perturbation
174	accuracy at a fraction of the cost. It is trained on small-molecule binding data, so
175	use it for protein-ligand and small-molecule work. It does not predict
176	protein-protein binding affinity; for protein binders, rely on interface confid

People who liked this also viewed

6 picks

Type