generative-design

Name: generative-design
Author: GPTomics

1.1kstars185forks

How to install generative-design?

gptomics/bioskills/generative-design

$npx -y skills add GPTomics/bioSkills --skill generative-design

Installs into the current project.

Summary

Designs novel molecules using REINVENT 4 (de novo, scaffold decoration, linker design, R-group, molecular optimization), MolMIM, Diffusion-based generators (DiGress, DiffSMol), and JT-VAE with explicit handling of multi-parameter optimization (MPO), goal-directed scoring function

Files · 1

View on GitHub

SKILL.md

1	## Version Compatibility
2
3	Reference examples tested with: REINVENT 4.0+, RDKit 2024.09+, PyTorch 2.1+, MolMIM (NVIDIA BioNeMo), chemprop 2.0+.
4
5	Before using code patterns, verify installed versions match. If versions differ:
6	- Python: `pip show <package>` then `help(module.function)` to check signatures
7
8	If code throws ImportError, AttributeError, or TypeError, introspect the installed
9	package and adapt the example to match the actual API rather than retrying.
10
11	# Generative Molecular Design
12
13	Generate novel molecules biased toward desired properties using deep generative models. REINVENT 4 (Loeffler et al. 2024, AstraZeneca) provides four generator families: Reinvent (de novo), Libinvent (scaffold decoration and library design), Linkinvent (linker design), and Mol2Mol (similarity-constrained molecular optimization). These support design tasks including R-group replacement and scaffold hopping and can be used with transfer learning, reinforcement learning, and curriculum learning. For specific niches: MolMIM (NVIDIA BioNeMo) for latent-space property optimization, DiffSMol / DiGress for diffusion-based generation, and JT-VAE for latent-space optimization. The art of generative design is in the scoring function: poorly-designed scoring rewards uninteresting molecules, while well-designed scoring captures both activity and developability.
14
15	For QSAR/scoring models that feed generative design, see `chemoinformatics/qsar-modeling`. For synthetic feasibility, see `chemoinformatics/retrosynthesis`. For library enumeration as alternative, see `chemoinformatics/reaction-enumeration`.
16
17	## Generator Mode Taxonomy
18
19	\| Mode \| Input \| Output \| Use case \| Fails when \|
20	\|------\|-------\|--------\|----------\|------------\|
21	\| De novo \| Empty seed or training set \| Novel molecules \| Wide chemical space exploration \| Synthetic feasibility weak \|
22	\| Scaffold decoration \| Scaffold + attachment points \| Decorated molecules \| Series expansion \| Generation diversity limited by scaffold \|
23	\| Linker design \| 2 fragments \| Linker molecules \| PROTAC, ternary complex \| Few linker geometric options \|
24	\| R-group replacement \| Scaffold + existing R-groups \| New R-group set \| Optimize one position \| Single-position only \|
25	\| Molecular optimization \| Lead molecule \| Improved analogs \| Lead optimization \| Improvement window narrow \|
26	\| Constrained generation \| Hard constraints (MW, fragments) \| Compliant molecules \| Patent / IP design \| Constraints overly restrictive \|
27
28	## Learning Algorithm Taxonomy
29
30	\| Algorithm \| Use \| Pro \| Con \|
31	\|-----------\|-----\|-----\|-----\|
32	\| Transfer learning (TL) \| Adapt prior model to focused training set \| Stable, simple \| Limited optimization power \|
33	\| Reinforcement learning (RL) \| Reward-driven generation \| Powerful for MPO \| Reward hacking risk \|
34	\| Curriculum learning (CL) \| Gradual constraint introduction \| Better convergence \| Slower; tuning sensitive \|
35
36	## Decision Tree by Scenario
37
38	\| Scenario \| Generator \| Algorithm \| Scoring \|
39	\|----------\|-----------\|-----------\|---------\|
40	\| New target, no SAR \| De novo \| Benchmark RL against simpler search \| Validated target evidence + developability objectives \|
41	\| Series expansion \| Scaffold decoration \| TL on series + RL \| QSAR ensemble + QED \|
42	\| PROTAC linker \| Linker design \| Project-specific constrained workflow \| Validated geometry/ternary-complex evidence; no generic DC50 surrogate \|
43	\| Lead optimization MPO \| Molecular optimization \| CL with staged constraints \| Multi-task: activity + ADMET \|
44	\| Diverse hit set \| De novo with diversity bonus \| RL + Tanimoto distance to known \| Activity + diversity \|
45	\| Patent space carve-out \| Constrained de novo \| RL + structural constraints \| Activity + novelty \|
46	\| Hit-to-lead \| R-group replacement \| TL on lead + RL \| Activity + Lipinski \|
47	\| ADMET-aware design \| De novo or optimization \| RL \| hERG + CYP + AMES + QED \|
48
49	## REINVENT 4 Setup
50
51	REINVENT 4 uses a TOML configuration file specifying generator, algorithm, prior model, and scoring functions.
52
53	Goal: Configure a reinforcement-learning REINVENT 4 run with a prior, agent, sampling parameters, and a QED scoring component.
54
55	Approach: Build a release-matched REINVENT 4 staged-learning TOML config with `[parameters]` for the prior/agent checkpoints, `[[stage]]` blocks, and one or more `[[stage.scoring.component]]` blocks.

People who liked this also viewed

6 picks

Type

generative-design

1.1kstars185forks

View on GitHub

How to install generative-design?

gptomics/bioskills/generative-design

$npx -y skills add GPTomics/bioSkills --skill generative-design

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	## Version Compatibility
2
3	Reference examples tested with: REINVENT 4.0+, RDKit 2024.09+, PyTorch 2.1+, MolMIM (NVIDIA BioNeMo), chemprop 2.0+.
4
5	Before using code patterns, verify installed versions match. If versions differ:
6	- Python: `pip show <package>` then `help(module.function)` to check signatures
7
8	If code throws ImportError, AttributeError, or TypeError, introspect the installed
9	package and adapt the example to match the actual API rather than retrying.
10
11	# Generative Molecular Design
12
13	Generate novel molecules biased toward desired properties using deep generative models. REINVENT 4 (Loeffler et al. 2024, AstraZeneca) provides four generator families: Reinvent (de novo), Libinvent (scaffold decoration and library design), Linkinvent (linker design), and Mol2Mol (similarity-constrained molecular optimization). These support design tasks including R-group replacement and scaffold hopping and can be used with transfer learning, reinforcement learning, and curriculum learning. For specific niches: MolMIM (NVIDIA BioNeMo) for latent-space property optimization, DiffSMol / DiGress for diffusion-based generation, and JT-VAE for latent-space optimization. The art of generative design is in the scoring function: poorly-designed scoring rewards uninteresting molecules, while well-designed scoring captures both activity and developability.
14
15	For QSAR/scoring models that feed generative design, see `chemoinformatics/qsar-modeling`. For synthetic feasibility, see `chemoinformatics/retrosynthesis`. For library enumeration as alternative, see `chemoinformatics/reaction-enumeration`.
16
17	## Generator Mode Taxonomy
18
19	\| Mode \| Input \| Output \| Use case \| Fails when \|
20	\|------\|-------\|--------\|----------\|------------\|
21	\| De novo \| Empty seed or training set \| Novel molecules \| Wide chemical space exploration \| Synthetic feasibility weak \|
22	\| Scaffold decoration \| Scaffold + attachment points \| Decorated molecules \| Series expansion \| Generation diversity limited by scaffold \|
23	\| Linker design \| 2 fragments \| Linker molecules \| PROTAC, ternary complex \| Few linker geometric options \|
24	\| R-group replacement \| Scaffold + existing R-groups \| New R-group set \| Optimize one position \| Single-position only \|
25	\| Molecular optimization \| Lead molecule \| Improved analogs \| Lead optimization \| Improvement window narrow \|
26	\| Constrained generation \| Hard constraints (MW, fragments) \| Compliant molecules \| Patent / IP design \| Constraints overly restrictive \|
27
28	## Learning Algorithm Taxonomy
29
30	\| Algorithm \| Use \| Pro \| Con \|
31	\|-----------\|-----\|-----\|-----\|
32	\| Transfer learning (TL) \| Adapt prior model to focused training set \| Stable, simple \| Limited optimization power \|
33	\| Reinforcement learning (RL) \| Reward-driven generation \| Powerful for MPO \| Reward hacking risk \|
34	\| Curriculum learning (CL) \| Gradual constraint introduction \| Better convergence \| Slower; tuning sensitive \|
35
36	## Decision Tree by Scenario
37
38	\| Scenario \| Generator \| Algorithm \| Scoring \|
39	\|----------\|-----------\|-----------\|---------\|
40	\| New target, no SAR \| De novo \| Benchmark RL against simpler search \| Validated target evidence + developability objectives \|
41	\| Series expansion \| Scaffold decoration \| TL on series + RL \| QSAR ensemble + QED \|
42	\| PROTAC linker \| Linker design \| Project-specific constrained workflow \| Validated geometry/ternary-complex evidence; no generic DC50 surrogate \|
43	\| Lead optimization MPO \| Molecular optimization \| CL with staged constraints \| Multi-task: activity + ADMET \|
44	\| Diverse hit set \| De novo with diversity bonus \| RL + Tanimoto distance to known \| Activity + diversity \|
45	\| Patent space carve-out \| Constrained de novo \| RL + structural constraints \| Activity + novelty \|
46	\| Hit-to-lead \| R-group replacement \| TL on lead + RL \| Activity + Lipinski \|
47	\| ADMET-aware design \| De novo or optimization \| RL \| hERG + CYP + AMES + QED \|
48
49	## REINVENT 4 Setup
50
51	REINVENT 4 uses a TOML configuration file specifying generator, algorithm, prior model, and scoring functions.
52
53	Goal: Configure a reinforcement-learning REINVENT 4 run with a prior, agent, sampling parameters, and a QED scoring component.
54
55	Approach: Build a release-matched REINVENT 4 staged-learning TOML config with `[parameters]` for the prior/agent checkpoints, `[[stage]]` blocks, and one or more `[[stage.scoring.component]]` blocks.

People who liked this also viewed

6 picks

Type