Research map / Probabilistic and scientific modeling / Molecules and proteins
A Chance-Constrained Generative Framework for Sequence Optimization
Xianggen Liu, Qiang Liu, Sen Song, Jian Peng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace chance constraint with learnable alpha (full CCGF) instead of linear validity reward (RL-Linear) (ablation) | sequence generator training · ZINC penalized LogP | Top-50 average penalized LogP: 10.15 vs 8.81 | Improved |
| Learn alpha in the chance-constraint bound (full CCGF) instead of fixing alpha (ablation) | CCGF · ZINC penalized LogP | Top-50 average penalized LogP: 10.15 vs 8.53 | Improved |
| Learn alpha (full CCGF) instead of fixed alpha (ablation) | CCGF · Python programs | Top-50 average score: -0.113 vs -0.139 | Improved |
| Replace chance-constrained objective with linear validity reward (RL-Linear) (ablation) | CCGF · Python programs | Top-50 average score: -0.131 vs -0.113 | Hurt |
| Use chance-constrained RL objective (full CCGF) instead of pure MLE training (ablation) | CCGF · ZINC penalized LogP | Top-50 average penalized LogP: 10.15 vs 3.17 | Improved |
| Chance-constrained CCGF vs sequence baselines (GVAE, SD-VAE, ORGAN) on molecule penalized LogP | sequence generator · ZINC penalized LogP | Top-1 penalized LogP: 12.32 vs 4.04 (SD-VAE best sequence baseline) | Improved |
| CCGF vs graph-based GraphAF on penalized LogP | sequence generator · ZINC penalized LogP | Top-1 penalized LogP: 12.32 vs 12.23 top-1; 10.15 vs GraphAF top-50 avg (not reported) | Mixed |
| CCGF vs GraphAF under Lipinski constraint | sequence generator · Lipinski-constrained solubility | Top-1 penalized LogP and Lipinski validity: 4.30 vs 4.15; 98.9% vs 97.4% | Improved |
| Chance-constrained CCGF vs grammar-based and VAE sequence methods on arithmetic expressions | sequence generator · Arithmetic expressions | Top-1 property score and validity: -0.04 vs -0.04 (GVAE) top-1; avg -0.52 vs -3.47; validity 99% vs 99% | Improved |
| CCGF vs SD-VAE on python programs | sequence generator · Python programs | Top-1 property score: -0.048 vs -0.121 | Improved |
| Add diversity term (adversarial discriminator) to property objective | CCGF · ZINC molecules diversity | Diversity (MACCS, PairFp, MorganFp, TopologicalFp) | Mixed |
Closest papers by meaning
- Variational Search Distributions (ICLR 2025 · similarity 0.84)
- How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data (ICLR 2024 · similarity 0.84)
- Constrained Discrete Diffusion (NeurIPS 2025 · similarity 0.84)
- Discrete Markov Probabilistic Models: An Improved Discrete Score-Based Framework with sharp convergence bounds under minimal assumptions (ICML 2025 · similarity 0.84)
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.84)
- Primal-Dual Guided Decoding for Constrained Discrete Diffusion (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
