Research map / Probabilistic and scientific modeling / Molecules and proteins
Hybrid Generative AI for De Novo Design of Co-Crystals with Enhanced Tabletability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents GEMCODE, a pipeline that hybridizes deep generative models (LSTM GAN, transformer VAE/CVAE) with multi-objective evolutionary optimization and ML tabletability predictors plus CCGNet to de novo design coformers for drug co-crystals with improved tabletability.
Problem
De novo design of pharmaceutical co-crystal coformers with predefined mechanical (plasticity/tabletability) properties, which is hard to address by experimental high-throughput screening.
Approach
Combine conditional transformer VAE generation conditioned on gradient-boosting-predicted plasticity parameters, multi-objective MOEA/D evolutionary refinement guided by those predictors, and a pretrained GNN (CCGNet) to rank drug-coformer pairs by co-crystallization probability.
Key results
- T-CVAE produced 5.63% target coformers versus 2.23% for GAN and 1.68% for T-VAE when generating 10,000 candidates.
- Evolutionary optimization increased median probability of the target class for orthogonal planes and H-bond bridging, e.g., H-bond bridging for GAN rose from 0.62 to 0.69 (+11.3%).
- Gradient boosting was the best mechanical property classifier, and CCGNet is reported with 98.6% average balanced accuracy for co-crystal formation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| condition generation on target mechanical properties using T-CVAE instead of T-VAE | Transformer coformer generation · Coformer generation | target coformers (%): 1.68% -> 5.63% | Helped |
| condition generation on target mechanical properties using T-CVAE instead of T-VAE | Transformer coformer generation · Coformer generation | novelty (%): 95.12 ± 0.11 -> 80.62 ± 0.25 | Hurt |
| condition generation on target mechanical properties using T-CVAE instead of T-VAE | Transformer coformer generation · Coformer generation | duplicates (%): 24.30 ± 0.45 -> 55.70 ± 0.19 | Hurt |
| apply evolutionary optimization (ablation) | GAN-generated coformers · GAN, unobstructed planes | median probability: 0.82 -> 0.82 (+0.0%) | No clear effect |
| apply evolutionary optimization (ablation) | GAN-generated coformers · GAN, orthogonal planes | median probability: 0.37 -> 0.39 (+5.4%) | Helped |
| apply evolutionary optimization (ablation) | GAN-generated coformers · GAN, H-bond bridging | median probability: 0.62 -> 0.69 (+11.3%) | Helped |
| apply evolutionary optimization (ablation) | T-VAE-generated coformers · T-VAE, unobstructed planes | median probability: 0.82 -> 0.82 (+0.0%) | No clear effect |
| apply evolutionary optimization (ablation) | T-VAE-generated coformers · T-VAE, orthogonal planes | median probability: 0.38 -> 0.40 (+5.3%) | Helped |
| apply evolutionary optimization (ablation) | T-VAE-generated coformers · T-VAE, H-bond bridging | median probability: 0.64 -> 0.69 (+7.8%) | Helped |
| apply evolutionary optimization (ablation) | T-CVAE-generated coformers · T-CVAE, unobstructed planes | median probability: 0.82 -> 0.83 (+1.2%) | Helped |
| apply evolutionary optimization (ablation) | T-CVAE-generated coformers · T-CVAE, orthogonal planes | median probability: 0.38 -> 0.39 (+2.6%) | Helped |
| apply evolutionary optimization (ablation) | T-CVAE-generated coformers · T-CVAE, H-bond bridging | median probability: 0.64 -> 0.69 (+7.8%) | Helped |
Surprising findings
- Fine-tuned Llama-3-8B with LoRA produced the highest diversity among tested approaches but only 0.34% target molecules, and GPT-2 reached 3.32% target molecules, comparable to GAN and T-VAE.
- The transformer CVAE, despite producing the most target coformers, was about 100x slower at generation than the GAN, making it impractical for many users.
Limitations
- Coformer dataset is small, limiting the molecular space explored.
- Gradient boosting model is biased toward predicting absence of orthogonal planes, causing false negatives.
- Predictions are computational only; no experimental co-crystal synthesis or tableting validation is performed in this work.
Experimental setup
Datasets: ChEMBL open co-crystal coformer dataset (6819 co-crystals) Cambridge Structural Database (CSD)
Benchmarks: coformer generation comparison (validity, novelty, duplicates, target coformers, diversity) mechanical property classification (accuracy, F1)
Metrics: validity novelty duplicates target coformer percentage diversity accuracy F1 median probability Mann-Whitney p-value CCGNet score
Baselines: GAN (LSTM) T-VAE GPT-2 SPEA-2
Builds on: MOEA/D SPEA-2 CCGNet GOLEM conditional VAE
Closest papers by meaning
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.86)
- Generative Hierarchical Materials Search (NeurIPS 2024 · similarity 0.85)
- De Novo Molecular and Crystal Design with Latent Space Bayesian Optimization (AAAI 2025 · similarity 0.85)
- CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models (NeurIPS 2026 · similarity 0.85)
- Fast Organic Crystal Structure Prediction with Unit Cell Flow Matching (NeurIPS 2026 · similarity 0.84)
- Scalable Diffusion for Materials Generation (ICLR 2024 · 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.