Research map / Probabilistic and scientific modeling / Molecules and proteins
UniGEM: A Unified Approach to Generation and Property Prediction for Molecules
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.
Proposes UniGEM, a diffusion-based unified model that generates atomic coordinates and predicts atom types and properties in a two-phase (nucleation/growth) process, using oversampling and a multi-branch EGNN to jointly improve 3D molecule generation and QM9 property prediction.
Problem
Molecular generation and property prediction are developed separately, and naive multi-task training of a diffusion model with property prediction degrades both tasks due to inconsistency between noisy and structured molecules.
Approach
Split the diffusion trajectory into a nucleation phase (coordinate denoising only) and a growth phase (where atom-type and property prediction losses are active), diffuse only coordinates while predicting atom types as a separate task, oversample growth-phase timesteps, and use a multi-branch EGNN so the two phases do not interfere.
Key results
- On QM9 generation, UniGEM reaches 89.8% molecule stability and 95.0% validity versus 82.0% and 91.9% for EDM.
- On six QM9 property prediction tasks, UniGEM achieves lower MAE than EGNN from scratch on all six (e.g., epsilon_LUMO 16.7 meV vs 25 meV), and beats listed pre-training methods on most properties.
- With BFN as the generator, UniGEM reaches 93.7% molecule stability on QM9, above GeoBFN's 90.9%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniGEM instead of GeoLDM | Molecular generation · QM9 | Validity (%): 93.8 -> 95.0 | Helped |
| Use UniGEM instead of GeoLDM | Molecular generation · GEOM-Drugs | Validity (%): 99.3 -> 98.4 | Hurt |
| Use UniGEM instead of 3D-EMGP | QM9 property prediction · QM9 | εLUMO MAE (meV): 18.2 -> 16.7 | Helped |
| Jointly train generation and property prediction | UniGEM framework · QM9 | Validity × uniqueness (%): 90.7 -> 88.8 | Hurt |
| Add property prediction to a frozen pretrained generation backbone | UniGEM framework · QM9 | εLUMO MAE: 25 -> 51.2 | Hurt |
| Add property prediction loss (ablation) | UniGEM · QM9 | Molecule stability (%): 82.0 -> 85.7 | Helped |
| Add atom type prediction loss (ablation) | UniGEM · QM9 | Molecule stability (%): 82.0 -> 89.3 | Helped |
| Add atom type prediction loss to property prediction loss (ablation) | UniGEM · QM9 | εLUMO MAE: 20.6 -> 16.7 | Helped |
| Oversample growth-phase time steps (ablation) | Single-branch UniGEM · QM9 | εLUMO MAE: 40.5 -> 19.2 | Helped |
| Oversample growth-phase time steps (ablation) | Single-branch UniGEM · QM9 | Validity × uniqueness (%): 92.9 -> 88.8 | Hurt |
| Use separate branches for different time-step ranges (ablation) | Oversampled UniGEM · QM9 | Validity × uniqueness (%): 88.8 -> 93.2 | Helped |
| Replace EDM coordinate generation with BFN (ablation) | UniGEM · QM9 | Molecule stability (%): 89.8 -> 93.7 | Helped |
Surprising findings
- Naive multi-task training of generation and property prediction degrades both tasks, and freezing a pretrained generator still fails to improve property prediction.
- Generating only coordinates and predicting atom types as a separate task yields lower atom-type error than joint diffusion (1.3E-5 vs 0.0077 in the reported table).
Limitations
- Nucleation time t_n is a hand-set hyperparameter that must be chosen from the noise schedule.
- Evaluation is limited to QM9 for property prediction (six properties) and relies on EDM-style bond-distance metrics for generation.
- Improvements on GEOM-Drugs validity are not reported for property prediction; the paper states the principles may extend to NLP and vision without experiments.
Experimental setup
Datasets: QM9 GEOM-Drugs
Benchmarks: QM9 GEOM-Drugs
Metrics: atom stability molecule stability validity validity*uniqueness MAE
Baselines: EDM GDM EDM-Bridge GeoLDM G-Schnet E-NF GeoBFN EGNN GraphMVP 3D Infomax GEM 3D-EMGP
Builds on: EDM EGNN GeoLDM GeoBFN InfoMax
Closest papers by meaning
- Unified Guidance for Geometry-Conditioned Molecular Generation (NeurIPS 2024 · similarity 0.86)
- Text-Guided Molecule Generation with Diffusion Language Model (AAAI 2024 · similarity 0.85)
- Geometric Representation Condition Improves Equivariant Molecule Generation (ICML 2025 · similarity 0.85)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.85)
- Data-Efficient Molecular Generation with Hierarchical Textual Inversion (ICML 2024 · similarity 0.85)
- Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.85)
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.