Research map / Probabilistic and scientific modeling / Molecules and proteins
Implicitly Guided Design with PropEn: Match your Data to Follow the Gradient
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 PropEn, which matches each training sample to a similar sample with higher property value and trains an encoder-decoder on the matched pairs to get implicit guidance for design optimization without a discriminator.
Problem
Guided design with limited data and non-smooth property landscapes, where explicit guidance (generator plus surrogate discriminator) is data-hungry and can drive designs off-distribution.
Approach
Match each sample with a nearby sample that has a better property value to build a much larger paired dataset; training an encoder-decoder to reconstruct the better matched sample implicitly learns the property gradient direction, so no discriminator is needed.
Key results
- On NACA airfoils with N=200, PropEn xy2xy reached 55.41 average improvement vs 15.04 for explicit guidance.
- In wet-lab antibody optimization, PropEn achieved 94.6% binding rate over 93 designs versus 62.9% for walk-jump and 44.2% for LaMBO.
- On toy data, PropEn improved up to ~85% of holdout designs versus ~50% for explicit AE guidance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PropEn x2x instead of explicit guidance | NACA airfoil optimization · NACA airfoils (N=500) | RI: 59.61 \pm 40.83 -> 96.31 $\pm$ 1.08 | Helped |
| use PropEn x2x instead of explicit guidance | NACA airfoil optimization · NACA airfoils (N=200) | RI: 6.53 \pm 6.35 -> 38.83 $\pm$ 8.80 | Helped |
| add reconstruction regularizer (ablation) | PropEn x2x · NACA airfoils (N=500) | RI: 96.31 $\pm$ 1.08 -> 74.06 \pm 14.65 | Hurt |
| add reconstruction regularizer (ablation) | PropEn x2x · NACA airfoils (N=200) | RI: 38.83 $\pm$ 8.80 -> 7.23 \pm 6.21 | No clear effect |
| use PropEn instead of walk-jump | Therapeutic antibody design · Therapeutic antibodies overall | binding rate: 62.9\% (105) -> 94.6\% (93) | Helped |
| use PropEn instead of diffusion (guided) | Therapeutic antibody design · Therapeutic antibodies overall | fraction improving seed: 22.6\% (84) -> 34.4\% (93) | Helped |
| use PropEn x2x instead of Explicit | Toy design optimization · 8-Gaussians (N=200, d=50) | ratio imp: 48.02 \pm 4.15 -> 94.36 \pm 5.64 | Helped |
| use PropEn x2x instead of Explicit | Toy design optimization · Pinwheel (N=200, d=50) | ratio imp: 50.12 \pm 4.64 -> 92.67 \pm 6.69 | Helped |
| add reconstruction regularizer (ablation) | PropEn x2x · 8-Gaussians (N=200, d=50) | uniqueness: 57.95 \pm 12.69 -> 77.78 \pm 6.01 | Helped |
| reconstruct both input features and property instead of input features only (ablation) | PropEn · 8-Gaussians (N=200, d=50) | ratio imp: 94.36 \pm 5.64 -> 78.39 \pm 6.92 | Hurt |
| reconstruct both input features and property, and add reconstruction regularizer (ablation) | PropEn · 8-Gaussians (N=200, d=50) | ratio imp: 94.36 \pm 5.64 -> 75.78 \pm 13.27 | Hurt |
Surprising findings
- Explicit guidance methods did not appear among the top three in binding affinity improvement in wet-lab results.
- PropEn xy2xy did not outperform PropEn x2x as sample size increased in toy experiments.
Limitations
- Supports only single-property optimization
- Requires a sufficient amount of measured training data, not zero-shot
- Matching distance metric choice requires domain knowledge and adds overhead
Experimental setup
Datasets: pinwheel toy data 8-Gaussians toy data NACA 4-digit airfoil coordinates in-house antibody SPR binding affinity data (HER2, T1, T2)
Benchmarks: pinwheel 8-Gaussians NACA airfoil Cl/Cd therapeutic antibody wet-lab binding
Metrics: ratio of improvement average improvement uniqueness novelty log-likelihood binding rate fraction improving seed pKD improvement
Baselines: Explicit guidance (AE with discriminator) walk-jump sampler LaMBO diffusion diffusion (guided) VAE-like baseline
Builds on: variational diffusion model matching estimators from causal inference iterative latent variable refinement
Closest papers by meaning
- Reliable Algorithm Selection for Machine Learning-Guided Design (ICML 2025 · similarity 0.85)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.84)
- Robust Model-Based Optimization for Challenging Fitness Landscapes (ICLR 2024 · similarity 0.84)
- Training Free Guided Flow-Matching with Optimal Control (ICLR 2025 · similarity 0.83)
- Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning (NeurIPS 2025 · similarity 0.83)
- BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.83)
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.