Research map / Probabilistic and scientific modeling / Molecules and proteins
Robust Model-Based Optimization for Challenging Fitness Landscapes
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 PPGVAE, a VAE whose latent space is structured by fitness so higher-fitness samples lie nearer the prior origin, enabling robust model-based optimization under data imbalance and separation in design space.
Problem
Model-based optimization for protein design fails when high-fitness training samples are rare and located far (in design space) from the abundant low-fitness samples; existing weighted-MLE generative methods are not robust to this imbalance and separation.
Approach
A VAE whose latent space is regularized by a relationship loss so that log prior probability of each encoded sample is linearly tied to its property value, making high-fitness samples lie closer to the origin and thus more likely under N(0,I); this avoids sample weighting and keeps all samples in training.
Key results
- PPGVAE maintains robust relative improvement across imbalance ratios and separation levels, while baselines (CbAS, RWR, CEM-PI) degrade as separation increases.
- CEM-PI is the most competitive baseline at low separation but its performance decays for small imbalance ratios as separation grows.
- PPGVAE converges to maximum relative improvement in fewer MBO steps and with smaller sampling budgets than baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PPGVAE instead of weighting-based methods | MBO · GMM | ΔYmax | Helped |
| Use PPGVAE instead of weighting-based methods | MBO · AAV | ΔYmax | Helped |
| Use PPGVAE instead of weighting-based methods | MBO · GB1 | ΔYmax | Helped |
| Use PPGVAE instead of weighting-based methods | MBO · PhoQ | ΔYmax | Helped |
| Use PPGVAE instead of weighting-based methods | MBO · PINN Poisson solutions | ΔYmax | Helped |
| Use PPGVAE instead of other methods | MBO · Benchmarked datasets | MBO steps to maximum relative improvement | Helped |
| Increase the imbalance ratio (more high-fitness samples in training) | MBO methods · All benchmark datasets | ΔYmax | Helped |
| Increase separation between low- and high-fitness samples | Weighting-based methods · All benchmark datasets | ΔYmax | Hurt |
| Reduce the sample-generation budget per MBO step (ablation) | PPGVAE · AAV | ΔYmax | Hurt |
| Reduce the sample-generation budget per MBO step (ablation) | PPGVAE · GB1 and PhoQ | ΔYmax | No clear effect |
| Increase sampling budget per MBO step (ablation) | PPGVAE and CEM-PI · AAV | ΔYmax | Helped |
| Vary temperature τ across log10(τ) ∈ [-1,1] (ablation) | PPGVAE · GMM, lowest imbalance and highest separation | ΔYmax | No clear effect |
Surprising findings
- Weighted-MLE methods can fail even with the same generative architecture because weights reduce effective sample size under severe imbalance.
- Training CEM-PI only on high-fitness samples helps it at higher separation, suggesting filtering can benefit weighting-based methods on harder tasks.
Limitations
- Does not use an additional exploration mechanism, so improvements rely on the latent structure alone.
- Robustness to oracle uncertainty is not addressed.
- The structural constraint may be restrictive for some design problems.
Experimental setup
Datasets: AAV GB1 PhoQ PINN-derived Poisson solutions GMM synthetic oracle toy MNIST
Benchmarks: GMM benchmark AAV GB1 (transformed) PhoQ (transformed) PINN Poisson solutions
Metrics: relative improvement of highest property (ΔY_max)
Baselines: CbAS RWR CEM-PI AdaLead Bombarelli VAE
Builds on: CbAS reward-weighted regression CEM-PI AdaLead Bombarelli VAE
Closest papers by meaning
- Improving protein optimization with smoothed fitness landscapes (ICLR 2024 · similarity 0.89)
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.88)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.88)
- A Variational Perspective on Generative Protein Fitness Optimization (ICML 2025 · similarity 0.88)
- Multi-Scale Representation Learning for Protein Fitness Prediction (NeurIPS 2024 · similarity 0.86)
- Bridging Model-Based Optimization and Generative Modeling via Conservative Fine-Tuning of Diffusion Models (NeurIPS 2024 · similarity 0.86)
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.