[Paper Review] AdaLead: A simple and robust adaptive greedy search algorithm for sequence design

Sam Sinai, Richard Wang+4 more|arXiv (Cornell University)|Oct 5, 2020

Evolutionary Algorithms and Applications41 references39 citations

TL;DR

AdaLead introduces a simple adaptive greedy algorithm for sequence design that competes with state-of-the-art methods, evaluated in the FLEXS environment across TF, RNA, and protein design tasks. It demonstrates robustness and consistency, performing well even with imperfect models.

ABSTRACT

Efficient design of biological sequences will have a great impact across many industrial and healthcare domains. However, discovering improved sequences requires solving a difficult optimization problem. Traditionally, this challenge was approached by biologists through a model-free method known as "directed evolution", the iterative process of random mutation and selection. As the ability to build models that capture the sequence-to-function map improves, such models can be used as oracles to screen sequences before running experiments. In recent years, interest in better algorithms that effectively use such oracles to outperform model-free approaches has intensified. These span from approaches based on Bayesian Optimization, to regularized generative models and adaptations of reinforcement learning. In this work, we implement an open-source Fitness Landscape EXploration Sandbox (FLEXS: github.com/samsinai/FLEXS) environment to test and evaluate these algorithms based on their optimality, consistency, and robustness. Using FLEXS, we develop an easy-to-implement, scalable, and robust evolutionary greedy algorithm (AdaLead). Despite its simplicity, we show that AdaLead is a remarkably strong benchmark that out-competes more complex state of the art approaches in a variety of biologically motivated sequence design challenges.

Motivation & Objective

• Motivate efficient design of biological sequences by leveraging model-guided exploration.
• Propose AdaLead, a simple adaptive greedy algorithm for sequence design.
• Benchmark AdaLead against state-of-the-art methods across diverse landscapes (TF binding, RNA, protein design).
• Evaluate robustness, consistency, and batch-sample efficiency using the FLEXS environment.

Proposed method

• Define an exploration framework with a ground-truth oracle phi and a surrogate phi' trained on available data.
• Introduce AdaLead: seed generation from near-maxima, recombination, rollout mutations, and selection by phi' over a batch.
• Use a threshold parameter kappa to control greediness and promote diversity when the surface is flat.
• Employ an abstract noise-corrupted model phi'_alpha to study robustness and consistency independent of a specific empirical model.
• Evaluate AdaLead with an ensemble of models (phi''), and compare to DynaPPO, CbAS/DbAS, CMA-ES across landscapes.
• Provide an open-source simulation environment FLEXS for benchmarking and reproducibility.

Experimental results

Research questions

• RQ1 How does AdaLead perform relative to other state-of-the-art sequence-design algorithms under model misspecification and noise?
• RQ2 Is AdaLead robust when the surrogate model is poor or biased?
• RQ3 Does AdaLead maintain consistency and diversity as the underlying model quality improves?
• RQ4 How does AdaLead scale and perform across different biologically motivated landscapes (TF binding, RNA, protein design)?

Key findings

• AdaLead frequently finds high-performing peaks more consistently than competing methods.
• In RNA landscapes, AdaLead often finds global optima and outperforms others on complex, highly epistatic surfaces.
• AdaLead shows robustness, maintaining strong performance even with uninformative models and ensemble CNN surrogates.
• Across TF, RNA, and protein design tasks, AdaLead demonstrates competitive optimization and favorable time evolution of best-found sequences.