Sequence-free landscape inference for directed evolution
- 1Department of Engineering Science, University of Oxford, Oxford, United Kingdom
Abstract
Directed evolution is a method for engineering biological systems or components, such as proteins, wherein desired traits are optimised through iterative rounds of mutagenesis and selection of fit variants. The process of protein directed evolution can be envisaged as navigation over high-dimensional optimisation landscapes with numerous local maxima. The performance of any strategy in navigating such a landscape is dependent on the ruggedness of that landscape. However, this information is generally unavailable at the outset of an experiment. Here we propose SLIDE, Sequence-free Landscape Inference for Directed Evolution, which consists of two parts. First, SLIDE provides an estimation of landscape ruggedness from a mutating population using only population-level phenotypic data and an estimate of the mutation rate. Such ruggedness information in itself is valuable in protein design, for instance in predicting evolutionary stability. Second, SLIDE offers a framework for using the estimated ruggedness metric to identify high-performing selection strategies for directed evolution. Using theoretical NK landscapes and four empirical protein fitness landscapes, we demonstrate consistent in silico improvement upon the performance of fixed-parameter strategies, using a pipeline that could also be combined with emerging AI-based methods for driving directed evolution. Author summary: Directed evolution (DE) navigates high-dimensional protein fitness landscapes to find improved variants, but the success of any strategy depends on the landscape’s ruggedness, which is typically unknown and often requires genetic sequencing to derive. We present Sequence-free Landscape Inference for Directed Evolution (SLIDE), a two-step framework that (i) estimates landscape ruggedness from population-level phenotypic decay curves together with a mutation rate estimate, and (ii) uses that ruggedness estimate to choose DE control parameters that better balance exploration and exploitation. In contrast to existing ruggedness metrics, SLIDE relies on spectral analysis to link the exponential decay rate of mean fitness under random, unbiased mutations to the dominant frequency content of the landscape. This decay rate is related to the normalised Dirichlet energy of the underlying genotype graph, and the framework naturally extends to biased mutation spectra. On homogeneous landscapes, such as NK landscapes, ruggedness can be accurately estimated from a single starting genotype. Heterogeneous empirical landscapes instead require sampling multiple starting points, naturally distinguishing local from global ruggedness estimates. We combine SLIDE with DE strategies developed in our previous work, but the method is strategy-agnostic and can be paired with any DE workflow, including emerging AI-driven workflows. In silico tests show that SLIDE reliably recovers landscape ruggedness and enables more effective selection of directed evolution strategies than fixed-parameter approaches across landscapes of varying ruggedness. Although heterogeneous landscapes require fitness decay measurements from multiple starting points, these can be generated in parallel using high-throughput mutagenesis methods, such as error-prone PCR or chemical mutagenesis, thereby avoiding the need for extensive sequencing or combinatorially complete genotype–phenotype maps while remaining scalable to large screens. Beyond DE, the approach can rapidly characterise landscapes arising in genetic circuits, enzyme libraries, and ecological models, providing a way to quantify genetic stability and guide experimental or computational search.
Related
- Directed evolution of a thermostable family-A DNA polymerase into a reverse transcriptase for 2'-modified RNA
- Optimization of landscape connectivity: A directed dynamic ecological network topology perspective
- GLOM TGITO: multiscale protein modelling and multiview landscape-guided sequence discovery
- From single-sequence structure prediction to protein fitness landscape through a composable, epistasis-aware mutation atlas
- Joint ancestry inference reveals the landscape of archaic introgression in admixed populations
BibTeX
@article{towers2026sequencefree,
title = {{Sequence-free landscape inference for directed evolution}},
author = {Sebastian Towers and Jessica James and Harrison Steel and Idris Kempf},
journal = {PLOS Computational Biology},
year = {2026},
month = sep,
doi = {10.1371/journal.pcbi.1014713},
url = {https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014713},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Sequence-free landscape inference for directed evolution
AU - Sebastian Towers
AU - Jessica James
AU - Harrison Steel
AU - Idris Kempf
PY - 2026
DA - 2026/09/03
JO - PLOS Computational Biology
DO - 10.1371/journal.pcbi.1014713
UR - https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014713
AB - Directed evolution is a method for engineering biological systems or components, such as proteins, wherein desired traits are optimised through iterative rounds of mutagenesis and selection of fit variants. The process of protein directed evolution can be envisaged as navigation over high-dimensional optimisation landscapes with numerous local maxima. The performance of any strategy in navigating such a landscape is dependent on the ruggedness of that landscape. However, this information is generally unavailable at the outset of an experiment. Here we propose SLIDE, Sequence-free Landscape Inference for Directed Evolution, which consists of two parts. First, SLIDE provides an estimation of landscape ruggedness from a mutating population using only population-level phenotypic data and an estimate of the mutation rate. Such ruggedness information in itself is valuable in protein design, for instance in predicting evolutionary stability. Second, SLIDE offers a framework for using the estimated ruggedness metric to identify high-performing selection strategies for directed evolution. Using theoretical NK landscapes and four empirical protein fitness landscapes, we demonstrate consistent in silico improvement upon the performance of fixed-parameter strategies, using a pipeline that could also be combined with emerging AI-based methods for driving directed evolution. Author summary: Directed evolution (DE) navigates high-dimensional protein fitness landscapes to find improved variants, but the success of any strategy depends on the landscape’s ruggedness, which is typically unknown and often requires genetic sequencing to derive. We present Sequence-free Landscape Inference for Directed Evolution (SLIDE), a two-step framework that (i) estimates landscape ruggedness from population-level phenotypic decay curves together with a mutation rate estimate, and (ii) uses that ruggedness estimate to choose DE control parameters that better balance exploration and exploitation. In contrast to existing ruggedness metrics, SLIDE relies on spectral analysis to link the exponential decay rate of mean fitness under random, unbiased mutations to the dominant frequency content of the landscape. This decay rate is related to the normalised Dirichlet energy of the underlying genotype graph, and the framework naturally extends to biased mutation spectra. On homogeneous landscapes, such as NK landscapes, ruggedness can be accurately estimated from a single starting genotype. Heterogeneous empirical landscapes instead require sampling multiple starting points, naturally distinguishing local from global ruggedness estimates. We combine SLIDE with DE strategies developed in our previous work, but the method is strategy-agnostic and can be paired with any DE workflow, including emerging AI-driven workflows. In silico tests show that SLIDE reliably recovers landscape ruggedness and enables more effective selection of directed evolution strategies than fixed-parameter approaches across landscapes of varying ruggedness. Although heterogeneous landscapes require fitness decay measurements from multiple starting points, these can be generated in parallel using high-throughput mutagenesis methods, such as error-prone PCR or chemical mutagenesis, thereby avoiding the need for extensive sequencing or combinatorially complete genotype–phenotype maps while remaining scalable to large screens. Beyond DE, the approach can rapidly characterise landscapes arising in genetic circuits, enzyme libraries, and ecological models, providing a way to quantify genetic stability and guide experimental or computational search.
SN - 1553-7358
ER -