FolDE in Evolutionary Design

• FolDE is a polysemous term defining methods where folding is targeted either through active-learning in protein engineering or via fold-or-die regimes in experimental evolution.
• In protein engineering, the framework uses naturalness-based warm-starting and constant-liar batch selection, achieving up to 55% higher odds of identifying top-performing mutants.
• In experimental evolution, FolDE implements a PCR-based fold or die selection of short oligonucleotides, offering clear genotype–phenotype–fitness mapping for adaptable systems.

Searching arXiv for papers using the term “FolDE” to ground the article in relevant sources. FolDE is a polysemous research term used in at least two distinct technical senses on arXiv. In protein engineering, it denotes “Foldy’s Directed Evolution”, an active-learning-assisted directed evolution workflow for low-$N$ optimization campaigns in which only a small number of mutants can be assayed per round (Roberts et al., 28 Oct 2025). In experimental evolution, it refers to the “fold or die” selection regime applied to Oli populations, where short oligonucleotides are selected for their ability to fold into a hairpin that self-primes under a defined temperature (Collins et al., 2012). A broader, looser usage also appears in active-matter research, where FolDE-style folding design describes the inverse design of active-force sequences that drive colloidal chains into target folded structures (Das et al., 2022). Across these usages, the common motif is that folding is treated not merely as a structural phenomenon but as a selectable, designable, or optimizable property.

1. FolDE as “Foldy’s Directed Evolution”

FolDE in the protein-engineering sense is an active-learning-assisted directed evolution framework designed for the low-$N$ regime, where a laboratory can measure only a small number of mutants per round (Roberts et al., 28 Oct 2025). Its stated objective is to maximize end-of-campaign success, rather than simply maximize immediate per-round exploitation. The motivating problem is that many existing ALDE workflows select the highest-predicted mutants in each round, yielding homogeneous training data that can degrade prediction quality in later rounds.

The method is built around two identified failure modes. The first is round-1 vs. round-2 tension: selecting the most natural mutants in round 1 produces stronger immediate hits than random selection, but these samples are concentrated in a narrow high-naturalness region and therefore provide poor training data for the next round. The second is homogeneous later-round batches: top-ranked mutants in later rounds are often near-duplicates of existing winners, so the batch contributes little exploration or new information (Roberts et al., 28 Oct 2025).

This formulation places FolDE within the broader ALDE literature, but with a specific emphasis on preserving model quality over short, measurement-limited campaigns. A plausible implication is that the framework is intended less as a single-step ranking scheme than as a sequential data-acquisition policy for rugged, epistatic protein fitness landscapes.

2. Methodological components of the low-$N$ protein-engineering framework

The first major component is naturalness-based warm-starting from protein LLM outputs. Naturalness is defined as a log-likelihood ratio between mutant and wild type at the mutated positions:

$$\text{Naturalness}(x', x) = \sum_{i \in T} \log P(x'_i \mid x, \theta) - \log P(x_i \mid x, \theta)$$

where $x$ is the wild-type sequence, $x'$ is the mutant, $T$ is the set of mutated positions, and $P(\cdot \mid x,\theta)$ is the PLM’s distribution over amino acids given the sequence context (Roberts et al., 28 Oct 2025). This quantity is computed with a single forward pass through the PLM. It is used both for zero-shot round-1 selection and for warm-start pretraining of the neural top layer on naturalness scores for all possible single mutants.

The paper identifies warm-starting as the critical mechanism preventing supervised model collapse in round 2. Without it, when round-1 data consist only of the highest-naturalness mutants, the supervised model is trained on a narrow and biased distribution. In the reported benchmarks, the paper states that round-2 Spearman correlation collapses from about 0.48 to 0.04 without warm-start (Roberts et al., 28 Oct 2025). The supervised model uses an ensemble of neural networks with a ranking loss; the top layer is an MLP over PLM embeddings, and training is based on pairwise Bradley–Terry ranking loss $\mathcal{L}_{BT}$ (Roberts et al., 28 Oct 2025).

The second major component is a constant-liar batch selector. After selecting a high-performing mutant, the algorithm inserts a pessimistic imagined observation and updates the covariance structure so that similar candidates are downweighted. The paper gives the covariance update as

$$\mathbf{\Sigma}' = \mathbf{\Sigma}_{M \setminus i} - \mathbf{v}\mathbf{v}^{T}/\sigma_i^2$$

and the mean update as

$N$0

The lie is pessimistic:

$N$1

and the method adds a scaled observation noise term $N$2 to stabilize selection (Roberts et al., 28 Oct 2025). The paper states that lower $N$3 yields a more aggressive lie and more exploration, while higher $N$4 approaches pure exploitation.

3. Evaluation and reported performance of “Foldy’s Directed Evolution”

FolDE was evaluated in silico on ProteinGym datasets using 9 proteins for workflow development/training, 17 single-mutation test proteins, and 3 multi-mutation test proteins, for 20 test protein targets in total (Roberts et al., 28 Oct 2025). The simulation protocol used 3 rounds, 16 mutants per round, and 48 total mutants per campaign. The primary campaign-level metrics were the cumulative number of top 10% mutants discovered and the probability of finding at least one top 1% mutant within 3 rounds.

Across all 20 test proteins, the paper reports that FolDE discovered 23% more top 10% mutants than the random-forest baseline and had a 55% higher likelihood of finding a top 1% mutant. The top-10% comparison is reported as significant with $N$5 by one-sided Wilcoxon signed-rank test, and the top-1% comparison with $N$6 (Roberts et al., 28 Oct 2025). Compared with zero-shot naturalness selection, FolDE found 6% more top 10% mutants and had a 15% higher probability of finding a top 1% mutant, with reported significance $N$7 and $N$8 respectively (Roberts et al., 28 Oct 2025).

The paper further states that zero-shot selection fails to find any top 1% mutant in 5 of 20 targets, whereas FolDE achieves nonzero probability for all targets. On the 17 single-mutation targets, FolDE found 13% more top 10% mutants than random forest and improved the probability of finding at least one top 1% mutant by 50%. On the 3 multi-mutation targets, it found a median of 96% more top 10% mutants than random forest and increased the median probability of finding a top 1% mutant by 2.25× (Roberts et al., 28 Oct 2025).

The ablation results identify round-1 zero-shot prediction, naturalness warm-start, and ranking loss as the largest contributors to performance. The constant-liar component had a smaller effect on the headline benchmark metrics, though the paper argues that it may matter more in larger multi-mutation campaigns where batch redundancy is more severe (Roberts et al., 28 Oct 2025).

4. FolDE as “fold or die” experimental evolution

A different use of FolDE appears in “Fold or die: experimental evolution in vitro”, where the term denotes a fold-or-die experimental evolution system or selection regime acting on Oli populations (Collins et al., 2012). Here the “organism” is Oli, short for short oligonucleotide. Oli is a single-stranded DNA hairpin designed so that, at the ancestral state, it can fold at about 45°C into a hairpin structure that brings its $N$9 and $N$0 ends into proximity. Selection acts on whether the molecule can fold into a hairpin that self-primes under the imposed temperature.

The system is a PCR-based analog of microbial experimental evolution and is described as tractable at the genetic, genomic, phenotypic and fitness levels (Collins et al., 2012). The genotype–phenotype–fitness mapping is explicit:

• Genotype: DNA sequence
• Phenotype: predicted/folded hairpin structure and ability to self-prime
• Fitness: success in subsequent PCR amplification

The selection cycle consists of self-priming under a defined temperature, qPCR-based amplification, restriction digestion, and transfer to the next round (Collins et al., 2012). In a single cycle, dilute Oli populations are incubated with dNTPs and polymerase at a chosen temperature, primers are added, qPCR amplification occurs, the extended $N$1 ends are removed by BsaI and HpyAV, and a diluted subsample is transferred to the next cycle. The restriction-digestion step prevents escape by extension products that would otherwise bypass the folding requirement.

This usage of FolDE differs sharply from the protein-engineering workflow of 2025. In the 2012 system, FolDE is not a computational acquisition policy but a laboratory selection regime in which folding itself is the selectable trait (Collins et al., 2012).

5. Fitness, dynamics, and findings in the Oli/FolDE system

Fitness in the Oli system is measured from the qPCR growth curve. The paper defines population-level fitness as the rate at which the population increases, operationalized as the slope of the qPCR curve during the log-linear phase, referred to in the Results as the relative slope (Collins et al., 2012). The narrative gives the usual PCR growth form

$N$2

and the fitness proxy

$N$3

The ancestral Oli molecule is a 107 bp single-stranded DNA sequence derived from a published HIV-1 hairpin. The experiment started from roughly 30,000 molecules per population and used 56 independent replicate populations under four regimes: Control, Sudden change, Rapid change, and Slow change. All temperature treatments total the same environmental shift of 15°C, from 55°C to 70°C, but differ in the rate of change (Collins et al., 2012).

The proof-of-principle experiment showed that adaptation occurs under rising temperature. All three temperature-increase treatments evolved higher fitness than the control, and slower environmental change led to higher final fitness (Collins et al., 2012). The study also reports visible phenotypic evolution in secondary structure, fluctuating self-priming ability across populations, and lineage diversification and extinction inferred from lineage-network analysis. Another notable finding is the evolution of genome length, including insertions, deletions, and apparent concatamerization or recombination effects, producing what the paper calls the “tyranny of short motifs” (Collins et al., 2012).

The authors frame the system as a model for building empirical genotype–phenotype–fitness maps and emphasize its tractability for studying mutation, epistasis, genome-length effects, and lineage success. This suggests that the fold-or-die meaning of FolDE is best understood as a deliberately minimal experimental platform in which folding is the object of selection rather than the target of prediction.

6. Broader and adjacent uses of the term

A broader, more analogical use appears in “Designing active colloidal folders”, which describes a FolDE-style folding design problem in active matter (Das et al., 2022). There the task is to choose a binary sequence of propulsion directions in a semiflexible colloidal chain so that the chain reliably collapses into a target folded structure. The model uses monomer self-propulsion perpendicular to the chain backbone, with sequence variables $N$4, and dynamics

$N$5

(Das et al., 2022). The paper reports a direct correspondence between the sequence of propulsion directions and the folded structure, and shows that target folds can be designed using a bias potential

$N$6

during sequence optimization (Das et al., 2022). This usage is conceptually aligned with the 2025 protein-engineering FolDE only in the broad sense that both treat folding as a design objective under data or control constraints.

Because these papers do not share a single common formalism, treating FolDE as a unitary term would be misleading. The evidence instead supports a context-dependent nomenclature: in 2012, FolDE denotes a fold-or-die experimental evolution system; in 2025, it denotes a low-$N$7 ALDE framework for protein activity optimization; and in 2022 it appears only as a looser descriptor for a folding-design problem in active colloids [(Collins et al., 2012); (Roberts et al., 28 Oct 2025); (Das et al., 2022)].

7. Conceptual commonalities and distinctions

Despite their different domains, the main FolDE usages share a common structural theme: folding is coupled to a measurable outcome. In the Oli system, the coupling is between hairpin self-priming and qPCR amplification (Collins et al., 2012). In Foldy’s Directed Evolution, the coupling is between sequence priors, activity prediction, and batch selection under a limited assay budget (Roberts et al., 28 Oct 2025). In active colloidal folders, the coupling is between active-force sequence and emergent folded conformation (Das et al., 2022).

The distinctions are at least as important as the similarities. The 2012 FolDE/Oli system is an in vitro experimental evolution platform centered on DNA folding phenotypes. The 2025 FolDE framework is a computational–experimental protein optimization workflow centered on active learning and PLM priors. The 2022 active-colloid usage is neither an evolutionary protocol nor a protein-engineering benchmark, but a nonequilibrium sequence-to-shape design problem [(Collins et al., 2012); (Roberts et al., 28 Oct 2025); (Das et al., 2022)].

This suggests that “FolDE” is best understood not as a single established technical object across disciplines, but as a recurring label attached to research programs that make folding operational: as a criterion for survival, as a target for optimization, or as a programmable design outcome.