Q-BIOLAT: Binary Protein Fitness Optimization

• Q-BIOLAT is a framework that casts protein fitness optimization as a QUBO-style combinatorial search problem using binary latent codes.
• The approach employs pretrained protein language model embeddings, PCA-based thresholding, and surrogate modeling to transform continuous data into optimization-friendly representations.
• Empirical findings show that combinatorial solvers like simulated annealing and genetic algorithms effectively explore rugged fitness landscapes in low-dimensional latent spaces.

Q-BIOLAT is a framework for protein fitness optimization in a discrete binary latent space, designed to turn protein design into a QUBO-style combinatorial optimization problem. In the 2026 arXiv preprints "Binary Latent Protein Fitness Landscapes for Quantum Annealing Optimization" and "Q-BIOLAT: Binary Latent Protein Fitness Landscapes for QUBO-Based Optimization," the framework is presented as a representation-learning and optimization pipeline that starts from pretrained protein LLM embeddings, compresses them into binary latent codes, learns a quadratic unconstrained binary optimization surrogate over those codes, and then applies classical combinatorial solvers such as simulated annealing, genetic algorithms, and greedy hill climbing to search for high-fitness regions of latent space (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

1. Definition, scope, and motivating problem

Q-BIOLAT is motivated by the observation that protein sequences are discrete, high-dimensional, and rugged / epistatic, while most learning-based protein design methods work in continuous latent spaces and are evaluated mainly by prediction metrics like Spearman correlation. The framework argues that, if the end goal is to optimize sequence fitness rather than merely predict it, then the representation should be optimized for searchability, not only predictive accuracy (Hy, 29 Mar 2026).

The core claim is that protein sequences live in a huge discrete space and that good prediction does not necessarily imply good optimization. Q-BIOLAT therefore constructs an optimization-aware binary representation where search can be done directly with discrete methods. Binary spaces are compatible with QUBO / Ising optimization, make combinatorial search explicit on the Boolean hypercube $\{0,1\}^m$, and allow the landscape to be analyzed in terms of unary effects, pairwise interactions, bit-flip dynamics, and local optima (Hy, 29 Mar 2026).

The framework is also presented as a bridge between three areas: protein representation learning via pretrained PLM embeddings, combinatorial optimization via binary search and QUBO objectives, and quantum-inspired or quantum-compatible optimization because QUBO / Ising formulations map naturally to annealers (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

2. Latent representation pipeline

The full pipeline is written as

$$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$

or, in the earlier paper,

$$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$

Here $s$ is a protein sequence, $e(s)$ is a pretrained PLM embedding, $z(s)$ is a reduced continuous latent vector, $x(s)$ is a binary latent code, and $f(x)$ or $\hat f(x)$ is the latent fitness surrogate (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

For a protein sequence of length $L$, the ESM-based model produces residue-level hidden states

$$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$0

A fixed-length embedding is then obtained by mean pooling: $$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$1 The extended paper also gives the mask-aware form

$$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$2

where $$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$3 is the attention mask (Hy, 29 Mar 2026).

The dense embedding is projected into a lower-dimensional latent space of size $$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$4. The papers consider random projection and PCA, and the extended paper also considers a deterministic autoencoder and a VAE. For random projection,

$$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$5

while PCA uses the principal-component projection of $$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$6 into $$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$7 (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

The reduced latent vector is binarized into

$$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$8

The main binarization rule is median thresholding: $$s \;\to\; e(s)\in\mathbb{R}^d \;\to\; z(s)\in\mathbb{R}^m \;\to\; x(s)\in\{0,1\}^m \;\to\; f(x),$$9 or equivalently

$$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$0

with the threshold taken as the median of the corresponding latent coordinate over the training set. This makes bit activations roughly balanced and defines a Boolean hypercube latent space (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

The extended paper reports a marked distinction between representation families after binarization:

Representation	Reported behavior
PCA + thresholding	high-entropy, decodable, optimization-friendly latent spaces
Random projection + thresholding	workable but weaker than PCA
Deterministic autoencoder	collapse after binarization
VAE	collapse after binarization

The reported collapse of AE and VAE means that, after thresholding, almost all bits become constant, bit entropy is near zero, and active latent dimensions disappear, even though reconstruction error remains low. The papers interpret this as evidence that reconstruction quality in continuous space does not translate automatically into usable binary codes for search (Hy, 29 Mar 2026).

3. QUBO surrogate and induced optimization landscape

The latent fitness function is modeled as a quadratic unconstrained binary optimization objective over binary variables $$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$1: $$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$2 or equivalently

$$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$3

and, in the extended paper,

$$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$4

with $$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$5 symmetric and zero diagonal. The coefficients $$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$6 are unary terms and the $$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$7 are pairwise interactions (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

Feature construction uses all linear terms $$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$8 and all pairwise terms $$s \;\rightarrow\; e(s) \in \mathbb{R}^{d} \;\rightarrow\; z(s) \in \mathbb{R}^{m} \;\rightarrow\; x(s) \in \{0,1\}^{m} \;\rightarrow\; \hat{f}(x).$$9. The total number of features is

$s$0

The parameters are estimated by ridge regression: $s$1 where $s$2 is the design matrix of unary and pairwise features, $s$3 is the vector of measured fitness values, and $s$4 is the $s$5 regularization coefficient (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

The optimization problem is then

$s$6

a standard discrete combinatorial optimization problem (Hy, 29 Mar 2026).

The extended paper makes the induced landscape itself a central object of study. For a single-bit flip, if $s$7 is $s$8 with bit $s$9 flipped, the fitness change is

$e(s)$0

with

$e(s)$1

A code $e(s)$2 is a single-bit local maximizer if

$e(s)$3

These expressions are used to characterize local search behavior (Hy, 29 Mar 2026).

The paper also gives smoothness bounds in Hamming space: $e(s)$4 and

$e(s)$5

This supports the paper’s claim that small $e(s)$6 corresponds to a smoother landscape, whereas larger interaction norms correspond to rougher, more rugged behavior (Hy, 29 Mar 2026).

A related point concerns identifiability. Since the QUBO feature map has dimension

$e(s)$7

if the design matrix is rank-deficient, then the QUBO parameters are not uniquely identifiable. The paper uses this as a formal explanation for why predictive accuracy and optimization can decouple: multiple parameterizations can fit observed data equally well but behave differently on unseen binary codes (Hy, 29 Mar 2026).

4. Search procedures, decoding, and evaluation

Q-BIOLAT studies several optimizers in binary latent space: simulated annealing, genetic algorithm, greedy hill climbing, random search, and latent Bayesian optimization. Simulated annealing uses single-bit flips and temperature-dependent acceptance; the genetic algorithm uses selection, crossover, and mutation; greedy hill climbing always takes the best improving single-bit flip; and latent BO is described as a lightweight Bayesian-style latent search baseline or a kernel-based uncertainty heuristic over Hamming space (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

The papers emphasize a trade-off between local and global search. Low-dimensional latent spaces are easier to optimize but may be less expressive; higher-dimensional latent spaces are more expressive but harder to optimize. The best trade-off is often around

$e(s)$8

Local methods such as greedy hill climbing or simulated annealing tend to keep solutions near the observed data manifold, while genetic algorithms are better at broad exploration in larger latent spaces (Hy, 18 Mar 2026).

In the earlier GFP-focused paper, mapping back to sequence space is retrieval-based rather than generative. Given an optimized code $e(s)$9, the nearest observed training sequences are retrieved by Hamming distance

$z(s)$0

The nearest neighbors are treated as real protein sequences associated with the optimized region of latent space. The paper is explicit that Q-BioLat does not generate new sequences directly in this setup (Hy, 18 Mar 2026).

Evaluation uses surrogate-level and retrieval-level metrics. The reported metrics include Spearman correlation, surrogate improvement, NN true fitness, and NN percentile. The earlier paper explains nearest-neighbor percentile as the rank percentile of the retrieved sequence’s true experimental fitness within the training distribution (Hy, 18 Mar 2026). The extended paper adds external sequence-level oracle evaluation using ridge regression, XGBoost, and Gaussian Process Regression, reporting that Gaussian process regression is usually the best oracle in low-data and moderate-data regimes, while ridge becomes competitive as data grows (Hy, 29 Mar 2026).

5. Empirical findings

The earlier paper evaluates on ProteinGym, specifically the GFP benchmark, using subset sizes

$z(s)$1

an $z(s)$2 train/test split, ESM-2 mean-pooled embeddings, latent dimensions

$z(s)$3

and both PCA and random projection (Hy, 18 Mar 2026).

For 10,000 samples, the reported Spearman correlations are 0.209 at dimension 8, 0.302 at dimension 16, 0.385 at dimension 32, and 0.413 at dimension 64. The paper interprets this as evidence that larger latent spaces can improve predictive fidelity when there is enough data, but that smaller datasets can suffer from overfitting in high dimensions (Hy, 18 Mar 2026).

Under PCA-based binary latents with 10,000 samples and 16 bits, the reported optimization results are:

• Simulated Annealing: Improvement $z(s)$4, NN True Fitness $z(s)$5, NN Percentile $z(s)$6.
• Genetic Algorithm: the same reported values as SA in the table.
• Random Search: Improvement $z(s)$7, NN True Fitness $z(s)$8, NN Percentile $z(s)$9.
• Greedy Hill Climb: Improvement $x(s)$0, NN True Fitness $x(s)$1, NN Percentile $x(s)$2.
• Latent BO: Improvement $x(s)$3, NN True Fitness $x(s)$4, NN Percentile $x(s)$5 (Hy, 18 Mar 2026).

These numbers support several of the paper’s main conclusions: SA and GA give the strongest surrogate improvement, greedy hill climbing can yield the best true nearest-neighbor fitness or percentile, random search is surprisingly competitive in relatively low-dimensional latent spaces, and latent BO underperforms in this setup (Hy, 18 Mar 2026).

A further result concerns representation geometry. For 10,000 samples and 16-bit latent codes, random projection and PCA have the same reported Spearman correlation of 0.302, but PCA gives better optimization performance: NN True Fitness $x(s)$6 and NN Percentile $x(s)$7, compared with $x(s)$8 and $x(s)$9 for random projection (Hy, 18 Mar 2026). This is used to support the representation-centric claim that predictive accuracy and optimization effectiveness are not the same thing.

The extended paper broadens the empirical picture to ProteinGym datasets mainly GFP and AAV, and reports that PCA-based binary codes have higher entropy, are more decodable, support stronger optimization, and perform better end-to-end; AE and VAE show near-zero binary entropy and no active dimensions after binarization; and PCA-based representations with 32 or 64 bits often yield the strongest final design scores in moderate-data regimes (Hy, 29 Mar 2026).

6. Significance, limitations, and terminological ambiguity

The conceptual significance of Q-BIOLAT lies in reframing protein fitness optimization as an explicitly discrete search problem in a learned binary space. The framework treats protein fitness as a structured binary energy landscape and argues that the latent landscape matters as much as prediction accuracy. A plausible implication is that evaluation protocols centered only on rank correlation or reconstruction error can miss the central difficulty of design, namely the geometry of the search space induced by the representation (Hy, 29 Mar 2026).

The limitations are also explicit. Binary latent codes are hand-crafted in the thresholding-based versions; the surrogate is quadratic and therefore models only unary and pairwise interactions, not higher-order epistasis directly; retrieval-based decoding is conservative; experiments in the earlier paper are centered on GFP in ProteinGym; and the quantum motivation remains prospective because the papers use classical optimization only, even though the objective is directly compatible with Ising or QUBO hardware (Hy, 18 Mar 2026, Hy, 29 Mar 2026).

The two papers propose several future directions: learned binary representations instead of thresholded projections, better surrogate models beyond pairwise QUBO, integration with quantum annealing hardware, and scaling toward more powerful quantum-assisted protein design pipelines (Hy, 18 Mar 2026).

The term itself also has a naming ambiguity in the arXiv record. In "Utilising a Quantum Hybrid Solver for Bi-objective Quadratic Assignment Problems," Q-BIOLAT refers not to protein fitness landscapes but to the bi-objective quadratic assignment problem studied under a quantum-hybrid constrained-quadratic-model solver. In that paper, Q-BIOLAT is the target use case for applying a quantum-hybrid solver to a bi-objective QAP and comparing scalarisation strategies (Ayodele, 2024). In current protein-design usage, by contrast, Q-BIOLAT denotes the binary latent protein fitness framework introduced in the 2026 preprints (Hy, 29 Mar 2026).