Protein Sequence Fitness Estimation

Updated 30 June 2026

Protein sequence fitness estimation is the quantitative determination of a protein’s ability to fold, remain stable, and function by assessing mutation impacts.
Methodologies include statistical-physics models, protein language models, and retrieval-augmented approaches that yield predictive metrics such as Spearman’s ρ ranging from 0.40 to 0.57.
Hybrid and surrogate models integrate sequence and structure data to enhance predictions, thereby advancing protein engineering and providing insights into evolutionary mechanisms.

Protein sequence fitness estimation is the quantitative inference of how mutations to an amino acid sequence impact a protein’s biological performance. Fitness, in this context, encapsulates a protein's ability to fold, remain stable, and perform its molecular function under evolutionary or synthetic engineering constraints. The estimation of fitness from sequence is a central task in protein engineering, evolutionary biology, and high-throughput mutagenesis, with methodological innovations drawing from statistical physics, unsupervised and supervised machine learning, population genetics, and combinatorial optimization.

1. Formal Definitions and Theoretical Frameworks

Fitness estimation links a protein sequence $x=(x_1, ..., x_L)$ to a scalar "fitness" $F(x)$ , representing a property such as catalytic activity, binding, or growth rate. Traditional formulations include:

Statistical-physics (maximum entropy) approaches: Model the sequence ensemble with a Boltzmann distribution, such that $P(x) \propto \exp(-E(x))$ , where $E(x)$ is an effective energy calculated from observed sequence statistics or physical energies (Miyazawa, 2016, Dai et al., 20 May 2025).
Population genetics approaches: The equilibrium distribution of sequences under mutation-fixation dynamics also follows a Boltzmann form, $P(x) \propto \exp(4N_e m(x))$ , where $m(x)$ is Malthusian fitness and $N_e$ the effective population size (Miyazawa, 2016).
Probabilistic LLM (PLM) approaches: Model fitness as the log-probability (or likelihood ratio relative to wild-type) of a sequence under a generative model trained on protein sequence databases (Fan et al., 8 Oct 2025, Notin et al., 2022, Kantroo et al., 2024). For masked LLMs (MLMs), the per-position pseudo-likelihood or pseudo-perplexity are commonly used as fitness proxies.

These frameworks share the postulate that sequence (and sometimes structure) statistics encode selective pressures, so that high-probability (or low pseudo-perplexity) regions in the learned model’s distribution correlate with biologically plausible, high-fitness sequences.

2. Fitness Estimation from Protein LLMs

Modern protein fitness estimation is dominated by pretrained PLMs:

Masked LLMs (MLMs) such as ESM and derivatives, trained to reconstruct masked amino acids, yield per-position likelihoods $p(x_i|x_{\setminus i})$ . The additive pseudo-log-likelihood (PLL) or its exponential (pseudo-perplexity, pPL), given by $F_{PLL}(x) = \frac{1}{L} \sum_{i=1}^L \log p(x_i|x_{\setminus i})$ , serves as a universal fitness score (Kantroo et al., 2024).
Autoregressive/causal PLMs (e.g. ProGen, Proust, Tranception) factorize $P(x) = \prod_{i=1}^L p_\theta(x_i|x_{<i})$ . Zero-shot fitness for a mutant sequence $F(x)$ 0 relative to wild-type $F(x)$ 1 is the log-likelihood difference $F(x)$ 2, where $F(x)$ 3 (Eris, 2 Feb 2026, Notin et al., 2022).
Retrieval-augmented PLMs incorporate homologous sequences at inference (e.g., via PSSMs, MSAs, or vectorized retrieval) to inject position-specific evolutionary context (Notin et al., 2022, Weitzman et al., 10 Jun 2025). The fusion is often a linear or log-probability sum.

Empirically, MLM pseudo-perplexity and autoregressive log-probability difference approaches yield Spearman’s $F(x)$ 4 in the range $F(x)$ 5– $F(x)$ 6 on the ProteinGym substitution benchmark, with indel tasks benefiting particularly from OFS (One-Fell-Swoop) pPL-based estimators, achieving $F(x)$ 7 (Kantroo et al., 2024, Eris, 2 Feb 2026).

Sequence-structure modalities are integrated through hybrid architectures:

Structure-aware fitness models: Combine sequence embeddings with structural graphs (e.g., SE(3)-transformers or GVPs), optionally leveraging molecular surface topology (Zhang et al., 2024, Sharma et al., 23 Apr 2025). Surface-aware models (e.g., S3F) achieve up to $F(x)$ 8 on ProteinGym, with structure+MSA or structure+sequence ensembles exceeding simple uni-modal predictors.
Profile fusion: EvoIF exemplifies combining within-family sequence/structure profiles and cross-family inverse-folding likelihoods in a compact transformer block, delivering state-of-the-art zero-shot predictive accuracy versus DMS-derived ground truth (Fan et al., 8 Oct 2025).
Practical issues: The inclusion of explicit structure aids mostly on stability and binding tasks. However, in disordered regions or regions masked for low AlphaFold confidence, sequence-only or MSA-only predictors often outperform hybrids (Sharma et al., 23 Apr 2025).

Table: Comparative Spearman’s $F(x)$ 9 (ProteinGym Substitutions) | Model | Type | Avg $P(x) \propto \exp(-E(x))$ 0 | |------------------|-------------------|--------------| | ESM-2-650M | MLM | 0.414 | | Proust | Causal LM | 0.390 | | S3F | Seq+Struct+Surf | 0.470 | | EvoIF (MSA-free) | Hybrid (profile) | 0.489 | | Protriever | Retrieval-aug PLM | 0.479 |

4. Supervised and Surrogate-Based Fitness Estimation

When high-throughput experimental fitness labels are available:

CNN/MLP-based surrogates: One-hot encoded sequences are input to shallow convolutional or fully-connected neural networks, trained by mean squared error relative to measured fitness (Bogensperger et al., 31 Jan 2025, Zong, 2023). Ensembles provide both mean and predictive variance for use in Bayesian optimization (Zong, 2023).
Graph signal smoothing: Fitness values are smoothed over a kNN graph (Hamming space) by Tikhonov regularization, yielding a closed-form update $P(x) \propto \exp(-E(x))$ 1 (Kirjner et al., 2023). Training a CNN on these smoothed labels yields more robust surrogates. Gibbs sampling with graph-based smoothing (GGS) exceeds other model-based optimization routines in simulated protein design.
Variational frameworks: Models such as VLGPO (Variational Latent Generative Protein Optimization) embed sequences in continuous latent spaces and employ a learned surrogate (CNN), whose gradients are incorporated into flow-matching sampling for protein design (Bogensperger et al., 31 Jan 2025).
Graph-convolutional networks for subspace exploration: In low-dimensional mutational spaces, protein fitness landscapes can be modeled as signals on a hypercube $P(x) \propto \exp(-E(x))$ 2 and denoised by wavelet transforms, with GCNs learning epistatic patterns (Daud et al., 20 Jun 2025).

Empirically, properly regularized surrogate models and graph-signal approaches can improve both rank correlation with ground truth and the convergence rate of optimization in sequence design applications (Kirjner et al., 2023).

5. Evolutionary, Statistical, and Biophysical Energy Models

Co-evolutionary models (Potts, DCA) remain foundational:

Potts/statistical energy: Potts models infer site and pairwise couplings $P(x) \propto \exp(-E(x))$ 3 from MSAs by maximum entropy. Change in statistical energy $P(x) \propto \exp(-E(x))$ 4 for a mutant is computed exactly, with group Lasso penalties and node-wise multinomial regression yielding both interpretability and improved convergence guarantees (Dai et al., 20 May 2025).
Integration with structure: Penalty weights can incorporate 3D residue distances, relaxing regularization for spatially proximate sites to reflect biophysical contact constraints (Dai et al., 20 May 2025).
Theoretical grounding: Protein-sequence ensembles under selection, inverse-Potts models, and physical folding energies relate via Boltzmann distributions with $P(x) \propto \exp(-E(x))$ 5, where $P(x) \propto \exp(-E(x))$ 6 is the folding free energy gap and $P(x) \propto \exp(-E(x))$ 7 the selective temperature (Miyazawa, 2016).
Inference of $P(x) \propto \exp(-E(x))$ 8 and $P(x) \propto \exp(-E(x))$ 9: Evolutionary statistical energies $E(x)$ 0 can be related to experimentally accessible mutational effects and evolutionary divergence rates (i.e., the nonsynonymous/synonymous substitution ratio), enabling direct estimation of population-genetic parameters from sequence data (Miyazawa, 2016).

6. Specialized and Hybrid Approaches

A range of problem-specific and emerging methods target aspects of the fitness estimation challenge:

Few-shot predictors: PRIMO employs in-context learning and test-time training to leverage a small number of experimentally labeled variants, integrating zero-shot PLM predictions and preference-based loss functions to outperform both naive transfer and classical regression baselines across diverse protein engineering tasks (Teufel et al., 2 Dec 2025).
Latent binary optimization: Q-BIOLAT projects continuous PLM embeddings to lower-dimensional binary codes, fits a quadratic surrogate (QUBO), and uses combinatorial optimization/quantum annealing to extract high-fitness candidates, offering new hardware-level integration (Hy, 18 Mar 2026).
End-to-end retrieval: Protriever replaces traditional MSA-based retrieval with learned vector search, optimizing both homolog retrieval and sequence modeling jointly, yielding zero-shot fitness estimates surpassing other sequence-based approaches on the ProteinGym benchmark and providing massive computational speedup (Weitzman et al., 10 Jun 2025).

These advances demonstrate generalization both to data-sparse and combinatorially large sequence spaces, multi-mutant and indel regimes, and scenarios limited by the depth or reliability of evolutionary information.

7. Benchmarking, Challenges, and Trends

Empirical benchmarking: The ProteinGym suite (217+ DMS assays, substitutions/indels) has become the de facto standard. Sequence-structure or hybrid models now reach $E(x)$ 1– $E(x)$ 2 in zero-shot settings; indel fitness is best predicted by OFS pPL and advanced causal LMs (Kantroo et al., 2024, Eris, 2 Feb 2026).
Data limitations: Disordered regions (low-confidence AF2 pLDDT), sequence–function epistasis, and high mutational depth pose ongoing challenges for all method classes (Sharma et al., 23 Apr 2025, Zhang et al., 2024).
Interpretability: High-performing models increasingly provide mechanism-aware diagnostics, e.g., entropy variance across positions predicting the value of homology retrieval for a given protein (Eris, 2 Feb 2026), or epistatic heatmaps resolving spatial couplings (Zhang et al., 2024).
Future directions: Scalability to full-proteome protein design, integration of biophysical surface and interaction features, online/adaptive learning (test-time training, active learning), and hardware-accelerated optimization (quantum, distributed) are active areas, leveraging both zero-shot generalization and sample-efficient supervised adaptation.

In summary, protein sequence fitness estimation unifies statistical, biological, and physical perspectives, leveraging unsupervised LLMs, graph-based methods, co-evolutionary statistics, and combinatorial search. Continuous benchmarking and methodological innovation have yielded robust, scalable surrogates and interpretable diagnostic tools, with accuracy now frequently limited by experimental constraints on sequence–function mapping rather than algorithmic sophistication (Eris, 2 Feb 2026 Kantroo et al., 2024 Sharma et al., 23 Apr 2025 Fan et al., 8 Oct 2025 Kirjner et al., 2023).