Proteins are often described as the machines of biology, but designing one from scratch remains an extraordinary challenge. A sequence of amino acids must fold into the right three-dimensional structure, bind the correct partner, avoid unwanted interactions and remain stable enough to function inside a real biological environment. Generative artificial-intelligence models have begun to produce protein sequences with many of the statistical features seen in nature, yet a sequence that looks biologically plausible is not necessarily useful. The new study “Aligning protein-generative models to experimental fitness with ProteinDPO,” published in Nature Methods by T. Widatalla, A. A. Borah, S. H. King and colleagues, presents a strategy for narrowing that gap by teaching protein-generative models to respond directly to experimental evidence. Called ProteinDPO, the approach is designed to align computational protein design with measured fitness, the laboratory-observed ability of a variant to perform a selected biological function.
Modern protein-generative models learn from enormous sequence databases. In much the same way that language models learn patterns of grammar, meaning and style from text, protein models learn patterns associated with amino-acid sequences that have survived evolution or have been observed experimentally. Once trained, these systems can generate new sequences, complete missing regions or propose variants of a known protein. Their central limitation is that the training data do not automatically tell the model which newly generated sequences will work best in a particular assay. A model may understand that certain residues commonly occur together, while lacking a precise sense of how a single mutation will alter catalytic activity, binding, expression or stability. ProteinDPO addresses this problem by using experimental comparisons between variants as a form of guidance. Rather than asking only whether a sequence resembles natural proteins, it asks whether one candidate performs better than another under a defined laboratory test.
The method is built around the idea of preference optimization, a family of techniques developed in artificial intelligence to make generative systems better reflect human or externally measured preferences. In a conventional preference-learning setup, a model receives pairs of outputs and information about which output is preferred. For proteins, the preferred sequence can be the variant that produces a stronger experimental signal, survives better under selection, binds a target more effectively or achieves another assay-specific measure of fitness. ProteinDPO uses these ordered comparisons to adjust the probability assigned to protein sequences. The goal is not simply to train a separate predictor and then search through the generator’s output. Instead, the generative model itself is modified so that its distribution shifts toward sequences associated with higher measured performance while retaining the biological patterns learned during pretraining.
This distinction matters because protein design is a multi-objective problem. A sequence can score highly in a computational prediction but fail when synthesized, expressed or tested. It may fold incorrectly, aggregate, degrade rapidly or become toxic to the host cell used in the experiment. Laboratory fitness measurements capture some of these effects in an integrated way, although no single assay represents every property required for a successful protein. By aligning generation with experimental outcomes, ProteinDPO aims to place the practical behavior of variants closer to the center of the design process. The framework can therefore be viewed as a feedback loop: a model proposes sequences, experiments measure their performance, the resulting comparisons become training information and the updated model proposes a new, more informed set of candidates.
The technical challenge is to improve fitness without destroying the model’s prior knowledge of protein sequence space. If optimization pushes too aggressively toward a small collection of high-scoring examples, the generator may lose diversity or produce sequences that exploit quirks of the assay rather than genuine biological function. This problem is related to distribution shift and reward hacking in machine learning. A model can learn to maximize the available score in ways that do not translate to the broader objective. ProteinDPO is intended to manage this tension by using preference-based updates rather than treating experimental fitness as an unlimited instruction to maximize. In practical terms, the method seeks a controlled movement away from the original model distribution, preserving plausible sequence patterns while increasing the probability of variants favored by experiments. That balance is especially important when experimental datasets are small, expensive and noisy.
Experimental fitness itself is not a universal quantity. It is a measurement defined by the biological system and protocol used to obtain it. A variant that grows rapidly in one selection experiment may not be more stable, more active in a purified system or more effective in a therapeutic context. Measurements can also contain technical variation arising from library construction, sequencing depth, expression differences and assay conditions. Pairwise preference data may nevertheless be useful because they can be more robust than trying to predict an exact numerical score for every sequence. If one variant consistently outperforms another, that relationship provides a directional signal for model training. ProteinDPO’s emphasis on experimentally ranked sequences reflects a broader movement in machine learning toward using comparisons, which can be easier to obtain and interpret than perfectly calibrated labels.
The reported framework is part of a rapidly expanding effort to connect generative biology with automated experimentation. Earlier generations of protein-design systems often treated generation as the final step: produce candidates, select a few using computational filters and send them to the laboratory. Newer approaches increasingly make experiments part of the model’s learning process. This creates the possibility of iterative design cycles in which each round of testing improves the next round of proposals. The appeal is clear for fields such as enzyme engineering, where millions of possible mutations may exist but only a small number can be tested directly. A model that learns from measured preferences could prioritize promising regions of sequence space and reduce the number of laboratory experiments required to reach a useful variant.
ProteinDPO also illustrates why generative models should not be evaluated only by the appearance of their outputs. A protein sequence can be novel, diverse and statistically convincing yet still fail its intended task. More meaningful assessments ask whether generated candidates express successfully, fold, retain activity and improve over appropriate baselines in prospective experiments. The study’s focus on experimental fitness places those questions at the forefront. It also highlights the importance of reporting how fitness is measured, how comparisons are constructed and whether the resulting designs generalize beyond the data used for alignment. Such details determine whether an apparent improvement represents genuine biological progress or simply better adaptation to one particular assay.
The approach could eventually support applications ranging from industrial biotechnology to medicine, but significant barriers remain. Protein function depends on cellular context, molecular partners and environmental conditions that may not be represented in a training dataset. Safety is another concern: a sequence optimized for a desired property could acquire unintended interactions or biological activity. Generative systems therefore require careful screening, containment and experimental validation, particularly when designs involve pathogens, toxins or clinically relevant targets. There is also a scientific question about interpretability. Preference optimization can reveal which sequences perform better, but it does not automatically explain the molecular mechanism responsible for the improvement. Structural modeling, biochemical characterization and targeted mutagenesis will still be necessary to understand why a design works.
By linking a protein generator to the outcomes of real experiments, ProteinDPO offers a practical answer to one of the central problems in AI-assisted biology: how to move from plausible sequences to useful molecules. Its significance lies less in replacing laboratory science than in making computational proposals more responsive to what laboratory science observes. The model begins with the broad evolutionary knowledge encoded in protein sequences, then receives a more focused signal from variants tested under defined conditions. If this feedback can be applied reliably across proteins and assays, generative systems may become adaptive design partners rather than static sequence predictors. The work points toward a future in which protein engineering is organized as a continuous dialogue between algorithms and experiments, with each informing the other and with biological performance—not merely computational plausibility—serving as the final judge.
Subject of Research: Protein-generative models aligned with experimentally measured protein fitness.
Article Title: Aligning protein-generative models to experimental fitness with ProteinDPO
Article References: Widatalla, T., Borah, A.A., King, S.H. et al. Aligning protein-generative models to experimental fitness with ProteinDPO. Nature Methods (2026). https://doi.org/10.1038/s41592-026-03137-3
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41592-026-03137-3
Keywords: Protein design, generative AI, ProteinDPO, direct preference optimization, experimental fitness, machine learning, protein engineering, synthetic biology
Tags: AI-driven biological function predictionalignment of computational protein models with laboratory databiologically plausible protein sequence generationdeep learning in protein structure and functionenhancing protein design accuracy with Fitness dataexperimental fitness measurements in protein designintegrating experimental data with generative modelsmachine learning for protein engineeringprotein sequence optimization using experimental evidenceprotein-generative modelsProteinDPO method for improving protein sequence relevance
