TL;DR: Researchers published a general method called Reinforcement Learning from eXperimental Feedback (RLXF) that steers protein language models toward experimentally measured functional goals, and demonstrated it by creating the most fluorescent CreiLOV variants reported to date.

On 12 September 2026, a peer-reviewed study in Nature Communications introduced functional alignment for protein design. According to Nature Communications, the work is titled “Functional alignment of protein language models via reinforcement learning” and is available as open access with permanent DOI 10.1038/s41467-026-77557-2. According to Scholars@Duke, the publication appeared in Nature Communications on September 12, 2026, from publisher Springer Science and Business Media LLC.

A published answer to a known mismatch

Protein language models, often called pLMs, can generate new protein sequences. According to Nature Communications, those models remain fundamentally misaligned with protein engineering goals. The reason stated in the paper is direct: they lack explicit understanding of function and often fail to improve properties beyond those found in nature.

The new study frames that gap as a training-data consequence. According to Nature Communications, protein language models are powerful tools for generative protein design, but because they are trained solely on natural sequences, they often struggle to generate proteins with enhanced or non-natural functions. The publication therefore describes its contribution as a general framework that aligns generative protein language models with experimentally defined functional objectives.

What RLXF is

The framework is named Reinforcement Learning from eXperimental Feedback, abbreviated RLXF. According to Nature Communications, RLXF is a general framework that aligns protein language models with experimentally measured functional objectives. The same abstract states that the approach draws inspiration from the methods used to align large language models like ChatGPT.

In RLXF, a reward function supplies the functional signal. A research curation note carried alongside the abstract states: “In RLXF, a reward function, such as a supervised sequence-function predictor or any sequence scoring model, provides feedback to the pLM, guiding it to generate sequences with improved function.” That description presents RLXF as a loop: the pLM proposes sequences, a scoring model evaluates predicted function, and reinforcement learning updates the generator toward higher-function outputs.

Tested in five protein families

The claim is not limited to one test protein. According to Nature Communications, the method was applied across five diverse protein families. In that cross-family test, RLXF improved generation of high-functioning variants beyond pre-trained baselines.

The five-family result matters because protein families differ in sequence, structure, and functional constraints. According to Nature Communications, the improvement over pre-trained baselines occurred across those diverse cases, supporting the description of RLXF as a general framework rather than a single-protein procedure. The related subjects listed for the article are protein design, protein engineering, proteins, and structural biology.

Demonstration in CreiLOV

For experimental demonstration, the team selected CreiLOV. According to Nature Communications, CreiLOV is an oxygen-independent fluorescent protein. The paper reports that RLXF-aligned models generated sequences with significantly enhanced fluorescence in that system.

The strongest result concerns record brightness for that reporter. According to Nature Communications, the RLXF-aligned models generated sequences including the most fluorescent CreiLOV variants reported to date. The statement identifies both a significant enhancement and a best-reported outcome within the CreiLOV literature known to the authors at publication.

Why evolutionary knowledge still matters

The study does not present RLXF as a replacement for pre-training. According to Nature Communications, RLXF-aligned models effectively integrate the evolutionary knowledge encoded in pre-trained pLMs with experimental observations. Pre-trained pLMs carry patterns learned from natural sequences, while experimental observations supply measured functional objectives for the target property.

That combination is linked to practical success. According to Nature Communications, the integration improved the success rate of generated sequences. Success rate here refers to the outcome reported by the authors: a higher share of generated sequences showing high function after alignment, relative to the pre-trained starting point.

Mutations that are hard to find otherwise

A central finding involves mutation combinations. According to Nature Communications, RLXF enabled the discovery of synergistic mutation combinations that are difficult to identify through zero-shot or evolutionary approaches. Zero-shot in this context means using a pre-trained model without task-specific experimental training, while evolutionary approaches refer to stepwise search procedures modeled on natural variation and selection.

The paper positions that capability as a bridge between two sources of information. According to Nature Communications, the aligned models combine evolutionary knowledge encoded in pre-trained pLMs with experimental observations. The reported effect is generation of high-functioning variants that neither source alone readily produced.

Beyond the limits of natural evolution

The authors describe the broader purpose in functional terms. According to Nature Communications, RLXF provides a scalable and accessible approach to steer generative models toward desired biochemical properties. The same summary states that the approach enables function-driven protein design beyond the limits of natural evolution.

That wording reflects the starting limitation. Because models trained solely on natural sequences often fail to improve properties beyond those found in nature, according to Nature Communications, steering with experimentally measured functional objectives is presented as the route to enhanced or non-natural functions. The framework is therefore described as function-driven rather than purely sequence-likelihood-driven.

Who reported the work

The author list contains seven names. According to Nature Communications, the authors are Nathaniel Blalock, Srinath Seshadri, Kensuke Nakamura, Agrim Babbar, Sarah A. Fahlberg, Ameya Kulkarni, and Philip A. Romero. According to Scholars@Duke, the citation order is Blalock, Seshadri, Nakamura, Babbar, Fahlberg, Kulkarni, and Romero.

Nathaniel Blalock is listed with two affiliations. According to Nature Communications, those affiliations are the Department of Chemical and Biological Engineering at the University of Wisconsin-Madison in Madison, WI, USA, and the Department of Biomedical Engineering at Duke University in Durham, NC, USA.

Srinath Seshadri, Ameya Kulkarni, and Philip A. Romero are associated with Duke engineering. According to Nature Communications, Seshadri, Kulkarni, and Romero are listed in the Department of Biomedical Engineering at Duke University in Durham, NC, USA, alongside Blalock’s Duke affiliation.

Kensuke Nakamura adds an industry affiliation. According to Nature Communications, Nakamura is listed with Duke Biomedical Engineering and with Modality Research Laboratories II, Daiichi Sankyo Co., Ltd., in Tokyo, Japan.

Agrim Babbar and Sarah A. Fahlberg represent biochemistry. According to Nature Communications, Babbar, Fahlberg, and Romero are listed in the Department of Biochemistry at the University of Wisconsin-Madison in Madison, WI, USA.

Correspondence is assigned to the senior author. According to Nature Communications, correspondence should go to Philip A. Romero at [email protected]. According to Scholars@Duke, Philip A. Romero is the Duke Scholars author associated with Biomedical Engineering for this publication.

How the publication is being shared

The article carries an early-access notice. According to Nature Communications, the site states: “We’re sharing this article early to provide faster access to peer-reviewed, accepted research.” The same page states that the article is citable, carries a permanent DOI, remains subject to further edits, and will be replaced automatically by the final Version of Record.

The journal record identifies the venue and access model. According to Nature Communications, the piece is labeled Article and Open access, with publication date 12 September 2026. According to Scholars@Duke, the journal title is Nature Communications, the ISSN is 2041-1723, and the publisher is Springer Science and Business Media LLC.

Funding, compute, and disclosures

Funding is explicitly acknowledged. According to Nature Communications, the authors acknowledge NSF Award Number 2226383 and NIH NIGMS Grant Number 5R01GM150929-02. A separate funding index associated with the paper also links the work to National Institute of General Medical Sciences support and to a National Science Foundation collaborative project described as integrating deep learning and high-throughput experimentation to navigate protein fitness landscapes for non-native enzyme catalysis.

Compute and discussion support are also named. According to Nature Communications, the authors thank Sameer D’Costa for server management and the High Throughput Computing Center at UW-Madison for compute resources. For scientific input, they thank Tom Sercu for insight into evaluating ESM-2 models and Adithyan Unni, Sam Gelman, and Chase Freschlin for ideas and discussions that aided the work.

The disclosure statement is brief. According to Nature Communications, the authors declare no competing interests. The publisher’s note adds that Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Why this fits protein-engineering practice

Protein engineering often proceeds by mutation and testing. A September 8, 2026 laboratory newsletter from AICell Lab summarized the classic directed-evolution loop as “mutate, test, keep the winners, repeat,” and noted that Frances Arnold shared a Nobel Prize for directed evolution. That newsletter is not part of the RLXF study, but it describes the field context in which RLXF is presented: learning which changes are likely to help, rather than testing blindly.

Machine-learning-guided evolution has been framed in similar terms for years. The same AICell Lab newsletter quotes a review by Yang, Wu and Arnold in Nature Methods in 2019 as stating that “Protein engineering through machine-learning-guided directed evolution enables the optimization of protein functions.” It further quotes that review as saying models “predict how sequence maps to function in a data-driven manner without requiring a detailed model of the underlying physics or biological pathways” and “accelerate directed evolution by learning from the properties of characterized variants and using that information to select sequences that are likely to exhibit improved properties.”

RLXF extends that logic to generative language models. Instead of only ranking existing candidates, the generator itself is updated. According to Nature Communications, the reward comes from experimentally measured functional objectives, implemented through a sequence-function predictor or other sequence scoring model. The generator retains evolutionary knowledge from pre-training while shifting its outputs toward measured function.

What to read carefully

Readers should preserve the study’s status distinctions. RLXF is described as a general framework, its five-family test is described as improvement beyond pre-trained baselines, and its CreiLOV experiment is described as producing significantly enhanced fluorescence including the most fluorescent CreiLOV variants reported to date, all according to Nature Communications. The article does not present those statements as promises about other proteins, other properties, or future products.

The record also distinguishes the shared early version from the final Version of Record. According to Nature Communications, the early shared version is citable and carries a permanent DOI but is subject to further edits and automatic replacement. Any later correction, formatting change, or editorial update would belong to the Version of Record process described on the journal page.

For U.S. science and technology readers, the practical takeaway is contained in the abstract’s final framing. According to Nature Communications, RLXF offers a scalable and accessible way to steer generative models toward desired biochemical properties, using experimental feedback to pursue function-driven protein design beyond the limits of natural evolution.

Sources referenced