Abstract Overview

Recently, a team led by Tie-Yan Liu at Microsoft Research published a study titled “Accelerating protein engineering with fitness landscape modelling and reinforcement learning” in Nature Machine Intelligence (a top-tier journal with an impact factor of 23.9). Addressing fundamental challenges in protein engineering—such as the enormous combinatorial sequence space, the difficulty of predicting epistatic interactions, and the low efficiency of experimental exploration—the authors proposed the μProtein framework.

This framework tightly couples a high-accuracy mutation effect prediction model, μFormer, with a reinforcement-learning-based exploration algorithm, μSearch. Remarkably, relying solely on single-point mutation experimental data, μProtein can efficiently discover higher-order functional variants and capture epistatic interactions between mutations.

In an application to optimizing antibiotic resistance of TEM-1 β-lactamase, μProtein designed multi-point mutant enzymes with exceptionally enhanced activity, surpassing the highest levels previously reported, and uncovered resistance mutations that have not been observed in nature. This work provides a novel and highly efficient strategy for large-scale protein sequence optimization.

Title: Accelerating protein engineering with fitness landscape modelling and reinforcement learning
Journal: Nature Machine Intelligence
DOI: https://doi.org/10.1038/s42256-025-01103-w
Code: https://github.com/microsoft/Mu-Protein

Background Expansion

The goal of protein engineering is to modify the amino acid sequence of natural proteins so that they exhibit improved performance under specific conditions, such as high temperature, extreme pH, or the presence of antibiotics. However, this task faces a fundamental challenge: the number of possible mutation combinations is astronomically large.

For a typical protein consisting of around 300 amino acids, even if mutations are introduced at only a few positions—with 20 possible amino acids at each site—the number of possible combinations can easily exceed trillions. Yet, truly high-performing variants are extremely rare within this vast “sequence space,” making the search akin to finding a needle in a haystack.

The challenge is further compounded by the fact that the effects of multiple mutations are often not additive. In some cases, two mutations with modest individual effects can jointly produce a dramatic functional enhancement; in other cases, a seemingly deleterious mutation can become critical in a specific mutational background. This nonlinear interaction between mutations is known as epistasis, which makes evolutionary trajectories toward optimal variants difficult to predict and renders traditional step-by-step experimental screening highly inefficient.

At present, high-throughput experimental techniques such as deep mutational scanning can typically evaluate only single-point mutations in a systematic manner. Exhaustively testing all double or triple mutants is prohibitively expensive. Therefore, a central bottleneck in protein engineering lies in how to accurately predict the functional effects of complex mutation combinations using limited experimental data, and how to efficiently identify the most promising candidates for experimental validation.

Figure 1. Protein fitness landscape.

A New AI-Driven Paradigm for Protein Engineering: Beyond Prediction to Active Exploration

In recent years, artificial intelligence has been widely applied to assist protein engineering. Mainstream predictive models typically encode amino acid sequences into numerical representations and train models to learn the mapping between sequence and function. Such models can predict functional scores for new sequences, but they suffer from notable limitations.

First, prediction accuracy—especially for multi-point mutations—remains insufficient, largely due to the difficulty of capturing epistatic effects. Second, most models are passive predictors: while they can score candidate sequences, they cannot actively propose sequences that are most worthwhile for experimental testing.

Against this backdrop, the study by Tie-Yan Liu’s team at Microsoft Research proposes a new framework named μProtein, aiming to address both challenges simultaneously. The core idea is to tightly integrate a high-precision mutation effect predictor (μFormer) with an intelligent search algorithm (μSearch), forming a dual-engine mechanism of “prediction + exploration.”

Specifically, μFormer is responsible for accurately predicting the function of every candidate mutant—effectively constructing a fitness landscape, where each point corresponds to a protein sequence and the height represents its functional performance. μSearch then acts as an intelligent navigator on this landscape, actively exploring and progressively moving toward regions of maximal fitness, rather than blindly sampling the entire space.

Thus, the success of the framework hinges on two pillars: the predictive accuracy of μFormer and the exploration efficiency of μSearch.

Figure 2. Workflow of the μProtein framework.

1. μFormer: Accurate Prediction of Complex Mutations Using Only Single-Point Data

To achieve sufficiently high predictive accuracy, μFormer employs a newly trained protein language model developed by the authors, leveraging their previously proposed paired masked language modeling (PMLM) pretraining strategy. This strategy simulates cooperative relationships between residue pairs during evolution, enabling the model to learn long-range dependencies and structural constraints among amino acids.

The key breakthrough is that μFormer requires only single-point mutation experimental data for fine-tuning, yet can accurately predict the functional effects of sequences containing multiple mutations, including insertions and deletions. Across multiple public benchmarks, μFormer outperforms state-of-the-art models such as DeepSequence and ProteinNPT. More importantly, it effectively captures epistatic interactions—a longstanding challenge for traditional linear models.

Figure 3. μFormer outperforms mainstream mutation effect prediction models across multiple benchmark datasets.

Figure 4. μFormer shows superior performance over DeepSequence and ProteinNPT on multiple higher-order mutant datasets.

The authors further visualized embeddings of TEM-1 β-lactamase mutants using t-SNE dimensionality reduction. High-activity variants, including extended-spectrum β-lactamases (ESBLs), form distinct clusters in the low-dimensional space, and align strongly with experimentally measured functional gradients along the principal components. This demonstrates that μFormer does not merely count mutations, but genuinely learns deep relationships between sequence and function.

Figure 5. t-SNE visualization of embedding vectors extracted by μFormer for TEM-1 mutants.

2. μSearch: Intelligent Exploration of Optimal Mutants

With a reliable predictive model in place, the next challenge is efficient search within an astronomically large sequence space. To this end, the authors designed μSearch, a reinforcement-learning-based search algorithm.

μSearch operates analogously to hill climbing. Starting from the wild-type protein, it iteratively proposes single amino acid substitutions—each step corresponding to a movement in sequence space. μFormer immediately evaluates the predicted fitness of the new sequence. If fitness improves, the search continues along that direction; if not, μSearch adapts its strategy and explores alternative paths.

Through continual trial-and-error and policy updates, μSearch avoids fitness valleys and rapidly converges toward high-activity regions, ultimately reaching the global optimum.

Comparative experiments on multiple simulated fitness landscapes demonstrate that μSearch significantly outperforms eight existing algorithms in terms of efficiency at discovering high-functioning sequences.

Figure 6. Performance comparison between μSearch and mainstream fitness landscape exploration algorithms.

Experimental Validation: Surpassing Naturally Occurring Resistance Mutations

To validate the practical utility of the μProtein framework, the authors conducted wet-lab experiments on TEM-1 β-lactamase, aiming to enhance resistance to cefotaxime, a third-generation cephalosporin antibiotic. Using only approximately 5,000 single-point mutation measurements to fine-tune μFormer, μSearch generated candidate sequences containing two to three mutations, from which highly functional variants were selected.

Experimental results showed that the double mutant G236S; T261V designed by μProtein exhibited resistance exceeding that of a well-known classical quadruple mutant (A40G; E102K; M180T; G236S), vividly demonstrating the power of epistasis-driven functional gains.

Notably, the framework also identified a previously unreported mutation, T261I, which has not been observed in naturally occurring resistant strains. Structural analysis revealed that this site lies near the active pocket and may reshape the hydrogen-bond network, thereby influencing substrate binding—offering new mechanistic insights into β-lactam antibiotic resistance.

Moreover, compared with random design strategies, μProtein-generated sequences showed a significantly higher experimental success rate, indicating that the framework not only predicts accurately but also effectively reduces experimental burden and improves overall engineering efficiency.

Figure 7. High-function TEM-1 sequences designed by μProtein.

Overall Summary

1. Core Contributions

The central innovation of the μProtein framework lies in its synergistic integration of high-accuracy prediction and high-efficiency exploration. μFormer enables accurate prediction of higher-order mutations—including double, triple, and even insertion/deletion variants—using only single-point mutation data, while effectively capturing epistatic interactions. μSearch, leveraging reinforcement learning, overcomes inefficient search in massive sequence spaces and dramatically improves the discovery rate of high-function mutants.

2. Limitations and Future Directions

Despite its impressive achievements, several open questions remain. From a personal perspective, the following points merit further investigation:

• Limited validation under extremely small sample sizes: While fine-tuning μFormer with ~5,000 single-point mutations is feasible in high-throughput settings, it remains unclear whether comparable performance can be achieved with only hundreds or even dozens of data points.
• Absence of a multi-round active learning loop: In this study, the predictive model is fine-tuned only once, and newly generated experimental data are not iteratively incorporated. Prior approaches such as ALDE and EvolvePro have demonstrated that closed-loop “design–experiment–feedback” strategies can substantially reduce data requirements.
• Untapped potential of multimodal integration: Future extensions could incorporate predicted protein 3D structures (e.g., AlphaFold2 outputs), conformational dynamics, or energy landscapes to enhance physical interpretability—an avenue also highlighted in the paper’s discussion section.