🧬 biology

From single-sequence structure prediction to protein fitness landscape through a composable, epistasis-aware mutation atlas

Cerebra-Epistasis is an end-to-end framework that integrates single-sequence structure prediction with explicit epistasis modeling to accurately predict multi-mutant protein fitness landscapes and enable rapid, one-shot inference of high-order mutation combinations.

Original authors: Yunxin Xu, Weizhe Wang, Zimu Yu, Endi Yang, Ziyu Shi, Shize Yu, Jian Hu, Haipeng Gong

Published 2026-09-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yunxin Xu, Weizhe Wang, Zimu Yu, Endi Yang, Ziyu Shi, Shize Yu, Jian Hu, Haipeng Gong

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Proteins are the molecular machines that keep life running, folding into intricate three-dimensional shapes to perform tasks ranging from digesting food to fighting infection. Scientists have long sought to redesign these proteins, tweaking their genetic blueprints to create better enzymes for industry or more effective medicines. The challenge lies in the sheer scale of the possibilities: a single protein can be altered in millions of different ways, and testing every single variation in a laboratory is impossible. To navigate this vast space, researchers rely on computer models to predict which changes will improve a protein's function and which will break it. However, a major hurdle has been understanding how multiple changes interact with one another. In biology, the effect of two mutations together is often not just the sum of their individual effects; they can amplify or cancel each other out in complex ways, a phenomenon known as epistasis. Predicting these non-linear interactions has been a persistent difficulty, especially when trying to forecast the behavior of proteins with many mutations at once.

A new framework called Cerebra-Epistasis, developed by researchers at Tsinghua University and Zhongnan University of Economics and Law, offers a fresh approach to this problem. Instead of treating each potential protein variant as a unique puzzle to be solved from scratch, the team built a system that first maps the protein's structure and then creates a reusable "atlas" of how every possible single change affects that structure. By combining a fast, single-sequence structure predictor with a network designed to learn how mutations interact, the system can assemble predictions for complex, multi-mutant proteins by simply pulling data from this pre-built map. This method allows the model to see the three-dimensional context of the protein, understanding that residues far apart in the genetic sequence might be close neighbors in the folded shape, influencing each other's behavior.

The researchers tested their system against existing methods using a wide variety of protein datasets. In standard tests involving millions of variants, Cerebra-Epistasis outperformed other leading models in predicting how well a protein would function. Its advantage became even more pronounced as the complexity of the mutations increased. While other models struggled to predict the behavior of proteins with three or more changes, Cerebra-Epistasis maintained high accuracy, suggesting it successfully captured the hidden rules of how mutations combine. The system was also able to predict unseen interactions between pairs of mutations that it had never seen during training, a capability that is crucial for designing new proteins from scratch. In specific case studies, such as analyzing the enzyme tryptophan synthase, the model correctly identified the strongest beneficial interactions between amino acids and mapped them to the correct physical locations within the protein's structure.

Beyond accuracy, the system introduced a significant leap in speed. Traditional methods require the computer to analyze the structure and sequence of every single candidate protein individually, a process that becomes exponentially slower as the number of possible variants grows. Cerebra-Epistasis, by contrast, encodes the original, unmutated protein just once to generate its atlas. Once this atlas is built, the system can instantly evaluate millions of potential combinations by simply retrieving and assembling the relevant pieces of data. The researchers estimated that screening a complete landscape of all possible three-mutation variants for a typical protein would take their system only a few minutes, whereas other top-tier models would require days or even years to complete the same task. This efficiency opens the door to exploring vast regions of protein space that were previously too computationally expensive to investigate.

The study also explored whether training the structure predictor and the fitness predictor together, rather than separately, would yield better results. They found that allowing the two parts of the system to learn from each other improved the accuracy of fitness predictions without degrading the quality of the structural maps. This suggests that the structural context is not just a static backdrop but an active component that can be fine-tuned to better understand how mutations affect function. While the system relies on the structure of the original protein and does not yet account for dramatic shape-shifting caused by mutations, it represents a powerful step forward in making protein engineering more predictable and efficient. By turning a massive, chaotic search problem into a structured assembly task, this work provides a new tool for scientists to design the next generation of biological tools.

Technical Summary: Cerebra-Epistasis

Problem Statement

Mapping the multi-mutant fitness landscape is essential for protein engineering but remains computationally prohibitive due to the vast combinatorial sequence space. A central challenge is the accurate modeling of non-additive epistatic effects (interactions between mutations), which arise from physical residue interactions defined by protein structure. Existing predictors often struggle with higher-order mutants because they either neglect epistasis, oversimplify it as sparse pairwise terms, or require explicit evaluation of every candidate variant. Furthermore, current methods typically lack explicit structural context or require computationally expensive Multiple Sequence Alignment (MSA) searches, hindering end-to-end optimization and efficient landscape-scale inference.

Methodology

The authors present Cerebra-Epistasis, an end-to-end framework that couples a single-sequence structure predictor with a downstream fitness network to construct a reusable, mutation-indexed atlas.

1. Structure-Aware Encoding (Cerebra-Seq)

The framework utilizes Cerebra-Seq, a lightweight, single-sequence structure predictor derived from the MSA-based Cerebra architecture.

Input: It processes a single wild-type (WT) sequence without MSA search. It integrates representations from frozen ESMC and ESM3 models via a manifold-constrained hyper-connection (mHC) mixing head.
Architecture: The model employs Evoformer blocks and a Path Synthesis Attention (PSA) structure module to generate residue-level node representations, residue-pair edge representations, and side-chain-resolved atomic coordinates.
Efficiency: Cerebra-Seq is designed for end-to-end trainability with downstream tasks, offering high structural accuracy with significantly lower parameter counts and training costs compared to models like ESMFold or OmegaFold.

2. Epistasis-Aware Fitness Modeling

The downstream fitness predictor decomposes multi-mutant fitness into additive single-mutation components and non-linear epistatic contributions.

SE(3)-Equivariant Encoding: The WT protein is encoded once using an SE(3)-equivariant transformer that integrates the predicted 3D coordinates, node/edge representations, and ESM-2 embeddings.
Tensor Construction: This encoding generates two reusable tensors indexed by sequence position and amino acid state:
- Single-Mutation Score Tensor ( $S$ ): Provides scalar scores for individual substitutions.
- Epistasis Representation Tensor ( $U$ ): Provides high-dimensional representations for modeling interactions.
Assembly Module: For any arbitrary-order mutant, the model retrieves the corresponding entries from $S$ and $U$ . The fitness is predicted by summing the single-mutation scores and aggregating the epistasis representations through a permutation-invariant summation followed by a shared MLP. This allows the model to infer fitness for unseen higher-order combinations without re-encoding the sequence or structure.

3. Training Strategy

The framework supports two training modes:

Downstream-only: The structure predictor (Cerebra-Seq) is frozen, and only the fitness network is trained.
End-to-End: Both Cerebra-Seq and the fitness network are optimized jointly, allowing structural representations to adapt specifically to mutation-effect prediction.

Key Results

Structure Prediction Performance

Evaluated on CAMEO 2024 and 2025 benchmarks, Cerebra-Seq achieved mean TM-scores of 0.83 and 0.81, respectively. It consistently outperformed direct ESM3 generation and SPIRED, ranking second only to ESMFold. Notably, it surpassed ESMFold on substantial subsets of targets while maintaining a significantly lower inference time (e.g., 1.84x faster than ESMFold for 200–1,000 residue proteins).

Fitness Prediction Accuracy

Standard Benchmarks: On the 1M+ to 1M+ benchmark (random cross-validation), Cerebra-Epistasis achieved a mean Spearman correlation of 0.85, outperforming state-of-the-art baselines like ProteinNPT (0.84).
Low-Data Extrapolation: In the Low- $N$ 1M/2M to 3M+ benchmark (training on <300 single/double mutants to predict higher-order variants), Cerebra-Epistasis achieved a mean Spearman correlation of 0.65, outperforming Kermut (0.62) and MULTI-evolve (0.61). Its advantage increased with mutation order, demonstrating superior extrapolation capabilities.
Epistasis Prediction: In a held-out double-mutant benchmark, the model achieved a Spearman correlation of 0.55 for unseen pairwise epistasis, significantly outperforming Kermut (0.35) and MULTI-evolve (0.19).

Biological Validity

The model successfully recovered biologically meaningful patterns:

TrpB Landscape: It correctly identified the strongest positive epistatic interaction (V227K/S228G) among 2,166 double mutants and reproduced the epistatic landscape across substitutions, consistent with the spatial proximity of these residues near the active site.
GB1 Landscape: The model recapitulated the distance-dependence of epistasis (stronger interactions at shorter spatial distances) and accurately modeled background-dependent epistasis (e.g., the G41–V54 interaction varying with site 39).

Computational Efficiency

The "one-shot" inference paradigm offers massive acceleration. By encoding the WT protein once, the framework avoids repeated sequence/structure processing. The authors estimate that screening a complete 3-mutant landscape for a 100-residue protein takes 3.70 minutes with Cerebra-Epistasis, compared to 2.57 days for Kermut and 2.60 years for ProteinNPT. This represents a speedup of approximately three orders of magnitude.

Significance and Claims

The paper claims that Cerebra-Epistasis provides a unified framework that bridges the gap between accurate structure-aware prediction and efficient landscape-scale inference. Its primary contributions are:

Explicit Epistasis Modeling: By separating additive and epistatic terms via reusable tensors, the model effectively captures non-linear interactions and enables reliable extrapolation from low-order to high-order mutants.
Structural Context: Integrating a lightweight, single-sequence structure predictor allows the model to leverage 3D physical context (packing, contacts) without the computational burden of MSA-based methods.
Scalability: The composable atlas design decouples the cost of structural encoding from the size of the candidate library, enabling the evaluation of combinatorial fitness landscapes that are currently intractable for conventional supervised predictors.
End-to-End Trainability: The framework demonstrates that joint optimization of structure and fitness prediction can improve fitness accuracy while largely preserving structural quality, challenging the convention of treating structure prediction as a fixed preprocessing step.

The authors conclude that while the WT-centered formulation limits the modeling of mutation-induced conformational rearrangements, Cerebra-Epistasis offers a significant step forward in making multi-mutant fitness landscape prediction both accurate and computationally feasible.