AI-enabled alkaline-resistant evolution of protein to apply in mass production

Liqi Kang, Banghao Wu, Bingxin Zhou, Pan Tan, Yun Chan Kang, Yongzhen Yan, Yi Zong, Shuang Li, Zhuo Liu, Liang Hong

Abstract

Artificial intelligence (AI) models have been used to study the compositional regularities of proteins in nature, enabling it to assist in protein design to improve the efficiency of protein engineering and reduce manufacturing cost. However, in industrial settings, proteins are often required to work in extreme environments where they are relatively scarce or even non-existent in nature. Since such proteins are almost absent in the training datasets, it is uncertain whether AI model possesses the capability of evolving the protein to adapt extreme conditions. Antibodies are crucial components of affinity chromatography, and they are hoped to remain active at the extreme environments where most proteins cannot tolerate. In this study, we applied an advanced large language model (LLM), the Pro-PRIME model, to improve the alkali resistance of a representative antibody, a VHH antibody capable of binding to growth hormone. Through two rounds of design, we ensured that the selected mutant has enhanced functionality, including higher thermal stability, extreme pH resistance and stronger affinity, thereby validating the generalized capability of the LLM in meeting specific demands. To the best of our knowledge, this is the first LLM-designed protein product, which is successfully applied in mass production.

Introduction

In laboratory settings, proteins are often delicate and require precise conditions to maintain their structure and function. However, in industrial bioprocessing, they must act as "warriors" capable of withstanding harsh environments. A common challenge is "Clean-in-Place" (CIP) protocols, which utilize 0.5 M Sodium Hydroxide (NaOH)—a solution with a pH exceeding 13—to remove contaminants. Most natural proteins, including the antibodies used to purify therapeutic agents like growth hormones, denature rapidly under these conditions.

The primary hurdle for AI in this domain is that most models are trained on natural protein sequences. Because nature rarely evolves proteins to survive in concentrated lye, these extreme traits are largely absent from training data. This study investigates whether an AI trained on "normal" biological data can predict mutations that enable a protein to survive an "unnatural" industrial gauntlet.

Methodology

The researchers employed a two-stage design strategy using Pro-PRIME, a transformer-based Large Language Model (LLM) architecture-optimized for protein engineering.

  1. Round 1 (Zero-Shot): The model performed saturated mutagenesis, evaluating every possible single-point mutation of a VHH antibody. Without additional experimental data, the model scored these based on their likelihood in natural contexts. The top 45 mutants were selected for experimental validation.
  2. Round 2 (Fine-Tuning): Experimental results from the initial 45 mutants were used to fine-tune Pro-PRIME, allowing it to map the specific landscape of alkali resistance. The model then predicted 20 multi-point mutants.

Validation involved affinity chromatography, thermal stability assays ($T_m$), and mass spectrometry to pinpoint structural vulnerabilities under alkaline stress.

Key Insights: Epistasis and Radical Design

A significant finding was the discovery of positive epistasis. The model identified combinations like A57D and P29T which, while individually inferior to the wild type, created a superior mutant when combined. Traditional "greedy" engineering approaches, which focus only on the best single-point mutations, would have overlooked these synergistic effects.

Furthermore, Pro-PRIME suggested "radical" mutations, such as Alanine to Proline (A15P). While such changes often break standard hydrogen bonding rules and are avoided in rational design, they significantly boosted stability in this context. Mass spectrometry confirmed that these AI-selected mutations often reinforced specific "breakpoints" where the wild-type antibody typically failed.

Results

The AI-driven optimization yielded substantial improvements across all key metrics. The most effective multi-point mutants retained over 50% of their binding capacity after 24 hours in 0.5 M NaOH, a stark contrast to the wild-type antibody, which lost nearly 85% of its capacity.

Normalized EC50 and thermal stability (Tm) of VHH antibody mutants compared to the wild type (WT) after alkaline treatment.

Thermal stability also saw a marked increase, with mutants achieving a melting temperature ($T_m$) boost of approximately 10°C. In a 5,000-liter production environment, the final designed protein maintained high yields for over 140 reuse cycles, more than doubling the 60-cycle lifespan of the wild type. Economically, this optimization is estimated to reduce manufacturing costs by over $1 million annually.

Conclusion and Future Directions

This study represents the first successful deployment of an LLM-designed protein in mass production. While the results demonstrate the power of AI in protein engineering, the underlying physical mechanisms—the "why" behind the success of specific mutations—remain partially obscured. Future work will utilize molecular dynamics simulations to bridge this gap. Additionally, while effective for the relatively short VHH antibody, further research is needed to determine if this approach scales to larger, more complex proteins where the computational complexity of multi-point mutation prediction increases exponentially.

Summary updated

Abstract

Artificial intelligence (AI) models have been used to study the compositional regularities of proteins in nature, enabling it to assist in protein design to improve the efficiency of protein engineering and reduce manufacturing cost. However, in industrial settings, proteins are often required to work in extreme environments where they are relatively scarce or even non-existent in nature. Since such proteins are almost absent in the training datasets, it is uncertain whether AI model possesses the capability of evolving the protein to adapt extreme conditions. Antibodies are crucial components of affinity chromatography, and they are hoped to remain active at the extreme environments where most proteins cannot tolerate. In this study, we applied an advanced large language model (LLM), the Pro-PRIME model, to improve the alkali resistance of a representative antibody, a VHH antibody capable of binding to growth hormone. Through two rounds of design, we ensured that the selected mutant has enhanced functionality, including higher thermal stability, extreme pH resistance, and stronger affinity, thereby validating the generalized capability of the LLM in meeting specific demands. To the best of our knowledge, this is the first LLM-designed protein product, which is successfully applied in mass production.
BibTeX
@inproceedings{kang2024ai,
  author = {Liqi Kang and Banghao Wu and Bingxin Zhou and Pan Tan and Yun Chan Kang and Yongzhen Yan and Yi Zong and Shuang Li and Zhuo Liu and Liang Hong},
  title = {{AI-enabled alkaline-resistant evolution of protein to apply in mass production}},
  booktitle = {eLife},
  year = {2024},
  doi = {10.7554/elife.102788},
}
Figure