Comprehensive prediction and analysis of human protein essentiality based on a pre-trained protein large language model

B. S. Kang, Rui Fan, Chunmei Cui, Qinghua CuibioRxiv 2024 · DOI

TL;DR

This paper introduces the Protein Importance Calculator (PIC), a novel deep learning model that leverages pre-trained protein language models to predict protein essentiality across human, cell line, and mouse levels. PIC significantly outperforms existing methods and introduces a new metric, the Protein Essential Score (PES), which can be used to identify prognostic biomarkers and quantify the essentiality of microproteins.

What Problem Does This Paper Address?

Understanding which proteins are "essential" – meaning critical for an organism's survival and development – is crucial for various applications, from identifying drug targets to understanding fundamental biological processes. While many computational methods exist to predict protein essentiality, they often fall short in comprehensively assessing it across different biological contexts. Specifically, existing methods struggle to simultaneously evaluate essentiality at the organism level (humans), cellular level (human cell lines), and in model organisms (mice). This limitation hinders a holistic understanding of protein essentiality and its implications. The authors aim to bridge this gap by developing a unified framework that can predict essentiality across these diverse levels.

Methods

The core of this work is the Protein Importance Calculator (PIC), a deep learning model built by fine-tuning a pre-trained protein language model (PLM), specifically ESM-2. The PIC architecture consists of three main modules:

  • Embedding: This module uses ESM-2 to convert raw protein sequences into fixed-size numerical feature vectors. ESM-2, pre-trained on a massive dataset of protein sequences, captures intrinsic features of proteins. The authors chose a moderate-sized ESM-2 model (esm2_t33_650M_UR50D) to balance performance and computational efficiency.
  • Attention: A multi-head attention mechanism is employed to weigh the importance of different amino acids within a protein sequence, allowing the model to focus on critical regions.
  • Prediction: This module uses an average-pooling layer to obtain a sequence-level representation, which is then fed into a multi-layer perceptron to generate a predicted essential probability.

A key innovation is the Protein Essential Score (PES), derived from the probability output of PIC. PES is designed to quantify protein essentiality and its functional divergence across human, cell line, and mouse levels. The authors also developed specialized PIC models for each level: PIC-human, PIC-cell (with 323 individual cell line models), and PIC-mouse.

Datasets

The study utilized several publicly available databases for data collection:

  • gnomAD: Used for human essentiality data, specifically employing the LOEUF (loss-of-function observed / expected upper bound fraction) metric. Proteins with LOEUF < 0.6 were classified as essential.
  • Project Score: Provided human cell line-specific binary essential scores derived from CRISPR-Cas9 knockout screening experiments across 323 cell lines.
  • Mouse Genome Informatics (MGI) database: Used for mouse essentiality data, with human-mouse homologous proteins assigned essentiality labels based on mouse data.

Key Insights and Ideas

  • Unified Essentiality Prediction: The ability to predict protein essentiality across human, cell line, and mouse levels simultaneously is a significant advancement, offering a more comprehensive view than previous single-context models.
  • Protein Language Models are Powerful: Fine-tuning pre-trained protein language models like ESM-2 proves highly effective for the task of essential protein prediction, outperforming existing methods.
  • PES as a Versatile Metric: The newly defined Protein Essential Score (PES) is not just a predictor of essentiality but also a valuable tool for exploring functional divergence and identifying potential biomarkers.
  • Microproteins Matter: The study demonstrates that PES can be applied to quantify the essentiality of human microproteins, a class of small proteins that have been understudied due to a lack of suitable evaluation methods.
  • Context is King: Protein essentiality is highly context-dependent, varying significantly across different cell types, tissues, and even species, underscoring the need for multi-level analysis.

Key Results

  • Superior Performance: PIC models consistently outperformed existing sequence-based methods. For instance, PIC achieved an AUROC increase of 5.13%-12.10% compared to other methods.
  • Cell Line Specificity: PIC improved the AUROC by an average of 9.64% compared to DeepCellEss, the leading cell line-specific method, across 323 human cell lines.
  • High AUROC Scores: PIC-human achieved the highest AUROC of 0.9132, followed by PIC-mouse (0.8736) and PIC-cell (median AUROC of 0.8579).
  • PES Correlation: PES showed strong positive correlations with various established biological metrics of protein importance, including PPI network degree, normal and cancer tissue expression levels, phyloP, and phastCons scores.
  • Prognostic Biomarkers: PES was successfully used to identify potential prognostic biomarkers for breast cancer, with the top 10 proteins ranked by PES-BRCA showing significant association with patient survival.
  • Microprotein Essentiality: The study quantified the essentiality of 617,462 human microproteins using PES, revealing their involvement in essential biological processes.

Limitations and Future Work

  • Interpretability of PES: The authors acknowledge that while PES is a powerful metric, a deep explanation of its core biological meaning is lacking, largely due to the "black box" nature of neural network models. They believe future advancements in deep learning interpretability will help address this.
  • Broader Biomedical Applications: The authors suggest that further exploration of PES and its application to broader biomedical problems is a promising future direction.
  • Other Cancers and Diseases: The success in breast cancer suggests that PES could be a valuable tool for identifying prognostic biomarkers and therapeutic targets in other cancers and diseases by combining various cell-level PIC models.

Summary updated

Abstract

Abstract Human essential genes and their protein products are indispensable for the viability and development of the individuals. Thus, it is quite important to decipher the essential proteins and up to now numerous computational methods have been developed for the above purpose. However, the current methods failed to comprehensively measure human protein essentiality at levels of humans, human cell lines, and mice orthologues. For doing so, here we developed Protein Importance Calculator (PIC), a sequence-based deep learning model, which was built by fine-tuning a pre-trained protein language model. As a result, PIC outperformed existing methods by increasing 5.13%-12.10% AUROC for predicting essential proteins at human cell-line level. In addition, it improved an average of 9.64% AUROC on 323 human cell lines compared to the only existing cell line-specific method, DeepCellEss. Moreover, we defined Protein Essential Score (PES) to quantify protein essentiality based on PIC and confirmed its power of measuring human protein essentiality and functional divergence across the above three levels. Finally, we successfully used PES to identify prognostic biomarkers of breast cancer and at the first time to quantify the essentiality of 617462 human microproteins. Key Points PIC outperformed existing computational methods for predicting essential proteins. PIC could comprehensively predict human protein essentiality at levels of human, human cell lines and mice orthologues at the same time. PES could serve as a potential metric to quantify the essentiality of both human proteins and human microproteins.
BibTeX
@misc{kang2024comprehensive,
  author = {B. S. Kang and Rui Fan and Chunmei Cui and Qinghua Cui and Qinghua Cui},
  title = {{Comprehensive prediction and analysis of human protein essentiality based on a pre-trained protein large language model}},
  year = {2024},
  doi = {10.1101/2024.03.26.586900},
  howpublished = {bioRxiv},
}
Figure