Understanding how genes are expressed in microbial genomes is crucial for unraveling the complexities of their biology and harnessing their potential for applications like biotechnology and medicine. However, studying gene expression in microbes has been challenging, particularly for species that haven’t yet been characterized at the transcriptional level. Researchers at the Beijing Institute of Technology have developed TXpredict, an innovative framework that uses cutting-edge machine learning to bridge this gap.

What is TXpredict?

TXpredict is a transformer-based computational tool designed to predict transcriptomes—the complete set of RNA transcripts produced by a genome. It leverages a powerful protein language model that has been trained on vast datasets, enabling it to analyze annotated genome sequences and predict how genes are expressed in microbes.

Prediction of transcriptome for new genome

a) Schematic of TXpredict for transcriptome prediction. b) Spearman correlation coefficients of predicted and measured gene expression in leave-one-genome-out scenarios. The TXpredict models were separately trained for bacterial and archaeal species. Results from five rounds of model training and testing are reported (n = 5). c) Transcriptome prediction for new genomes. d) Genome sizes and gene counts for the new genomes (n = 900). e) Number of genera in the new genomes (light colors) compared to training genomes (dark colors). f) Distributions of predicted gene expression levels for new bacterial genomes and archaeal genomes (n = 2.34M and 0.77M). g) Top 20 GO terms of the genes in the new genomes calculated by the DeepFRI model²⁵ (left) and their predicted gene expression levels (right). h) Predicted and measured gene expression in the leave-one-genome-out test for B. hinzii. Genes associated with the top 20 GO terms are shown and colored as in g). Spearman correlation coefficient between measured and predicted expression values is indicated in the top left corner. i) t-SNE plots of the protein embeddings calculated by the ESM2 model²⁴ (n = 3.11M). Each dot represents a gene in the new genomes, colored by taxa (left), predicted gene expression (middle), and GO terms (right).

How Accurate is TXpredict?

In tests, TXpredict demonstrated an average Spearman correlation of 0.53 when predicting gene expression in new bacterial genomes. While this may not seem perfect, it’s a significant step forward in accurately modeling microbial transcriptomes, especially considering the diversity and complexity of bacterial species.

Expanding Our Knowledge of Microbial Transcriptomes

One of the most exciting aspects of TXpredict is its scalability. The researchers extended the framework to predict transcriptomes for 900 additional microbial genomes, spanning 276 genera. Remarkably, 69% of these genomes had never been characterized at the transcriptional level before. This achievement opens up new possibilities for exploring microbial diversity and understanding how these organisms function at a molecular level.

A Tool for Context-Specific Insights

TXpredict doesn’t stop at general predictions. It can also forecast condition-specific gene expression, revealing how microbes adapt to different environments. This capability is invaluable for studying microbial behavior under various conditions, such as stress, nutrient availability, or the presence of antibiotics.

Implications for Science and Industry

The ability to predict gene expression in microbes has far-reaching implications:

• Basic Research: TXpredict can help scientists explore microbial adaptation and evolution.
• Biotechnology: By understanding gene regulation, researchers can design synthetic gene sequences for industrial processes, such as biofuel production or pharmaceuticals.
• Medical Applications: Insights from TXpredict could inform the development of antimicrobial therapies by identifying key genes involved in pathogenicity or resistance.

Conclusion

TXpredict represents a major advancement in microbial genomics. By combining powerful machine learning techniques with annotated genome data, it provides researchers with a versatile tool for predicting transcriptomes and understanding microbial gene regulation. As this technology continues to evolve, it holds the promise of unlocking new discoveries and innovations in microbiology and beyond.

Availability – Our trained model and inference codes are available from GitHub: https://github.com/lingxusb/TXpredict