Deep learning driven biosynthetic pathways navigation for natural products with BioNavi-NP

Shuangjia Zheng, Tao Zeng, Chengtao Li, Binghong Chen, Connor W. Coley, Yuedong Yang, Ruibo WuNature Communications 2022 · DOI

TL;DR

This paper introduces BioNavi-NP, a deep learning-powered tool that predicts biosynthetic pathways for natural products (NPs). By combining general organic and specific biosynthetic reactions, BioNavi-NP can identify plausible pathways for a vast majority of tested compounds, significantly outperforming existing rule-based methods and offering a valuable resource for NP discovery and engineering.

What Problem Does This Paper Address?

Natural products (NPs) are a rich source of novel drugs and bioactive compounds, but their complex structures and low natural abundance make their production and study challenging. A major hurdle is that the complete biosynthetic pathways for most of the over 300,000 known NPs are unknown. While some NPs can be synthesized chemically, this is often difficult and expensive. Therefore, predicting how NPs are made in nature (biosynthesis) from simpler building blocks is crucial for both understanding their origins and for enabling their production through synthetic biology or semi-synthesis. Existing computational tools for this task, known as bio-retrosynthesis, are often limited by their reliance on manually curated rules or incomplete reaction databases, making them less effective for complex or novel pathways.

Methods

BioNavi-NP employs a two-pronged approach:

  1. Single-Step Bio-Retrosynthesis Prediction: The core of the system is a deep learning model, specifically an end-to-end transformer neural network. This model is trained to predict the immediate precursor(s) of a given molecule. Crucially, it's trained on a combined dataset of both general organic reactions and known biosynthetic reactions. This allows it to learn broader chemical transformations while also capturing the specific nuances of biological synthesis. Data augmentation with a large set of organic reactions (USPTO_NPL) significantly improved performance, suggesting that learning general chemical principles aids in predicting biological transformations. An ensemble of four transformer models was used to further enhance robustness and accuracy.
  2. Multi-Step Pathway Planning: Once the single-step prediction model is established, BioNavi-NP uses a deep learning-guided AND-OR tree-based search algorithm (inspired by Retro*) to navigate and plan complete multi-step biosynthetic pathways. This search algorithm efficiently explores the vast combinatorial space of possible pathways, starting from the target NP and working backward to known building blocks.

Datasets

  • BioChem Dataset: A curated dataset of 33,710 unique pairs of precursors and metabolites, derived from MetaCyc, KEGG, and MetaNetX. This dataset was used for training and evaluating the single-step prediction model.
  • USPTO_NPL Dataset: A dataset of 60,000 organic reactions, selected from the USPTO database for their similarity to biochemical reactions. This was used for data augmentation to improve the transformer model's robustness.
  • Internal Test Set: 368 diverse NPs with known complete biosynthetic pathways, used for evaluating the multi-step planning performance of BioNavi-NP. These were categorized into five classes: AA/MA, AAs, CA/SA, MVA/MEP, and Others.
  • External Test Set: 25 unseen NPs from recent publications, used to assess the generalizability of BioNavi-NP.
  • Benchmark Datasets (LASER and Golden): Datasets previously used to evaluate rule-based tools like RetroPathRL, used for direct comparison.

Key Insights and Ideas

  • The Power of Combined Data: Training on both general organic and specific biosynthetic reactions is key. This suggests that while biological pathways have unique features, they also share fundamental chemical principles with organic reactions, and learning both leads to superior prediction.
  • Organic Retrosynthesis Tools Aren't Enough: The paper explicitly shows that models trained only on organic reactions fail to predict biosynthetic precursors, highlighting the distinct nature of biological synthesis despite shared chemistry.
  • Ensemble Learning for Robustness: Combining multiple transformer models (ensemble) significantly boosted performance, demonstrating the benefit of averaging predictions from diverse models to reduce variance and improve reliability.
  • AND-OR Tree Search for Efficiency: The Retro*-inspired search algorithm is crucial for navigating the complex, branching nature of biosynthetic pathways, making the multi-step planning computationally feasible.
  • Building Block Hit Rate Matters: The paper emphasizes that correctly identifying the building blocks is as important as finding the pathway itself, especially since the exact pathway might not be unique.

Key Results

  • Single-Step Prediction: BioNavi-NP's best single-step model achieved a top-10 accuracy of 60.6%, which is 1.7 times more accurate than the previous rule-based model.
  • Multi-Step Planning (Internal Test):
    • Success Rate: BioNavi-NP identified plausible biosynthetic pathways for 90.2% of 368 test compounds. This is a significant improvement over RetroPathRL (52.7%).
  • Building Block Hit Rate: BioNavi-NP recovered reported building blocks for 72.8% of compounds when using user-defined building blocks (UDBs), compared to 5.1% for RetroPathRL_UDB.
  • Pathway Hit Rate: BioNavi-NP achieved a pathway hit rate of 26.1% (with UDBs), significantly outperforming RetroPathRL_UDB (4.1%).
  • Efficiency: BioNavi-NP is considerably faster than RetroPathRL, with a much lower computational time for comparable results.
  • Generalizability: The tool performed well on unseen external test cases, with an 88% success rate for pathway prediction.
  • Category Performance: BioNavi-NP showed the highest performance for the AA-MA and CA/SA categories, while the AAs category (especially RiPPs and NRPs) presented a challenge due to complex building blocks.

Limitations and Future Work

  • Complex Structures: The model struggles with predicting biosynthetic starting materials for some very complex NPs that involve many building blocks or reaction steps. The authors suggest that dividing such complex structures into segments for prediction might be a future approach.
  • Intermediate Accuracy: While overall pathways are predicted, individual intermediate predictions can sometimes be incorrect or missing. The authors note that these missing parts can often be inferred from other predicted pathways or by considering the entire network holistically.
  • Anomalous Shortcuts: Some predicted pathways may take "shortcuts" through small cofactor-like molecules, which the authors attribute to a small fraction of anomalous data. Future work will focus on integrating atom-mapping approaches to improve mass balance and reduce such unjustified shortcuts.
  • Enzyme Prediction Integration: While BioNavi-NP predicts pathways, identifying the specific enzymes for these reactions is a separate step. The authors suggest further integrating advanced enzyme prediction and design tools to enhance the practical utility of their predictions.
  • Pathway Scoring: The current pathway scoring function is acknowledged as not perfect due to data limitations and search algorithm constraints. Future work could involve incorporating more detailed information about enzyme promiscuity, fidelity, and diversity.

Summary updated

Abstract

The complete biosynthetic pathways are unknown for most natural products (NPs), it is thus valuable to make computer-aided bio-retrosynthesis predictions. Here, a navigable and user-friendly toolkit, BioNavi-NP, is developed to predict the biosynthetic pathways for both NPs and NP-like compounds. First, a single-step bio-retrosynthesis prediction model is trained using both general organic and biosynthetic reactions through end-to-end transformer neural networks. Based on this model, plausible biosynthetic pathways can be efficiently sampled through an AND-OR tree-based planning algorithm from iterative multi-step bio-retrosynthetic routes. Extensive evaluations reveal that BioNavi-NP can identify biosynthetic pathways for 90.2% of 368 test compounds and recover the reported building blocks as in the test set for 72.8%, 1.7 times more accurate than existing conventional rule-based approaches. The model is further shown to identify biologically plausible pathways for complex NPs collected from the recent literature. The toolkit as well as the curated datasets and learned models are freely available to facilitate the elucidation and reconstruction of the biosynthetic pathways for NPs.
BibTeX
@inproceedings{zheng2022deep,
  author = {Shuangjia Zheng and Tao Zeng and Chengtao Li and Binghong Chen and Connor W. Coley and Yuedong Yang and Ruibo Wu},
  title = {{Deep learning driven biosynthetic pathways navigation for natural products with BioNavi-NP}},
  booktitle = {Nature Communications},
  year = {2022},
  doi = {10.1038/s41467-022-30970-9},
}
Figure