F-CPI: Prediction of activity changes induced by fluorine substitution using multimodal deep learning

Qian Zhang, Weizhen Yin, Xinyao Chen, Aimin Zhou, Guixu Zhang, Zhi Zhao, Zhiqiang Li, Yan Zhang, Jingshan Shen, Weiliang Zhu, Xiangrui Jiang, Zhijian XubioRxiv 2024 · DOI

TL;DR

This paper introduces F-CPI, a novel multimodal deep learning framework designed to predict how fluorine substitution affects drug activity. By leveraging sequence and graph-based features, F-CPI achieves high accuracy in predicting these changes, offering a valuable tool for accelerating drug discovery and optimization, as demonstrated by a significant increase in activity for a SARS-CoV-2 3CLpro inhibitor.

What Problem Does This Paper Address?

Fluorine substitution is a widely used strategy in drug discovery to fine-tune a molecule's physicochemical properties, often leading to improved efficacy and metabolic stability. However, predicting the precise impact of fluorine substitution on a drug's activity is challenging. Traditional methods struggle with the nuanced electronic effects of fluorine, and existing AI models for compound-protein interactions (CPI) are not specifically tailored to account for the unique challenges posed by fluorine's influence. This gap means that many promising fluorine-containing drug candidates might be overlooked or require extensive, costly experimental validation. The authors aim to bridge this gap by developing a specialized deep learning model that can accurately predict activity changes upon fluorine substitution.

Methods

The core of this work is the F-CPI framework, which is designed to be a general paradigm for predicting fluoro-substitution compound-protein interactions (CPI-FS). F-CPI utilizes a multimodal approach, incorporating various feature representations for both compounds and proteins. The authors explore three main model architectures:

  • F-CPI (S): Sequence-based: This model uses sequence features (SMILES strings for compounds, amino acid sequences for proteins) and employs Transformer encoders with self-attention mechanisms for feature extraction.
  • F-CPI (G): Graph-based: This architecture treats molecules as graphs, with atoms as nodes and bonds as edges. It uses Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) to process compound and protein structural information.
  • F-CPI (M): Multimodal: This model combines multiple feature types, including sequence features, Morgan fingerprints for compounds, and pre-trained embeddings (like TAPE) and PSSM matrices for proteins.

All F-CPI models incorporate a decoder that fuses compound and protein features to predict binary outcomes (significant activity change or not). The models also leverage techniques like focal loss to handle imbalanced datasets and auxiliary tasks (predicting activity values) to improve learning.

Datasets

The authors curated a large dataset of 111,168 paired samples, each containing an original compound and its fluoro-substituted analog, along with protein information and activity measurements (IC50, Ki, etc.). This dataset was derived from ChEMBL and other sources. For comparison with traditional CPI models, a new dataset was constructed by filtering for IC50 values, containing 1965 proteins, 42406 compounds, and 48387 fluoride-substituted control groups.

Key Insights and Ideas

  • Specialization is Key: The study highlights that a model specifically designed for fluorine substitution (F-CPI) outperforms general CPI models, underscoring the unique challenges and opportunities presented by fluorine chemistry.
  • Multimodality Matters: Combining different types of molecular and protein features (sequence, graph, fingerprints) leads to more robust predictions, especially in the F-CPI (M) model.
  • Attention Mechanisms are Powerful: The use of Transformer and GAT architectures, which rely on attention mechanisms, proves effective in capturing complex interactions and subtle changes induced by fluorine.
  • Activity Cliffs are Predictable: The F-CPI framework demonstrates a strong ability to predict significant activity changes, even when they are substantial (e.g., >100-fold increase), which is crucial for identifying promising drug candidates.
  • Pre-trained Embeddings Offer a Trade-off: While pre-trained embeddings can speed up training and improve performance, their effectiveness depends on the specific task and data. For compounds, they might overemphasize similarity and miss subtle fluorine effects, whereas for proteins, they can effectively capture complex structural information.

Key Results

  • The best-performing model, F-CPI (M), achieved an accuracy of 89.46%, a recall of 43.59%, and a precision of 67.27% on the constructed dataset.
  • F-CPI (M) outperformed traditional machine learning methods (SVM, RF) and other deep learning CPI models (DeepDTA, GraphDTA) in accuracy, precision, and recall.
  • In a case study, F-CPI successfully predicted a significant increase in activity for a SARS-CoV-2 3CLpro inhibitor, where the introduction of a single fluorine atom led to a more than 100-fold increase in activity (IC50: 22.99 nM vs. 28190 nM).
  • Ablation studies showed that introducing Morgan fingerprints and auxiliary tasks significantly improved model performance, while the PSSM-400 feature had a smaller impact.

Limitations and Future Work

  • Dataset Imbalance: While focal loss was used, the dataset still exhibits an imbalance between positive and negative samples, which can affect model performance.
  • Protein Feature Engineering: The authors note that expressing complex protein information through manual prior design is challenging, suggesting that further advancements in protein feature representation could improve model performance.
  • Generalizability: While the model shows promise, further validation on diverse datasets and real-world drug discovery pipelines would be beneficial.
  • Experimental Validation: The paper highlights the importance of wet experiments to verify predictions, suggesting that integrating experimental feedback loops could further refine the model.
  • Exploring Other Fluorine Effects: The current work focuses on single fluorine substitutions. Future work could explore the impact of multiple fluorine atoms or different types of fluorinated groups.

Summary updated

Abstract

Abstract There are a large number of fluorine (F)-containing compounds in approved drugs, and F substitution is a common method in drug discovery and development. However, F is difficult to form traditional hydrogen bonds and typical halogen bonds. As a result, accurate prediction of the activity after F substitution is still impossible using traditional drug design methods, whereas artificial intelligence driven activity prediction might offer a solution. Although more and more machine learning and deep learning models are being applied, there is currently no model specifically designed to study the effect of F on bioactivities. In this study, we developed a specialized deep learning model, F-CPI, to predict the effect of introducing F on drug activity, and tested its performance on a carefully constructed dataset. Comparison with traditional machine learning models and popular CPI task models demonstrated the superiority and necessity of F-CPI, achieving an accuracy of approximately 89% and a precision of approximately 67%. In the end, we utilized F-CPI for the structural optimization of hit compounds against SARS-CoV-2 3CL pro . Impressively, in one case, the introduction of only one F atom resulted in a more than 100-fold increase in activity (IC 50 : 22.99 nM vs. 28190 nM). Therefore, we believe that F-CPI is a helpful and effective tool in the context of drug discovery and design.
BibTeX
@misc{zhang2024f,
  author = {Qian Zhang and Weizhen Yin and Xinyao Chen and Aimin Zhou and Guixu Zhang and Zhi Zhao and Zhiqiang Li and Yan Zhang and Jingshan Shen and Weiliang Zhu and Xiangrui Jiang and Zhijian Xu},
  title = {{F-CPI: Prediction of activity changes induced by fluorine substitution using multimodal deep learning}},
  year = {2024},
  doi = {10.1101/2024.07.17.602844},
  howpublished = {bioRxiv},
}
Figure