Scholarlink

<p>Drug discovery is a complex and costly process. Efficiently predicting drug–target interactions is crucial for streamlining this endeavor. Traditional methods, such as high-throughput screening, are impractical due to their high costs and time demands. deep learning (DL) has emerged as promising alternative, demonstrating significant potential in predicting drug–target interactions. However, these models often face challenges related to their high computational complexity. Transformer architectures, which utilize attention mechanisms, are particularly affected by its quadratic complexity which becomes computationally expensive, especially as input sequence lengths increase, posing significant barriers to scalability and practical application in large-scale drug discovery. This paper introduces FFT-DTBA, a novel transformer architecture for drug–target binding affinity (DTBA) prediction. FFT-DTBA leverages Fourier transforms to efficiently overcome the intensive resources need of conventional transformers in large-scale drug discovery. FFT-DTBA avoids using Attention layers, replacing them with Fast Fourier transform and permutations, significantly reducing the model complexity as Fourier transforms are neither computationally intensive, nor resources consuming. Our model adheres to the same design and architecture of transformers, though it integrates Fourier transforms with permutations as effective technique for token mixing. Furthermore, to enhance the model capabilities, the encoded representation undergoes a weighting aggregation step which assigns weights to different positions within the representation, effectively focusing on the most relevant information for predicting binding affinities, and consequently, providing the same benefits as attention mechanisms without the drawbacks. This improvement allowed FFT-DTBA to achieve similar or better performance compared to transformer models based on attention mechanisms, while reducing complexity significantly. Rigorous testing on the Davis and KIBA datasets demonstrates FFT-DTBA superior performance compared to existing models. By addressing the computational complexities of existing models, FFT-DTBA offers a promising solution for large-scale drug discovery efforts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fourier transform and permutation-based transformer for drug–target binding affinity prediction

  • 期刊论文

  • 发表日期：

    2025年7月19日

查看全文
  • Nassima Aleb

摘要

Drug discovery is a complex and costly process. Efficiently predicting drug–target interactions is crucial for streamlining this endeavor. Traditional methods, such as high-throughput screening, are impractical due to their high costs and time demands. deep learning (DL) has emerged as promising alternative, demonstrating significant potential in predicting drug–target interactions. However, these models often face challenges related to their high computational complexity. Transformer architectures, which utilize attention mechanisms, are particularly affected by its quadratic complexity which becomes computationally expensive, especially as input sequence lengths increase, posing significant barriers to scalability and practical application in large-scale drug discovery. This paper introduces FFT-DTBA, a novel transformer architecture for drug–target binding affinity (DTBA) prediction. FFT-DTBA leverages Fourier transforms to efficiently overcome the intensive resources need of conventional transformers in large-scale drug discovery. FFT-DTBA avoids using Attention layers, replacing them with Fast Fourier transform and permutations, significantly reducing the model complexity as Fourier transforms are neither computationally intensive, nor resources consuming. Our model adheres to the same design and architecture of transformers, though it integrates Fourier transforms with permutations as effective technique for token mixing. Furthermore, to enhance the model capabilities, the encoded representation undergoes a weighting aggregation step which assigns weights to different positions within the representation, effectively focusing on the most relevant information for predicting binding affinities, and consequently, providing the same benefits as attention mechanisms without the drawbacks. This improvement allowed FFT-DTBA to achieve similar or better performance compared to transformer models based on attention mechanisms, while reducing complexity significantly. Rigorous testing on the Davis and KIBA datasets demonstrates FFT-DTBA superior performance compared to existing models. By addressing the computational complexities of existing models, FFT-DTBA offers a promising solution for large-scale drug discovery efforts.