Scholarlink

<p>Transformer-based, template-free SMILES-to-SMILES translation models for reaction prediction and single-step retrosynthesis are of interest to computer-aided synthesis planning systems, as they offer state-of-the-art accuracy. However, their slow inference speed limits their practical utility in such applications. To address this challenge, we propose speculative decoding with a simple chemically specific drafting strategy and apply it to the Molecular Transformer, an encoder-decoder transformer for conditional SMILES generation. Our approach achieves over 3X faster inference in reaction product prediction and single-step retrosynthesis with no loss in accuracy, increasing the potential of the transformer as the backbone of synthesis planning systems. To accelerate the simultaneous generation of multiple precursor SMILES for a given query SMILES in single-step retrosynthesis, we introduce Speculative Beam Search, a novel algorithm tackling the challenge of beam search acceleration with speculative decoding. Our methods aim to improve transformer-based models’ scalability and industrial applicability in synthesis planning.</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

Accelerating the inference of string generation-based chemical reaction models for industrial applications

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年3月10日

查看全文
  • Mikhail Andronov,
  • Natalia Andronova,
  • Michael Wand,
  • Jürgen Schmidhuber,
  • Djork-Arné Clevert

摘要

Transformer-based, template-free SMILES-to-SMILES translation models for reaction prediction and single-step retrosynthesis are of interest to computer-aided synthesis planning systems, as they offer state-of-the-art accuracy. However, their slow inference speed limits their practical utility in such applications. To address this challenge, we propose speculative decoding with a simple chemically specific drafting strategy and apply it to the Molecular Transformer, an encoder-decoder transformer for conditional SMILES generation. Our approach achieves over 3X faster inference in reaction product prediction and single-step retrosynthesis with no loss in accuracy, increasing the potential of the transformer as the backbone of synthesis planning systems. To accelerate the simultaneous generation of multiple precursor SMILES for a given query SMILES in single-step retrosynthesis, we introduce Speculative Beam Search, a novel algorithm tackling the challenge of beam search acceleration with speculative decoding. Our methods aim to improve transformer-based models’ scalability and industrial applicability in synthesis planning.