Scholarlink

<p>Discovery of novel therapeutic small molecules remains one of the most challenging tasks in pharmaceutical research, as it is described by high attrition rates, well over $2 billion per candidate, and timelines often surpassing a decade. Generative AI has transformative potential to probe the enormous chemical space estimated at 10<sup>60</sup> drug-like molecules, but current computational approaches are fragmented across multiple tools and require enterprise-grade hardware, thus creating a sort of “computational divide” that excludes many academic groups and smaller laboratories. We introduce in this study a unified end-to-end Generative AI Assistant, specifically designed for constrained hardware environments, optimized to run on widely available consumer-grade GPUs like the NVIDIA GTX 1650 with 4&#xa0;GB VRAM. With our system, we integrated a lightweight LSTM-based generative model using SELFIES tokenization to ensure 100% syntactic validity, a multi-task XGBoost classifier for toxicity prediction across 12 biological assays, hybrid property prediction with molecular fingerprints, and an API-based module for 3D protein structure prediction via ESMFold. Using benchmark testing, we have established a robust performance level with a reliable convergence of our generative model in conjunction with an overall decrease in training loss from 2.15 to 1.19 and a stable validation loss of 1.43, as well as a weighted average AUC of 0.790 for our toxicity classifier. Toxic-class recall significantly increased after threshold tuning, improving the framework’s appropriateness for early-stage safety screening applications. In addition, the analysis of chemical space validates the model’s ability to generate previously unexplored novel molecules that possess desirable drug-like characteristics (mean LogP = 2.04) and to enhance safety, we are integrating both ML-based toxicity screening and a rule-based PAINS filtering into a novel hybrid prototype. As an additional contribution to overall accessibility, we also developed and tested a CPU fallback feature to allow for automatic fallback of the generative model in instances of hardware incompatibility. Together these contributions enable the efficient and cost-effective democratization of modern drug discovery workflows on affordable hardware (93% reduction in infrastructure costs) without sacrificing the scientific integrity or predictive capability of our developed models.</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

A lightweight, integrated generative AI assistant for accelerated early-stage drug discovery on constrained-resource hardware

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月18日

查看全文
  • Tarandeep Kaur Bhatia,
  • Varun Singh Thakur,
  • Keshav Kaushik,
  • Renu Kumawat

摘要

Discovery of novel therapeutic small molecules remains one of the most challenging tasks in pharmaceutical research, as it is described by high attrition rates, well over $2 billion per candidate, and timelines often surpassing a decade. Generative AI has transformative potential to probe the enormous chemical space estimated at 1060 drug-like molecules, but current computational approaches are fragmented across multiple tools and require enterprise-grade hardware, thus creating a sort of “computational divide” that excludes many academic groups and smaller laboratories. We introduce in this study a unified end-to-end Generative AI Assistant, specifically designed for constrained hardware environments, optimized to run on widely available consumer-grade GPUs like the NVIDIA GTX 1650 with 4 GB VRAM. With our system, we integrated a lightweight LSTM-based generative model using SELFIES tokenization to ensure 100% syntactic validity, a multi-task XGBoost classifier for toxicity prediction across 12 biological assays, hybrid property prediction with molecular fingerprints, and an API-based module for 3D protein structure prediction via ESMFold. Using benchmark testing, we have established a robust performance level with a reliable convergence of our generative model in conjunction with an overall decrease in training loss from 2.15 to 1.19 and a stable validation loss of 1.43, as well as a weighted average AUC of 0.790 for our toxicity classifier. Toxic-class recall significantly increased after threshold tuning, improving the framework’s appropriateness for early-stage safety screening applications. In addition, the analysis of chemical space validates the model’s ability to generate previously unexplored novel molecules that possess desirable drug-like characteristics (mean LogP = 2.04) and to enhance safety, we are integrating both ML-based toxicity screening and a rule-based PAINS filtering into a novel hybrid prototype. As an additional contribution to overall accessibility, we also developed and tested a CPU fallback feature to allow for automatic fallback of the generative model in instances of hardware incompatibility. Together these contributions enable the efficient and cost-effective democratization of modern drug discovery workflows on affordable hardware (93% reduction in infrastructure costs) without sacrificing the scientific integrity or predictive capability of our developed models.