A lightweight, integrated generative AI assistant for accelerated early-stage drug discovery on constrained-resource hardware
- 1Amity School of Engineering and Technology, Amity University Punjab
- 2Center for Cyber Security and Cryptology, Sharda School of Computing Science & Engineering, Sharda University
- 3Department of IoT and Intelligent Systems, Manipal University Jaipur
Abstract
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.
Related
- Knowledge distillation-based lightweight generative large language models for standardized quality control of Chinese radiology reports
- A Pragmatic Randomized Trial of an EHR-Integrated Generative AI Chart Summarization Tool for Ambulatory Clinicians
- Gen Z Medical and Physician Assistant Student Recommendations for Medical School Wellness Programming
- AFP assistant: a retrieval-augmented generation and large language model-powered multilingual polio chatbot for low-resource language communities
- Prospective evaluation of a real-time AI assistant for outpatient documentation and quality control
BibTeX
@article{bhatia2026lightweight,
title = {{A lightweight, integrated generative AI assistant for accelerated early-stage drug discovery on constrained-resource hardware}},
author = {Tarandeep Kaur Bhatia and Varun Singh Thakur and Keshav Kaushik and Renu Kumawat},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-61237-8},
url = {https://doi.org/10.1038/s41598-026-61237-8},
note = {CC BY},
}
RIS
TY - JOUR
TI - A lightweight, integrated generative AI assistant for accelerated early-stage drug discovery on constrained-resource hardware
AU - Tarandeep Kaur Bhatia
AU - Varun Singh Thakur
AU - Keshav Kaushik
AU - Renu Kumawat
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-61237-8
UR - https://doi.org/10.1038/s41598-026-61237-8
AB - 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.
SN - 2045-2322
ER -