GenMolVS by Persistent: GPU-Accelerated Drug Discovery on Amazon EKS

GenMolVS (Generative Molecules & Virtual Screening) is an end-to-end, AI-powered drug discovery accelerator jointly developed by Persistent Systems and NVIDIA, orchestrating multiple NVIDIA BioNeMo NIM microservices into a seamless, production-grade pipeline on Amazon Elastic Kubernetes Service (Amazon EKS), powered by GPU-accelerated Amazon EC2 instances with NVIDIA H100 and A10G Tensor Core GPUs.

Drug discovery typically takes 10–15 years and costs upward of $2.6 billion per new drug. The preclinical phase alone—target identification, candidate generation, and screening—accounts for 3–6 years and a disproportionate share of the cost. GenMolVS fundamentally reshapes this by leveraging generative AI to predict protein conformations, generate novel drug-like molecules, and simulate protein–ligand interactions—before a single wet-lab experiment.

• 🔬 Protein Structure Prediction: Predicts a target protein’s 3D conformation with atomic-level accuracy from an amino acid sequence using AlphaFold2, OpenFold, and ESMFold NIMs.
• 🧫 De Novo Molecule Generation: Generates novel, drug-like small molecules optimized for solubility, toxicity, and bioavailability using MolMIM and GenMol NIMs.
• 🎯 Virtual Screening & Molecular Docking: Predicts ligand binding poses, ranks candidates by binding affinity (ΔG), and filters the most promising drug candidates using DiffDock and Boltz-2 NIMs.
• 📊 Visualization & Analysis: Interactive 3D (Molstar) and 2D (SmilesDrawer) viewers plus a sortable results dashboard for binding poses, ADMET profiles, and composite scores.

• ☁️ Amazon EKS orchestrates NIM pods across GPU node groups in an Amazon VPC.
• ⚙️ GPU Node Groups on p5.48xlarge (8x H100) and g5.48xlarge (8x A10G) for high-throughput inference.
• 🧠 FastAPI Orchestrator coordinates the multi-stage pipeline and intermediate data transfer via Amazon S3.
• 🧩 NVIDIA GPU Operator manages drivers, runtime, device plugin, and DCGM exporter.
• 📈 Karpenter Autoscaler provisions right-sized GPU nodes on demand and scales to zero when idle.
• 🗄️ Amazon S3 for protein databases (BFD, UniRef), molecule libraries, and pipeline outputs.
• 📦 Amazon ECR private registry for NIM container images pulled from NVIDIA NGC.
• 🗃️ Amazon RDS PostgreSQL for experiment metadata, run history, and FDA 21 CFR Part 11 audit trails.
• 🔐 AWS Secrets Manager + IRSA for secure handling of NGC API keys and credentials.
• 📡 Amazon CloudWatch + Prometheus/Grafana for GPU and cluster-level observability.

GenMolVS incorporates an agentic AI layer powered by NVIDIA AgentIQ that autonomously selects optimal structure prediction methods, dynamically tunes molecule generation parameters, iteratively refines generation and docking cycles, and produces natural-language result summaries—using an LLM NIM as the reasoning backbone.

• ⚡ Up to 3x faster pipeline execution vs. traditional sequential computation.
• 📉 50–60% reduction in preclinical timeline.
• 💰 Up to 50% cost savings through optimized GPU utilization and Karpenter autoscaling.
• 🧪 End-to-end pipeline in 4–8 weeks vs. 3–6 months preclinical traditionally.