Persistent Systems, NVIDIA Launch AI Drug Discovery Tool GenMolVS

Persistent Systems and NVIDIA have launched GenMolVS, a groundbreaking AI drug discovery tool designed to significantly accelerate and reduce the cost of developing new medicines. This cloud-native pipeline, built on NVIDIA BioNeMo microservices, streamlines the most expensive research phases by predicting protein structures, generating novel drug-like molecules, and simulating their interactions in days instead of years.

How the GenMolVS Pipeline Works

GenMolVS provides an end-to-end AI pipeline for drug discovery. It accepts a raw protein sequence and uses a series of integrated AI models to predict the protein's structure, generate novel molecules, and simulate how they bind, ultimately producing a ranked shortlist of promising drug candidates.

The process begins with a raw protein sequence and concludes with a ranked shortlist of synthesizable drug candidates. As detailed in the official blog, the GenMolVS pipeline integrates several core AI modules:

• Protein structure prediction using AlphaFold2 and OpenFold2
• De novo molecule generation with NVIDIA MolMIM and GenMol models
• Diffusion-based docking via DiffDock for accurate binding poses
• Agentic reasoning, powered by the NeMo Agent Toolkit, to schedule simulations and flag high-confidence hits

Persistent reports that this approach can reduce a typical five-to-six-year preclinical cycle to approximately two years, saving tens of millions of dollars before the first animal study.

Why GenMolVS Matters for Drug Discovery

In drug development, speed is critical as the patent clock starts ticking early. By shifting costly wet-lab screening to GPU-accelerated simulations, GenMolVS allows chemists to test hypotheses virtually, dramatically reducing the need for experimental chemical synthesis. While 70% of large pharma teams aim to adopt generative AI by 2027, half are blocked by "production readiness." GenMolVS addresses this directly with containerized BioNeMo NIM services, audit logging, and LIMS integrations to ensure deployments meet stringent internal validation requirements.

Real-World Momentum for BioNeMo

GenMolVS enters a rapidly maturing ecosystem. Roche is building an AI factory with over 3,500 NVIDIA Blackwell GPUs to accelerate target validation, according to the Roche AI factory announcement. Similarly, an Eli Lilly co-innovation lab has committed up to $1 billion to scale BioNeMo-driven autonomous chemistry, as reported by HLTH.

Balancing Speed with Regulatory Compliance

Life sciences regulators demand traceable models, reproducible data, and documented decision-making. GenMolVS addresses this by including agentic document generation, which maps every simulation back to its underlying parameters. This feature directly answers a key critique of AI discovery tools. However, industry analysts caution that AI's true success will be judged in Phase III trials, where historical failure rates remain near 90%.

What to Watch Next

Persistent has announced that the next GenMolVS release will integrate multimodal safety predictors, allowing chemists to optimize for both potency and toxicity in a single loop. The team is also piloting closed-loop connections to robotic synthesis partners, advancing the vision of the self-driving lab that is gaining traction across pharmaceutical R&D.

---

What is GenMolVS and which NVIDIA technologies power it?

GenMolVS (Generative Molecules & Virtual Screening) is a production-ready AI pipeline that Persistent Systems built on NVIDIA BioNeMo NIM micro-services and the NeMo Agent Toolkit. The stack folds three core jobs into one seamless workflow:

• Protein Structure Prediction - AlphaFold2/OpenFold2 create 3-D target models from plain sequence
• De-novo Molecule Generation - MolMIM & GenMol networks design novel, drug-like candidates
• Generative Docking - DiffDock diffusion model ranks binding poses in minutes instead of hours

By chaining these modules with agentic orchestration, researchers move from raw sequence to a short-list of experiment-ready compounds without writing a single line of GPU code.

How much faster and cheaper can pre-clinical discovery become?

Customers report that GenMolVS can shrink the 5-6 year pre-clinical timeline to roughly two years and compress individual design-make-test-analyze cycles from months to days. Early adopters already see:

• 30-40 % reduction in wet-lab iterations
• 16-20 % hit rate for antibody leads vs. the historical 0.1 %
• Up to 70 % cut in structure-determination and HTS budgets before the first flask is touched

Persistent quotes a "months to days" acceleration for lead identification, a figure echoed in NVIDIA-backed projects at Roche and Eli Lilly.

Is the platform validated for regulated pharma environments?

Yes. The architecture is deployed as containerized NIM micro-services that run on-premises or in private cloud, giving IT and QA teams full audit trails, version pinning and rollback. Persistent couples this with:

• GxP documentation templates
• Model-cards that capture training data, metrics and limitations
• Human-in-the-loop checkpoints before every wet-lab hand-off

These controls satisfy early FDA feedback that AI models must be "explainable, reproducible and locked" once they enter the critical path toward an IND filing.

Which wet-lab steps still require human intervention?

AI can narrow millions of possibilities to a few hundred high-confidence molecules, but regulators still want physical data on solubility, permeability, tox and stability. GenMolVS therefore exports:

• Ranked SDF libraries with predicted QED, logP and synthetic accessibility scores
• A recommended "minimum viable assay" list so chemists can focus spend on 20-50 instead of 20,000 compounds
• Iteration loops that feed assay results back into the agent, letting the model refine its chemical space without drifting outside pre-set safety filters

Where is the product available and what does onboarding look like?

GenMolVS ships through Persistent's life-science practice and NVIDIA's partner portal. A typical kick-off:

1. One-week discovery workshop to lock target class, assay cascade and compliance scope
2. Two-week container deployment on client GPU cluster or NVIDIA DGX Cloud
3. Parallel four-week fine-tune using proprietary sequences and historic assay data
4. Go-live with SLA-backed inference throughput (≥1,000 ligands scored per hour on a single A100)

Support includes quarterly model updates, security patches and an optional "closed-loop lab" add-on that wires robotic instrumentation into the agent for 24/7 autonomous cycles.