NVIDIA BioNemo Model Use case: Using GenMol in a Molecule Optimization Workflow

This workflow on Vecura uses GenMol – a model belonging to the BioNeMo collection, to turn a single flagged compound into a batch of drug-like alternatives in seconds, then validates every one of them so a chemist can act on the results with confidence.

Lead optimization has a chemistry-idea problem. When a promising compound carries a toxicity liability, the medicinal chemist's job is to imagine alternative structures that keep the good properties and shed the bad ones, one careful redesign at a time. Generative models change the pace of that step. Instead of proposing analogs by hand, a chemist can hand a scaffold to a model and get a batch of novel, drug-like alternatives to react to. In the NVIDIA BioNeMo model collection, GenMol does exactly that, and it does it fast enough that the idea-generation step stops being the bottleneck.

This case study follows one real run on Vecura. We started from a NEK2 inhibitor that docked well but carried a serious toxicity flag, used GenMol to generate a batch of optimized alternatives, and then ran every alternative back through toxicity prediction and docking. The story here is what GenMol contributes to that loop, and why pairing fast generation with fast validation is the part that makes it useful.

 

The starting point: a good compound with a bad flag

The parent compound is an isoxazole-based NEK2 inhibitor. On Vecura, it docked into the NEK2 structure (PDB 2W5A) with a top pose near minus eight kcal per mole, a moderate-to-strong predicted affinity, and it scored well on drug-likeness with a QED of 0.83, a molecular weight of 313, a logP of 2.08, and high predicted intestinal absorption.

The problem showed up in the toxicity screen. The compound carries a diacylhydrazide motif, and this acylhydrazide chemical class is associated with metabolic liability and, in some cases, liver toxicity through reactive-metabolite formation. That single structural feature is enough to stall a compound that otherwise looks advanceable. This is the exact situation optimization is meant to solve, and the exact situation a generative model can accelerate.

What GenMol contributed

A batch of ideas, in seconds

GenMol took the parent scaffold and generated a batch of candidate molecules in a single pass, with the sampling depth set to 200 steps. Inference completed in about 16 seconds, and the full job, including setup and teardown, finished in under 80 seconds. Of 19 molecules generated, 18 were chemically valid, and the batch was genuinely drug-like: the top candidate reached a QED of 0.95, and the strongest ten candidates clustered in the 0.65 to 0.95 range.

Just as useful as the scores was the direction the batch pointed. The top candidates were dominated by sulfonamide, sulfone, and phosphonate linkages in place of the flagged diacylhydrazide. For a chemist, that is an immediately actionable read: GenMol is suggesting these functional-group families as the most promising replacements for the liable motif, turning an open-ended redesign question into a short list of concrete chemistry to explore.

This is the core value. GenMol does not replace medicinal-chemistry judgment. It front-loads it. In the time it takes to read a single ADMET report, the model proposes a batch of novel, valid, drug-like alternatives and surfaces which structural directions are worth a chemist's attention.

Why generation alone is not enough

A batch of ideas is only useful if you can tell the good ones from the flawed ones quickly, and this is where the Vecura loop matters. Every GenMol molecule was passed back through ADMET-AI for toxicity and re-docked into NEK2, so each suggestion arrived with the evidence needed to accept, reject, or refine it. The most instructive example is the batch's standout candidate, Molecule 6, a phosphorus-linked analog of the isoxazole scaffold.

On the toxicity endpoints that motivated the whole exercise, Molecule 6 is a striking improvement over the parent.

Table 1. ADMET-AI predictions, parent compound versus GenMol Molecule 6.

Predicted mutagenicity, clinical-toxicity, and cardiotoxicity risk all fell sharply, and Molecule 6 re-docked into NEK2 at essentially the same score as the parent (near minus eight kcal per mole, within the noise of the method). On its face, GenMol proposed a molecule that keeps the binding and clears three of the toxicity flags, in seconds.

Figure 1. Docking poses in the NEK2 binding site (PDB 2W5A). (A) Parent compound. (B) Molecule 6.

Reading the result honestly, which is the point of the loop

The validation layer also does what a good validation layer should: it flags what still needs work, and it does so before anyone invests in synthesis. Two things stand out for Molecule 6. Its DILI (liver injury) probability, at 0.77, remains elevated, so the specific liability the project set out to remove is improved but not resolved. And the way GenMol achieved the tox reduction, by replacing the amide-hydrazide linker with a phosphorus-based group, is a substantial change to the scaffold that also alters the hydrogen-bonding chemistry the parent relied on. A charged phosphorus linker in particular can affect permeability and absorption in ways a single QED value does not capture.

That is not a mark against GenMol. It is the workflow working. A generative model that only ever produced perfect molecules would not need validation. The value of running GenMol on Vecura, with ADMET-AI and docking in the same loop, is that a bold suggestion like Molecule 6 comes back annotated with both its wins and its open questions in one screen. The chemist gets an idea worth pursuing and a precise list of what to check next, instead of a black-box answer to take on faith.

Why run Genmol on Vecura

The model is only as useful as the loop around it, and that loop is what Vecura provides. Three things made this run work.

Generation and validation in one place. GenMol, ADMET-AI, and docking against NEK2 ran together on Vecura, so every generated molecule was scored and docked without moving data between tools. Ideas and evidence lived in the same pipeline.

Fast enough to iterate. Sixteen-second inference and sub-eighty-second jobs mean a chemist can regenerate, re-screen, and refine within a working session rather than across days.

Results a chemist can act on. Because every suggestion returned with toxicity predictions and a docking pose, the batch could be triaged immediately, keeping human judgment on the decisions and handing the model the legwork.

Beyond GenMol: The BioNeMo Collection on Vecura

GenMol is just one of many BioNeMo models now available on Vecura. Researchers can also run CodonFM and RNAPro for RNA understanding and structure prediction, La‑Proteina, Proteina, Proteina‑Complexa, and ProtComposer for protein generation and binder design, and optimization models like KERMT, ReaSyn, and DualBind for property prediction, synthesizability, and binding affinity.

Together, these models span proteins, RNA, and small molecules - enabling scientists to generate, validate, and triage candidates in a single pipeline, and accelerate discovery with GPU‑powered inference speeds.

Notes: This was a single exploratory run, and the standout candidate trades one set of liabilities for a new set of questions, as early optimization suggestions often do. Predicted ADMET and docking scores are directional signals to prioritize experiments, not substitutes for them. What the run demonstrates is the operational result this piece is about: on Vecura, GenMol turned one flagged compound into a batch of valid, drug-like, fully-validated alternatives in seconds, and pointed a chemist at the chemistry worth pursuing next.

Bring your optimization problem to Vecura and pair BioNeMo models like GenMol with the complete suite of integrated workflows, from ADMET prediction and docking to structure generation, property scoring, and beyond. On Vecura, you can generate, validate, and triage a full batch of alternatives in one place, without moving data between tools, and accelerate your next design decision with actionable results.