Drug discovery is changing faster than at any point in its history. At Proxima, we are building agentic workflows to accelerate how new proximity-based medicines are discovered, from understanding disease biology to designing and testing molecules in the lab.

We design proximity modulators, or ProMods: molecules that induce, modulate, or block protein interactions, including molecular glues, PROTACs, and PPI inhibitors. Success in this field requires advances across many disciplines at once: foundation model training, chemistry, biology, proteomics, structural modeling, and experimental design. AI agents are helping us connect these domains more tightly, turning data into hypotheses, hypotheses into experiments, and experiments back into better models.

But agents are only as powerful as the scientific tools they can use. NVIDIA BioNeMo and the BioNeMo Agent Toolkit provide a growing set of these tools built to bring agentic workflows to the demands of proximity-based drug discovery.

Across our platform, our collaboration with NVIDIA is already accelerating work at multiple layers of the discovery stack. The examples below are not exhaustive, but they illustrate how advances in accelerated biology, large-scale modeling, and inference infrastructure compound across our pipeline.

Learning from proprietary interactomics data, at unmatched scale. Our models learn from the vast space of protein interactions we are able to sample with NeoLink, our proprietary platform for mapping protein interactions across the proteome. To train on this corpus, the majority of which is absent from any public database, we need to generate and update sequence alignments continuously as new data arrives. The faster we align and learn from new data, the faster our agents can design the next round of experiments. BioNeMo's GPU-accelerated Multiple Sequence Alignment (MSA Search NIM) speeds up our MSA prep by ~20x, unlocking the model scaling that lets us go after ever-harder targets and better understand disease biology.

Modeling the large biological assemblies that matter for proximity modulation. Proximity modulators often work by reshaping or stabilizing multi-protein complexes, not simply by binding single proteins in isolation. These assemblies can involve thousands of residues and exceed the practical limitations of existing folding workflows. NVIDIA Fold-CP context parallelism in BioNeMo allows us to model much larger biological systems, helping us study ProMods in a more complete biological context.

Accelerating model training and inference for harder targets. Speed matters at every stage of model development and deployment. NVIDIA TensorRT delivers a 1.6 to 1.9x throughput gain for Neo inference, while NVIDIA cuEquivariance roughly doubles the speed of both our multi-node training and inference workflows. This acceleration lets us train larger models, evaluate more biological hypotheses, and design more molecules against difficult targets. Traditionally, standing up a new inference engine could take days; with the BioNeMo Agent Toolkit and its bundled skills, we integrated ours in just a few hours.

By accelerating the discovery of proximity modulators, NVIDIA BioNeMo is helping us turn new biology into therapeutic hypotheses faster and move promising programs toward patients sooner. We are proud to build that future alongside NVIDIA as they launch the BioNeMo Agent Toolkit at BIO 2026.