How Neptune Underpins Bioptimus’ Decisions in Training Biology Foundation Models

Watch now

Bioptimus is a Paris-based startup founded in 2023 with the ambition to build a multimodal, multiscale foundation model for biology. Their goal is to bridge different levels of biological data, from DNA, proteins, cells, tissue, to patient-level data, within one unified model.

The technical team of ~15 researchers and engineers (many from DeepMind and computational biology) runs multi-day training and ablation experiments, tracking everything from hardware utilization and throughput to model internals and checkpoints.

The challenge

Bioptimus started out using Weights & Biases (W&B) for experiment tracking. While useful early on, they quickly ran into limitations around pricing (tied to training hours and users) and foresaw scalability bottlenecks as their workloads grew. They needed a tool that could keep pace with long-running training jobs, multi-GPU logging, and the ability to resume or fork experiments without friction.

When you train foundation models, you need an experiment tracker. We run experiments that are extremely costly and time-consuming, and we want to extract as much information as possible from all of them. So we need an extremely precise tracker that works under various conditions, online and offline, and allows us to connect experiments in non-trivial ways. That’s why Bioptimus decided to use Neptune.

Zelda Mariet VP of Research

Faced with these needs, Bioptimus chose Neptune because it was purpose-built for foundation model training, with flexible APIs, offline support, and responsive performance at scale.

Migration with minimal disruption

Switching from W&B to Neptune was straightforward. The team continued logging training metrics, hyperparameters, and system statistics with little code change. Neptune’s technical support team helped guide the process, adapt it to Bioptimus’s infrastructure, and support the migration of historical data. The quality of the API also made integration seamless: it felt like plug and play within their codebase.

I remember the migration well because I’m the one who did it. Changing your experiment tracker is a major thing, especially for a foundation model company. But the documentation from Neptune was straightforward, and for something that central, migrating was surprisingly easy.

Zelda Mariet VP of Research

Monitoring long training runs end-to-end

Neptune is now central to Bioptimus’s entire workflow: multi-day pretraining, fine-tuning, systematic ablations, and benchmarking against competitors. 

They log training losses, hyperparameters, GPU usage, throughput, and system-level metrics like CPU, memory, and disk I/O. This allows Bioptimus to ensure proper hardware utilization, as well as guide decisions to use resources efficiently.

When branching experiments, they rely on Neptune’s forking functionality: resuming training from a checkpoint while preserving full lineage, or spawning multiple variants from the same base run. This allows them to test hyperparameters, diagnose infrastructure restarts, and compare trajectories without losing context.

They also use Neptune to share results with stakeholders who have different levels of technical expertise. Some want detailed panels with every metric, while others prefer high-level summaries. Neptune’s reports make it easy to provide both views.

And all that happens without interrupting or delaying their work. 

In terms of performance, I was pleasantly surprised. Neptune never showed up in profiles in an obtrusive way, and logging overhead was minimal, which is necessary at scale. The UI is very fast, plots work well, and everything is logged almost immediately in real time. Speed-wise, I’m very happy with Neptune.

Alexander Immer Senior Research Scientist

Offline logging for restricted environments

Not all of Bioptimus’s compute environments allow internet access. In particular, some of their patient-data clusters run fully offline for compliance and security reasons. With Neptune, they can log experiments locally and sync later, preserving a complete record without breaking security constraints.

We work with a variety of experimental environments and cloud providers, and some don’t have internet access. We needed the ability to run the tracker offline and sync later. Neptune listened to our feedback and built this feature for us.

Felipe Llinares-López VP of AI

The results

Since adopting Neptune, Bioptimus has:

• Made Neptune a central system of record underpinning model development decisions.
• Increased trust across the team that experiment results are accurate, reproducible, and tied to the correct model version.
• Reduced wasted GPU time by catching bottlenecks and hardware failures during multi-day training.
• Enabled secure offline logging on restricted patient-data clusters without losing experiment history.
  

Neptune is central in a lot of what we do. And a lot of our decision-making relies on experimentation. If you cannot trust your results, you cannot make decisions. We rely on Neptune’s API and visualizations to compare runs, assess new model features, and revisit past versions. It underpins a lot of the decision-making of the company.

Rodolphe Jenatton CTO

A big thank you to Mathilda, Zelda, Rodophe, Felipe, Alexander, Dasha, and the rest of the amazing Bioptimus team for helping create this case study!

Product resource

Neptune vs Weights & Biases

Read more Product resource

How Navier AI uses Neptune to Rapidly Iterate on Physics Foundation Models

Read more