Skip to main content
proto-tools is an open source infrastructure layer, that provides access to a large collection of computational biology and biological AI tools behind a single, uniform Python interface. This guide covers how to set-up the package and run a tool. Open as a runnable notebook

​
Installation

To begin, pip install the package in a virtual environment. The base environment of proto-tools is intentionally light for maximum compatibility. The base package requires Python 3.10 or later.

​
Example tool call

All models and tools have their standard operations divided into functions that operate over the following abstractions:
  • Inputs: Contain the actual biological entities the tool operates over.
  • Configs: Contain runtime settings that can be adjusted to control tool behavior.
  • Outputs: The object that is produced by the tool.
Input plus Config feed into run_*(), which returns an Output These abstractions are Pydantic models, which declare each field with a type and the values it is allowed to take. Pydantic enforces those declarations the moment the object is constructed, so a value outside a permitted range is rejected immediately and names the offending field, rather than failing part-way through a run that has already loaded a model onto a GPU. Let’s show a quick example of how to call a tool. Here we will call Protenix to fold a protein complex. To begin, it may be helpful to look at our documentation page for Protenix, which contains a section on the Input, Config, and Output models and the various fields they define. Protenix — Structure Prediction On the documentation page above, we can see the run function and the various models it takes in as arguments. Let’s import them below.
python
python
And that’s it! All environment set-up is handled automatically for us under the hood. The next time this tool is called, inference will be subsantially faster as the environment and weights will already be set up on this machine.

​
Go deeper

For the complete runtime reference, covering identifier resolution, the registry and CLI surfaces, schema and documentation extraction, gated model weights, and calling conventions, consult the developer notes in the proto-tools repository: Finding and Calling ToolsIdentifier resolution, registry methods, schemas, docs extraction, gated weights, and calling patterns.

​
Next Steps

Continue with Tool Environments, the next guide in this series.

Tool Environments

How a tool’s isolated environment is built on first call and cached afterward.

Tool Persistence

Keep a model loaded across calls to skip repeated load times.

Device Management

How tools are placed on GPUs, with LRU eviction and CPU offload.

Cloud Inference

Dispatch tool runs to remote compute.