Copyright (C) 2021 Upinder S. Bhalla, National Centre for Biological Sciences, Tata Institute of Fundamental Research, Bangalore, India.
All code in HOSS is licensed under GPL 3.0 or later.
HOSS provides a set of methods for performing hierarchical optimization of signaling and other models. The key idea is that many signaling and cellular processes are modular and signal flow in them is hierarchical. This makes it possible to break up large optimization problems into small modules such that each module that is optimized depends only on its own parameters, and the properties of upstream blocks. The upstream blocks are evaluated first, and held fixed as one goes successively deeper into the model. Our colleagues Radulescu and co-workers have shown mathematically that this is both more efficient, and gives better fits, than 'flat' optimization where the whole set of parameters is fitted at the same time. We have demonstrated that this is also true in practice on messy real-world optimization problems.
The HOSS code performs simple hierarchical optimization, flat optimization, and two multi-start methods for optimization which use hierarchical optimization internally. The multi-start methods are even better than plain HOSS, but require more resources.
The structure of an optimization pipeline is defined using a configuration file in JSON. The schema for this configuration file is provided as part of the HOSS project.
The HOSS project has been written up and is on bioRxiv:
"Hierarchical optimization of biochemical networks"
Nisha Ann Viswan, Alexandre Tribut, Manvel Gasparyan, Ovidiu Radulescu, Upinder S. Bhalla
https://doi.org/10.1101/2024.08.06.606818
HOSS takes a configuration file argument, and has numerous other options. The configuration file specifies all aspects of the optimization, notably a start model, a set of experiments defined in FindSim format, a list of parameters to tweak, and optional bounds on the parameters. These are all organized into the hierarchy chosen for the optimization.
Here is a typical command-line invocation of HOSS:
hoss Config/D3_b2AR_hoss.json --algorithm COBYLA --outputDir OPT_D3_b2AR_COBYLA
For examples for running HOSS, including scripts, experimental datasets and configuration files, see the repository for generating the figures for the paper: hossFigs
HOSS depends on FindSim, HillTau and MOOSE.
FindSim is the Framework for Integrating Neuronal Data and Signaling Models.
FindSim defines experiments, specially biochemcial experiments, in a standard JSON format. A FindSim file specifies:
- The design, stimuli, and experimental conditions of an experiment
- The readouts of an experiment.
- A model on which this experiment can be run.
The FindSim code does the following:
- Reads a FindSim file, a model, and various other optional arguments.
- Runs the model on the experiment definition
- Compares the model output to the experimental readouts defined in the file and returns a score which says how well the model output matched experiments.
- Optionally, it modifies parameters in the model, as required by the optimization routine.
It is this score which the HOSS script uses to compute the value of the objective function for its optimization.
Note that FindSim is agnostic to model type, and hence so is HOSS. At present they can work with HillTau and MOOSE models.
HillTau is a format and program for specifying and running abstracted models of cellular signaling. These abstracted models retain a direct mapping to molecules, hence it is easy to use them in optimization calculations.
MOOSE is the Multiscale Object Oriented Simulation environment. It is good for running ODE-based signaling models defined in the standard SBML, as well as multiscale models combining electrical and chemical signaling.