loopless sampling

[hariszaf]

Thermodynamics constraints are rather important in metabolic modeling.
If not considered, they would lead to thermodynamically infeasible/implausible flux loops within flux sample.

Loopless space though is a non-convex one.
Approaches such as the LooplessFluxSampler could be used for more valid sampling from the biological point of view.

Further literature:

• Elimination of Thermodynamically Infeasible Loops in Steady-State Metabolic Models
• Extreme Pathways and Kirchhoff's Second Law

[vikasgrewal16]

I want to contribute to this issue but i am having a difficulty to understand this. Can you please tell me how can I do this @hariszaf

[Tushar7012]

Hi @hariszaf and @vissarion ,

I took some time to carefully go through this issue, the referenced literature, and Dingo’s current sampling workflow to make sure I understand the problem correctly before starting any work.

As I understand it, the main concern is that standard steady-state flux sampling can generate thermodynamically infeasible internal cycles. While these flux vectors are mathematically valid under mass-balance constraints, they violate thermodynamic feasibility and therefore reduce the biological interpretability of the samples. Enforcing looplessness addresses this, but at the cost of moving to a non-convex feasible space, which makes direct integration with existing convex polytope samplers non-trivial.

Rather than attempting a full implementation upfront, my intention is to proceed incrementally. First, I want to clearly identify how sampling is currently structured in Dingo and what assumptions are made about convexity. From there, I plan to explore how loopless approaches (such as those described in the cited work) could conceptually fit into the existing design — for example, whether this would make more sense as an optional sampler, a constrained variant, or an approximate/post-processing approach.

As an initial step, I’m aiming to put together a small design-level proposal that outlines a feasible integration path, explicitly states limitations, and avoids overpromising on performance or generality. Based on feedback, this could then evolve into a minimal prototype or experimental feature rather than a fully optimized solution.

I’d appreciate any guidance on whether this direction aligns with your expectations for this issue, and if there are particular parts of the codebase or prior design decisions I should focus on while exploring this further.

Thanks for your time and for pointing out this important aspect of biologically valid sampling.