The "max difference of decontaminated expressions"

[Peevin]

The "max difference of decontaminated expressions"

[zijianni]

Hi @Peevin , there is no standard way to define the appropriate maxit value for different samples.

The convergence is defined numerically by the other argument tol. Even with the same tol value, the convergence speed is also dependent on the properties of input expression matrix (e.g. for one matrix you might reach numerical convergence within 20 iterations, but for another one you might need 50 iterations).

In practice you can keep using the same default value across your samples - it's the first few iterations that make the most significant changes to the gene expression. When the number of iterations grows bigger, the changes will become smaller, as you can see from the changes in likelihood value and max difference.

Let me know if that explains it or if you have additional questions.

[Peevin]

Hi @Peevin , there is no standard way to define the appropriate maxit value for different samples.

The convergence is defined numerically by the other argument tol. Even with the same tol value, the convergence speed is also dependent on the properties of input expression matrix (e.g. for one matrix you might reach numerical convergence within 20 iterations, but for another one you might need 50 iterations).

In practice you can keep using the same default value across your samples - it's the first few iterations that make the most significant changes to the gene expression. When the number of iterations grows bigger, the changes will become smaller, as you can see from the changes in likelihood value and max difference.

Let me know if that explains it or if you have additional questions.

Thanks very much for your reply.

Based on your response, I checked the output of spotclean in the RStudio console. As you mentioned, with increasing iterations, the change in max difference starts to become very small. In my previous runs, when it reached around 30 iterations, the change in max difference was already very small. Does this mean it has converged, or should I increase the number of iterations to ensure it stops when the max difference is below 1? Another question is, if I increase the number of iterations excessively to ensure it stops when the max difference is below 1, would it result in excessive data correction?

[zijianni]

Again, the definition of convergence is numerical and depends on tol. If you increase tol you could have a numerical convergence within 30 iterations.

For your second question, No - What the algorithm does is to find the parameter value that maximizes the likelihood (locally, to be rigorous). The iteration process is to move the current estimation closer and closer to the maximizer, but won't move beyond it.

In practice, 30 iterations should be sufficient, as any future iterations are making smaller and smaller changes to the estimation, but they still take the same amount of time to compute, which is not fast, to be honest.

[Peevin]

@zijianni Thank you very much! After careful consideration, we have made the decision to set a bigger maxit parameter and utilize the default tol setting for running spotclean. This will help us obtain "cleaner" spatial transcriptome data, which is crucial for our downstream analysis.