Questions about using HVG vs all genes for FICTURE/punkst factor selection and multisample approach

[ccruizm]

Hi, and thank you for continuing to develop and support the FICTURE pipeline. I’ve tested FICTURE as a standalone tool and it has worked well for our cohort. I’m now exploring punkst, particularly because the multi-sample functionality may help with integrating data from multiple tissue types in a diverse cancer cohort.

I had a question regarding gene selection for factorization. Have you tested or would you recommend running FICTURE/punkst using only highly variable genes versus using the full gene set? I am working with a CosMx 6000 gene panel. When increasing the number of K, I expect to observe factors that appear redundant; nevertheless, I’m wondering whether restricting to highly variable genes could help reduce noise or improve factor quality, or whether it is generally better to keep all genes.

I’d also appreciate any guidance on best practices for using punkst in a multi-sample, heterogeneous cohort. For context, we have cancer TMAs with ~3 TMAs per donor across ~10 slides, and cores may come from different tissue types. My understanding is that the multi-sample preparation could train a model per core and then integrate across slides to learn a shared factor set. Previously, I ran FICTURE separately for each core, across different K values (e.g., 4–20), then compared factors across runs to identify robust, recurring gene programs. Am I correct in assuming that the multi-sample approach is intended to reduce the need for these extensive cross-run comparisons by jointly learning from all samples?

Any guidance or best practices would be greatly appreciated. Thanks!

[Yichen-Si]

1. Based on my limited experience, using all the genes is generally better with targeted gene panels; HVG might help for sequencing based platforms where all transcripts are captured. Doesn't hurt to try both though.
2. What is a core? What is TMA? You are right that multi-sample pipeline is intended to jointly learn factors to help comparison. I'm not sure what happens when you have different tissue types though, since there are many tissue-specific cell types, you might find some factors only occur in a subset of samples, which should be expected. As long as there is no strong technical artifacts contributing to the difference between samples it may be fine to learn the model jointly.

[ccruizm]

Thanks for the response! That's helpful.

Sorry for not explaining earlier. A Tissue Microarray (TMA) is a slide containing multiple small tissue punches (called "cores") arranged in a grid. Each core is punched out from a tumor biopsy, so a single TMA slide can contain cores from different patients or different regions of the same tumor. In our case, we have ~3 cores per donor across ~10 TMA slides, with each core spanning a few FOVs.

You mentioned that some factors may only occur in a subset of samples when different tissue types are present, which makes sense for our cohort. Since our cores can vary substantially in cellular composition (tumor-rich vs. stromal vs. immune-infiltrated), I'm planning to run a broad range of K values, roughly from 5 up to 50, to capture that diversity. At lower K I'd expect broad factors shared across cores, while at higher K I'd expect to resolve tissue-specific or rare cell type factors that only appear in a subset of samples. Does that reasoning sound ok? Also, would you happen to have any recommendations or criteria for selecting the best K? For instance, is there a diagnostic metric within PunkST that helps evaluate model fit, or do you typically rely on the biological interpretability of the factors?

I would also like some advice on which parameters I should change from the default values. Since the tissue sections are relatively small, I want to make sure the parameters I'm using are reasonable. I'd appreciate any input on the following:

1. For the hexagonal binning, I'm currently using a hex_grid_dist of 12 and min_count_hex of 50. Given that CosMx has a pixel size of ~0.12 µm/px, would you recommend adjusting the hex grid distance? Also, is 50 a reasonable minimum count threshold for the hex bins with a ~6K gene panel, or should it be higher or lower?
2. For the topic model training, I'm using n_epochs set to 1, min_count_train at 50, and minibatch_size of 512. My main concern is whether 1 epoch is enough for convergence, especially for small cores that may have relatively few hexagonal units. Is there a rule of thumb for the minimum number of epochs, or does it depend on the dataset size? Similarly, should min_count_train be adjusted for CosMx transcript density?
3. For pixel decoding, I'm using pixel_res of 0.5 and min_init_count of 5. Are these reasonable defaults for CosMx, or would you suggest going finer or coarser on the pixel resolution?
4. Lastly, when using multisample-prepare to jointly model across cores, is there a recommended min_total_count_per_sample to filter out low-quality cores before joint training?

Any rules of thumb for these would be very helpful. Thanks again!