Runtime of Fast Higashi - ma-compbio/Higashi GitHub Wiki
Unless stated, all numbers reported at running Fast-Higashi at 500kb resolution on all autosomal chromosomes. All numbers are consistent with the uploaded tutorial notebook.
Runtime for processing
| Dataset | Number of cells | Runtime - contact pairs 2 sparse matrix | Runtime - sparse matrix to sparse tensor |
|---|---|---|---|
| Tan et al. Cell 2021 (mousebrain) | 1954 | 72.79s | 28.50s |
| Tan et al. Cell 2021 (mousebrain + visual cortex) | 3646 | 122.20s | 50.50s |
| Lee et al. Nature Methods 2019 (human PFC) | 4238 | 100.99s | 75.21s |
| Liu et al. Nature 2021 (mouse hippocampus) | 5398 | 85.35s | 65.91s |
| Heffel et al. biorxiv 2022 (human developing brain) | 53063 | 955.64s | 841.25s |
| Kim et al. Plos Compbio 2020 (subset) | 6388 | 26.59s | 40.39s |
| Kim et al. Plos Compbio 2020 (all) | 16707 | 43.39s | 95.93s |
| Ramani et al. Nature Methods 2017 (ML1 + ML3 - 1mb) | 620 | 5.04s | 7.19s |
| Ramani et al. Nature Methods 2017 (all - 1mb) | 2271 | 16.51s | 18.21 s |
Runtime for tensor decomposition
main factor that decides are hyperparameters:
- do linear convolution or partial random walk with restart or not
- tolerance. (Suggested 3e-4 for high-cov data with medium number of cells, 2e-5 for low-cov data or data with large number of cells. Overall, one should try to run with 3e-4 first, inspect the results, and fine tune the model with 2e-5 or even smaller threshold when needed. Fast-Higashi will resume previous training. )
| Dataset | Number of cells | Hyperparam: Tol | Hyperparam: conv / rwr | Runtime - tensor decomposition |
|---|---|---|---|---|
| Tan et al. Cell 2021 (mousebrain) | 1954 | 3e-4 | F/F | 137.82s |
| Tan et al. Cell 2021 (mousebrain) | 1954 | 2e-5 | F/F | 527.38s |
| Tan et al. Cell 2021 (mousebrain) | 1954 | 3e-4 | F/T | 150.15s |
| Tan et al. Cell 2021 (mousebrain + visual cortex) | 3646 | 3e-4 | F/T | 147.13s |
| Lee et al. Nature Methods 2019 (human PFC) | 1954 | 3e-4 | F/F | 228.38s |
| Liu et al. Nature 2021 (mouse hippocampus) | 5398 | 2e-5 | F/T | 559.02s |
| Heffel et al. biorxiv 2022 (human developing brain) | 53063 | 2e-5 | F/F | 899.69s |
| Kim et al. Plos Compbio 2020 (subset) | 6388 | 2e-5 | T/T | 1157.30s |
| Kim et al. Plos Compbio 2020 (all) | 16707 | 2e-5 | T/T | 1375.53s |
| Ramani et al. Nature Methods 2017 (ML1 + ML3 - 1mb) | 620 | 2e-5 | T/T | 71.32s |
| Ramani et al. Nature Methods 2017 (all - 1mb) | 2271 | 2e-5 | T/T | 230.51s |