parametrization of diffsplice to get meaningful results

[albamasmalavila]

Hello,

I performed a differential splicing analysis between two conditions for which we have 7 biological replicates per condition. I used the default parameters of the diffSplice command and I got 11 significant events (p-value < 0.05). However, when I plot the PSI values for these events, most of them don't really make sense or are not biologically relevant. In one case we have all samples in the two conditions with PSI of 0, so a delta PSI of 0, and however a p-value of 0.03. In other cases we have all samples with a PSI around 1 except one of them with a PSI around 0, with a delta PSI around 0.15 and a p-value around 0.03. I understand this is due to the way the significance is calculated with the empirical method and using the mean to calculate the delta PSI.

I'm thinking in changing some parameters like increasing the --lower-bound, increasing the --tpm-threshold, or using the median instead of the mean. Is there a standard cut-off that is recommended to use to avoid such cases, or should it be determined based on the distribution in each dataset? Also, would you recommend to filter the events before analysis based on the PSI values, similar to what is done in differential gene expression analysis where lowly expressed genes are filtered out?

Thank you very much for your help,

Alba

[EduEyras]

Hi Alba, I performed a differential splicing analysis between two conditions for

which we have 7 biological replicates per condition. I used the default parameters of the diffSplice command

Which method did you use for the differential splicing? "classical" or "empirical"?

and I got 11 significant events (p-value < 0.05). However, when I plot the PSI values for these events, most of them don't really make sense or are not biologically relevant. In one case we have all samples in the two conditions with PSI of 0, so a delta PSI of 0, and however a p-value of 0.03. In other cases we have all samples with a PSI around 1 except one of them with a PSI around 0, with a delta PSI around 0.15 and a p-value around 0.03. I understand this is due to the way the significance is calculated with the empirical method and using the mean to calculate the delta PSI. I'm thinking in changing some parameters like increasing the --lower-bound, increasing the --tpm-threshold, or using the median instead of the mean. Is there a standard cut-off that is recommended to use to avoid such cases, or should it be determined based on the distribution in each dataset? Also, would you recommend to filter the events before analysis based on the PSI values, similar to what is done in differential gene expression analysis where lowly expressed genes are filtered out?

You can use a filter for the expression of the transcript: *-f | --total-filter: Minimum total expression of the transcripts involved in the event (Default is zero). If used, it will filter out the events that do not reach this total expression value for the transcripts defining the event (the denominator of the PSI calculation).* You could also use a cutoff for the delta PSI, e.g. remove all cases with |delta PSI| < 0.05 The fact that you have 7 reps per condition, but very few clear significant cases could be that the reps have lots of variability. Did you check that they are group together or are similar to each other? I hope this helps Eduardo

…

Thank you very much for your help, Alba — Reply to this email directly, view it on GitHub <#224?email_source=notifications&email_token=ADCZKB6HH4LUE2Y42QO4D7D46FJSBA5CNFSL4Z3JMQ5C6L3HNF2C22DVMIXUS43TOVSS6NBVHA4DCNBXHEYDLJTSMVQXG33OVJZXKYTTMNZGSYTFMSSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCZKB34D6EBHLSLY3KCY2D46FJSBAVCNFSM6AAAAACZ2GC6OKVHI2DSMVQWIX3LMV43ASLTON2WKOZUGU4DQMJUG44TANI> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/ADCZKBZRG6JNVEXJLVKQVYT46FJSBA5CNFSL4Z3JMQ5C6L3HNF2C22DVMIXUS43TOVSS6NBVHA4DCNBXHEYDLJTSMVQXG33OVJZXKYTTMNZGSYTFMSSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/ADCZKBY335JIJPDCLSMHMJD46FJSBA5CNFSL4Z3JMQ5C6L3HNF2C22DVMIXUS43TOVSS6NBVHA4DCNBXHEYDLJTSMVQXG33OVJZXKYTTMNZGSYTFMSSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[albamasmalavila]

Hi Eduardo,

Thanks for your reply.

I used the empirical method.

About the -f option of the psiPerEvent command, I understand the units are TPM. Is this parameter equivalent to --tpm-threshold in the diffSplice command? Would you recommend a specific threshold for low expression?

About the sample similarity I didn't look at it, but it's a good point. Would you recommend doing PCA or clustering using the expression values (at the transcript level or event level) or using the PSI values?

Thanks,

Alba

[EduEyras]

Hi, They are similar, but one is calculated per sample and the other is an average across samples and conditions. Usually, it's good to plot the distribution of TPMs (usually in log10(TPM+1) scale or (TPM)^(1/10) scale to see the zeroes :-) ) to have an idea of a TPM value that could be considered too low to be a real expression. Yes, a PCA using gene or transcript expression values would help pinpoint samples that might be outliers. You'd expect the PCA to cluster the samples by condition. A pairwise correlation analysis of transcript or gene expression values can also help. Best Eduardo

…

On Fri, 5 Jun 2026 at 00:16, Alba Mas Malavila ***@***.***> wrote: *albamasmalavila* left a comment (comprna/SUPPA#224) <#224 (comment)> Hi Eduardo, Thanks for your reply. I used the empirical method. About the -f option of the psiPerEvent command, I understand the units are TPM. Is this parameter equivalent to --tpm-threshold in the diffSplice command? Would you recommend a specific threshold for low expression? About the sample similarity I didn't look at it, but it's a good point. Would you recommend doing PCA or clustering using the expression values (at the transcript level or event level) or using the PSI values? Thanks, Alba — Reply to this email directly, view it on GitHub <#224?email_source=notifications&email_token=ADCZKB5DFZ7EVT2VOURSU5D46GADRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINRSGI4TQMJWGUZKM4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4622981652>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCZKBZVSMSEFDXJMTAF4OL46GADRAVCNFSM6AAAAACZ2GC6OKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DMMRSHE4DCNRVGI> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/ADCZKB5NAARSXMEHWNCCGAL46GADRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINRSGI4TQMJWGUZKM4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/ADCZKB2CPXRHBMPUHTKKKVD46GADRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINRSGI4TQMJWGUZKM4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>

[albamasmalavila]

Thank you very much.

Alba