Different number of reads for same pod5 with different models

[KunFang93]

Hi,

I tried iterative strategy to finer tuning the model. I trained three models with
round0

#!/bin/bash

MODEL_PATH="dna_r10.4.1_e8.2_400bps_hac@v5.0.0"
REFERENCE="/mnt/data/kfang/reference/sacCer3_mm/sacCer3.mmi"
MIN_ACC=0.95

# Create necessary directories if they do not exist
#mkdir -p round1/test

# Loop over parts from 00 to 12
#for i in $(seq -w 0 12)
#do
#  # Create the directory for each part if it doesn't exist
#  mkdir -p "round1/part${i}"
#  bonito basecaller "$MODEL_PATH" \
#    "data_pod5_lig" \
#    --reference "$REFERENCE" \
#    --recursive \
#    --save-ctc \
#    --min-accuracy-save-ctc "$MIN_ACC" \
#    --read-ids "all_part${i}" \
#    > "round0/part${i}/basecalls_${i}.bam"
#done

# Create a comma-separated list of part directories for combining
#combine_list=$(seq -w 0 12 | sed 's/^/round1\/part/' | paste -sd "," -)
#python bonito_basecall_parts_combine_large.py -s "$combine_list" -o ./round0_combine

bonito train --epochs 1 --lr 5e-4 --directory ./round0_combine/ ./round0_combine/fine-tuned-model --pretrain "$MODEL_PATH"

round1

#!/bin/bash

MODEL_PATH="./round0_combine/fine-tuned-model/"
REFERENCE="/mnt/data/kfang/reference/sacCer3_mm/sacCer3.mmi"
MIN_ACC=0.95

# Create necessary directories if they do not exist
#mkdir -p round1/test

# Loop over parts from 00 to 12
#for i in $(seq -w 0 12)
#do
#  # Create the directory for each part if it doesn't exist
#  mkdir -p "round1/part${i}"
#  bonito basecaller "$MODEL_PATH" \
#    "data_pod5_lig" \
#    --reference "$REFERENCE" \
#    --recursive \
#    --save-ctc \
#    --min-accuracy-save-ctc "$MIN_ACC" \
#    --read-ids "all_part${i}" \
#    > "round1/part${i}/basecalls_${i}.bam"
#done

# Create a comma-separated list of part directories for combining
#combine_list=$(seq -w 0 12 | sed 's/^/round1\/part/' | paste -sd "," -)
#python bonito_basecall_parts_combine_large.py -s "$combine_list" -o ./round1_combine

bonito train --epochs 1 --lr 5e-4 --directory ./round1_combine/ ./round1_combine/fine-tuned-model --pretrain "$MODEL_PATH"

And I exported round0_combine/fine-tuned-model and round1_combine/fine-tune-model with bonito export

bonito export --output dna_bonito_model_r0 round0_combine/fine-tuned-model/
bonito export --output dna_bonito_model_r1 round1_combine/fine-tuned-model/

When I check the performance of different models, I found that the total number of reads are different from different model:
dorado official support sup model 2297227 in total

dorado basecaller sup,5mC_5hmC,6mA --kit-name SQK-RBK114-96 -r --output-dir ./ --reference sacCer3.mmi --mm2-opts "-k 15 -w 10" pod5/

samtools flagstat -@ 40 ../calls_2025-02-22_T21-36-58.bam
2297227 + 0 in total (QC-passed reads + QC-failed reads)
1894467 + 0 primary
327805 + 0 secondary
74955 + 0 supplementary
0 + 0 duplicates
0 + 0 primary duplicates
1717024 + 0 mapped (74.74% : N/A)
1314264 + 0 primary mapped (69.37% : N/A)
0 + 0 paired in sequencing
0 + 0 read1
0 + 0 read2
0 + 0 properly paired (N/A : N/A)
0 + 0 with itself and mate mapped
0 + 0 singletons (N/A : N/A)
0 + 0 with mate mapped to a different chr
0 + 0 with mate mapped to a different chr (mapQ>=5)

round0_model 1880743 in total

dorado basecaller /dna_bonito_model_r0 --kit-name SQK-RBK114-96 -r --output-dir ./ --reference sacCer3.mmi --mm2-opts "-k 15 -w 10" pod5/

samtools flagstat -@ 40 calls_2025-03-12_T15-08-54.bam
1880743 + 0 in total (QC-passed reads + QC-failed reads)
1554593 + 0 primary
266328 + 0 secondary
59822 + 0 supplementary
0 + 0 duplicates
0 + 0 primary duplicates
1589597 + 0 mapped (84.52% : N/A)
1263447 + 0 primary mapped (81.27% : N/A)
0 + 0 paired in sequencing
0 + 0 read1
0 + 0 read2
0 + 0 properly paired (N/A : N/A)
0 + 0 with itself and mate mapped
0 + 0 singletons (N/A : N/A)
0 + 0 with mate mapped to a different chr
0 + 0 with mate mapped to a different chr (mapQ>=5)

and round1_model 1753039 in total

dorado basecaller /dna_bonito_model_r1 --kit-name SQK-RBK114-96 -r --output-dir ./ --reference sacCer3.mmi --mm2-opts "-k 15 -w 10" pod5/

samtools flagstat -@ 40/bonito_r1/calls_2025-03-12_T15-45-24.bam
1753039 + 0 in total (QC-passed reads + QC-failed reads)
1420365 + 0 primary
265031 + 0 secondary
67643 + 0 supplementary
0 + 0 duplicates
0 + 0 primary duplicates
1595665 + 0 mapped (91.02% : N/A)
1262991 + 0 primary mapped (88.92% : N/A)
0 + 0 paired in sequencing
0 + 0 read1
0 + 0 read2
0 + 0 properly paired (N/A : N/A)
0 + 0 with itself and mate mapped
0 + 0 singletons (N/A : N/A)
0 + 0 with mate mapped to a different chr
0 + 0 with mate mapped to a different chr (mapQ>=5)

I wondered if

1. these differences are as expected?
2. what's the potential cause for these?
3. how could I mitigate the reduction of the total reads number?
4. the procedure of training looks reasonable?

Apologize for so many questions. Thank you so much for helping!

Best,
Kun