Search

Additional file 5: Figure S5. Accuracies of MLR + maSigPro versus CPMTPp + maSigPro. The bars, top whisker, and bottom whisker represent mean, maximum, and minimum values of accuracies by threefold cross-validation, respectively. a HCV dataset. b MS dataset.

Additional file 1: Figure S1. The pseudo-code of CPMTPp. This code predicted a therapy response of a patient. $${{\varvec{x}}}^{\left({t}_{r}\right)}=({x}_{1}^{{(t}_{r})},\dots ,{x}_{l}^{{(t}_{r})}) (r=1,\dots ,R)$$ x t r = ( x 1 ( t r ) , ⋯ , x l ( t r ) ) ( r = 1 , ⋯ , R ) : gene expression levels collected by the patient at time point “ $${t}_{r}$$ t r ”. “ $$l$$ l ” was the number of genes in the gene subset. $${{\varvec{w}}}^{({t}_{r})}={({w}_{1}^{\left({t}_{r}\right)},\dots ,{w}_{l}^{\left({t}_{r}\right)})}^{\top }$$ w ( t r ) = ( w 1 t r , ⋯ , w l t r ) ⊤ : wights of the logistic regression at time point “ $${t}_{r}$$ t r ”. $${w}_{0}^{{t}_{r}}$$ w 0 t r was a constant term. $$\widehat{y}$$ y ^ was the predicted “sensitive” or “not sensitive” of the patient.

Additional file 2: Figure S2. The pseudo-code of CPMTPg:step1. This code created a gene pool by the step1 of CPMTPg. $${{\varvec{X}}}^{({t}_{r})}=\left({{\varvec{x}}}_{1}^{\left({t}_{r}\right)},\dots ,{{\varvec{x}}}_{l}^{\left({t}_{r}\right)}\right); {{\varvec{x}}}_{j}^{\left({t}_{r}\right)}={\left({x}_{j}^{\left(1,{t}_{r}\right)},\dots ,{x}_{j}^{\left(N, {t}_{r}\right)}\right)}^{\top } (j=1,\dots ,p)$$ X ( t r ) = x 1 t r , ⋯ , x l t r ; x j t r = x j 1 , t r , ⋯ , x j N , t r ⊤ ( j = 1 , ⋯ , p ) : gene expression levels of “ $$p$$ p ” genes × “ $$N$$ N ” subjects at time point “ $${t}_{r}$$ t r ”. $${y}^{(i)} (i=1,\dots ,N)$$ y ( i ) ( i = 1 , ⋯ , N ) : the therapy response of the $${i}^{th}$$ i th patient. $${\lambda }^{(k)} (k=1,\dots ,K)$$ λ ( k ) ( k = 1 , ⋯ , K ) : the $${k}^{th}$$ k th values of lambda in Elastic Net in Stability Selection. The $${\varvec{G}}$$ G was the gene pool having genes selected by step1 of CPMTPg.

Additional file 3: Figure S3. The pseudo-code of CPMTPg:step2. This code created a gene pool by the step1 of CPMTPg. $${{\varvec{X}}}^{({t}_{r})}=\left({{\varvec{x}}}_{1}^{\left({t}_{r}\right)},\dots ,{{\varvec{x}}}_{l}^{\left({t}_{r}\right)}\right); {{\varvec{x}}}_{j}^{\left({t}_{r}\right)}={\left({x}_{j}^{\left(1,{t}_{r}\right)},\dots ,{x}_{j}^{\left(N, {t}_{r}\right)}\right)}^{\top } (j=1,\dots ,p)$$ X ( t r ) = x 1 t r , ⋯ , x l t r ; x j t r = x j 1 , t r , ⋯ , x j N , t r ⊤ ( j = 1 , ⋯ , p ) : gene expression levels of “ $$p$$ p ” genes × “ $$N$$ N ” subjects at time point “ $${t}_{r}$$ t r ”. $${y}^{(i)} (i=1,\dots ,N)$$ y ( i ) ( i = 1 , ⋯ , N ) : the therapy response of the $${i}^{th}$$ i th patient. $${\lambda }^{(k)} (k=1,\dots ,K)$$ λ ( k ) ( k = 1 , ⋯ , K ) : the $${k}^{th}$$ k th values of lambda in Elastic Net in Stability Selection. The $${\varvec{G}}$$ G was the gene pool having genes selected by step1 of CPMTPg. The $${{\varvec{G}}}_{subset}$$ G subset was the gene subset of CPMTPp.

Additional file 4: Figure S4. Results of CPMTPp + CPMTPg using artificial data. The artificial gene expression data (1000 genes × 40 subjects × 5 time points; “#” in this figure means “number”) was created. This data subjects were 20 sensitive and 20 not sensitive responders. Gene expression levels of “Gene1”, “Gene2”, and “Gene3” were created by adding noise following a normal distribution (center:0; standard deviation:0.5) to each baseline. The baseline of “Gene1” had the different rising/ falling trends of gene expression levels between sensitive and not sensitive responders at all time points, while the baseline of “Gene2” and “Gene3” had it at a part of time points. Gene expression levels of the other genes were created by uniform distribution (maximum:1; minimum:5). To evaluate CPMTPp + CPMTPg, the threefold cross-validation was performed using this artificial data. As a result, CPMTPg selected “Gene1” from all genes as the gene subsets at all validation. These mean accuracies were 92.8%(CP1: $${t}_{0}\sim {t}_{1}$$ t 0 ∼ t 1 ), 97.6%(CP2: $${t}_{0}\sim {t}_{2}$$ t 0 ∼ t 2 ), 100%(CP3: $${t}_{0}\sim {t}_{3}$$ t 0 ∼ t 3 ), and 100%(CP4: $${t}_{0}\sim {t}_{4}$$ t 0 ∼ t 4 ), respectively. The accuracy at the early term was higher than 90%, and this value increased along with the time progressing. Similar trends were observed using actual datasets in this paper.