6 1 COMPOSICIÓN DEL LÉXICO CHAMORRO
6.3. ADAPTACIÓN DE ANGLICISMOS
The best model with transporter-expression correction was generated from applying feature selection to the full set of features (M5), contrarily to the best model with no expression correction, which was trained with feature selection applied to both physiological descriptors and molecular descriptors, separately (M4). Despite being subject to feature selection alongside a much larger number of molecular features (304 molecular descriptors and 11 physioloical descriptors), some of the physiological descriptors were selected in the final subset (pBCRP1_c, pOATP1B1_c and pMRP2_c). Even though a relatively small percentage of compounds was affected by any of these three transporters, with pBCRP1_c showing the largest descriptor importance (13.9%), looking into the composition of the random forest model actually shows that both pMRP2_c and pOATP1B1_c were found at the top node in 16 and 4 trees, respectively (out of a total of 600 trees). Finding a descriptor close to (or at) the top node is a strong indicative of the meaningful role of such descriptor towards the modelled output variable (Freitas, 2013).
Table 7.2. Summary of predictive performance measured on the test set (N=134) for the best model
in this work (M5) and the best model from Chapter 6 (8a).
Best Models
conditions Feature content
R2 RMSE MAE GMFE MFE
Within 2-FE (%) Within 3-FE (%) 8a RF-GA FS-MDs + FS-PDs 0.560 0.4497 0.3391 2.18 2.99 56.0 73.1
M5 RF-GA FS-All features 0.582 0.4353 0.3237 2.11 2.84 59.7 72.4
Table 7.2 shows that using protein expression-corrected transporter data led to an improvement in the ability to predict Vss, showing an MAE of 0.3237 vs 0.3391 when no expression correction was applied (model 8a). Beyond MAE, all performance measures, except for % predictions within 3-FE, were better for the current chapter’s best model, M5. Additionally, 58.2% of the prediction errors were reduced when compared to the best model trained in the absence of expression correction. As seen in Figure 7.3, these 58.2% (shown as the filled circles) include the instances associated with the largest prediction errors. In fact, almost all compounds in the right-hand side of the plot, which fall outside the 3-FE threshold (indicated by the dashed lines), are predicted with higher accuracy when using transporter-expression correction (when compared to model 8a). The improved predictions include four problematic compounds pointed out in previous works as well as in Section 6.3 as being particularly challenging to predict due to extensive binding to different tissue structures(WHO, 2013, Watts and Diab, 2010, Barbour et al., 2009, Zheng et al., 2011) (pentamidine, chloroquine, risedronic acid and tigecycline).
As the relationship between chemistry and Vss has already been extensively discussed in the literature, when discussing the chemical descriptors of the best model (M5) the focus will be placed on discussing the presence of physiological descriptors in the model. However, it should be pointed out that, besides the high importance in model M5 of expected descriptors such as ionized fraction and lipophilicity (which occupied the top
positions in the random forest’s trees in Chapter 6), a new descriptor implemented in MOE
(h_pavgQ) has been found to have the highest feature importance in model M5. This is a descriptor calculated using Extended Hückel Theory, which is a semi-empirical quantum mechanics method that takes into account local resonance and electron withdrawing effects (Labute et al., 2014). In particular, h_pavgQ is the average total (formal) charge, a pH dependent parameter calculated based on the relative concentration of various protonation states of the molecule. This parameter conveys similar information to fractions of anionic, cationic, zwitterionic and unionized forms of a molecule at different pH values, which are calculated from the acidic and basic pKa values (Ghafourian et al., 2006). These fractions
were found to be major predictors of VD as distribution of compounds may be limited for acidic compounds such as nonsteroidal anti-inflammatory drugs with strong plasma protein binding, whereas basic compounds may be able to accumulate in the phospholipid membranes (Ghafourian et al., 2006, Freitas et al., 2015).
Figure 7.3. Scatter plot of test set predictions obtained by the best model with transporter-correction
expression (model M5). Filled circles indicate predictions which show a smaller error compared to the best model with no transporter-expression correction in Chapter 6 (model 8a). Compare to Figure 6.4 to see the improvement for the outlier predictions.
In the previous attempt to model this same dataset without transport-expression correction in Chapter 6, even though exactly the same training set was used as well as the same feature selection and regression algorithms, there was a marked difference in the physiological descriptors that were selected into the model. In Chapter 6 and pPEPT1, pPL, pMRP1, pMRP2, pBCRP1 and pOATP1B1 were selected and used in the best achieved model (see Appendix IV Table A4.1 for full list of descriptors of model 8a), while in the current work the first three of these descriptors were not selected in the best model (M5). This difference is especially significant, as one of them (pPL) encodes information on experimental and predicted phospholipidosis (completed missing experimental data), and its absence associated with improved performance goes against the observations in Chapter 6, where pPL consistently improved the majority of the models produced. However, while the selection of a descriptor is evidence of its informative value, the failure to select a descriptor does not imply a lack of informative value. It might be the case that pPL was not selected due to the selection of other descriptors which were correlated with pPL or even
more informative than this descriptor (making the selection of pPL unnecessary), as M5 (contrarily to 8a) resulted from feature selection applied to the full set of features (making pPL redundant).
Table 7.3. Full list of descriptors used in the best model (M5), and their relative importance (in
parenthesis), calculated as the percentage of correctly predicted training compounds, over the total number of training compounds that go through a decision node containing each of the descriptors in the model. Descriptors in model M5 h_pavgQ (42.2) LogD(5_5) (16.5) FiA (35.2) chi1 (16.4) LogD(10) (35.1) vsurf_W7 (16) vsurf_HL1 (26.4) dipole (15.7) FiB (25.3) vsa_acc (15.1) ASA_P (24.9) vsurf_Wp5 (14.3) vsa_pol (23.6) vsurf_IW5 (14) vsurf_HB2 (23) pBCRP1_c (13.9) LogP (21.7) vsurf_ID8 (13.7) FASA_P (21.6) VAdjMa (13.2) PEOE_VSA_FPPOS (21.1) AM1_dipole (13.2) PM3_HOMO (20.5) vsurf_DD13 (12.3) vsurf_HB1 (20.4) pMRP2_c (12.2) PEOE_VSA_FPOL (20.2) vsurf_Wp6 (11.9) Q_VSA_FPOL (20.1) PEOE_VSA+5 (11.4) a_ICM (20) a_nO (11.2) SMR_VSA0 (19) vsurf_DD23 (10.6) AM1_HOMO (18.6) a_don (10.2) SlogP_VSA1 (18.6) Num_Rings (9) vsurf_HB3 (18.4) PEOE_VSA-2 (8) Q_RPC- (18) Halogen_ratio (6.5) PEOE_VSA-0 (17.6) FiAB (5.4) vsurf_ID2 (17.5) SlogP_VSA6 (5) Q_VSA_FPNEG (17.5) chiral_u (4.8) PEOE_VSA_PPOS (17.3) pOATP1B1_c (3.6) Surface_Tension (17.2) Num_Rings_4 (3.6) LogD(6_5) (16.8) Rule_Of_5 (3.4) density (16.6) b_triple (1.1) Kier3 (16.6)
Lastly, similarly to what was done in Chapter 6, in order to challenge the contribution of the three physiological descriptors in the model, the model was retrained under the same conditions as the best model, with the only change being the removal of physiological features. As the best model consists of a very large ensemble (a random forest of 600 trees), it is possible that the presence of the physiological features is not advantageous and merely results from the combination of chance and the fact that the random forest is built with unpruned trees. Despite this possibility, removing the PDs yielded degradation of the
model’s performance (validation set MAE = 0.3003 vs 0.3044, with and without PDs, respectively). Curiously, removing three molecular descriptors of the same level of
importance as these three physiological features improved the model (validation set MAE = 0.3003 vs 0.2951, for the original and removed-descriptors model, respectively).