Thursday, November 24, 2016

Which method is more accurate? or Errors have error bars

2017.01.10 update: this blogpost is now available as a citeable preprint

This post is my attempt at distilling some of the information in two papers published by Anthony Nicholls (here and here). Anthony also very kindly provided some new equations, not found in the papers, in response to my questions.

Errors also have error bars
Say you have two methods, $A$ and $B$, for predicting some property and you want to determine which method is more accurate by computing the property using both methods for the same set of $N$ different molecules for which reference values are available. You evaluate the error (for example the RMSE) of each method relative to the reference values and compare. The point of this post is that these errors have uncertainties (error bars) that depend on the number of data points ($N$, more data less uncertainty) and you have to take these uncertainties into consideration when you compare errors. 

The most common error bars reflect 95% confidence and that's what I'll use here.  

The expression for the error bars assume a large $N$ where in practice "large" in this context means roughly 10 or more data points.  If you use fewer points or would like more accurate estimates please see the Nicholls papers for what to do.

Root-Mean-Square-Error (RMSE)
The error bars for the RMSE are asymmetric.  The lower and higher error bar on the RMSE for method $X$ $(RMSE_X)$ is
$$ L_X = RMSE_X - \sqrt {RMSE_X^2 - \frac{{1.96\sqrt 2 RMSE_X^2}}{{\sqrt {N - 1} }}} $$
$$ = RMSE_X \left( 1- \sqrt{ 1- \frac{1.96\sqrt{2}}{\sqrt{N-1}}}  \right) $$

$$ U_X =  RMSE_X \left(  \sqrt{ 1+ \frac{1.96\sqrt{2}}{\sqrt{N-1}}}-1  \right) $$

Mean Absolute Error (MAE)
The error bars for the MAE is also asymetric. The lower and higher error bar on the MAE for method $X$ $(MAE_X)$ is

$$ L_X =  MAE_X \left( 1- \sqrt{ 1- \frac{1.96\sqrt{2}}{\sqrt{N-1}}}  \right)  $$

$$ U_X =  MAE_X \left(  \sqrt{ 1+ \frac{1.96\sqrt{2}}{\sqrt{N-1}}}-1  \right)  $$

Mean Error (ME) 
The error bars for the mean error are symmetric and given by 
$$ L_X = U_X =  \frac{1.96 s_N}{\sqrt{N}} $$

where $s_N$ is the standard population deviation (e.g. STDEVP in Excel).

Pearson’s correlation coefficient, $\textbf{r}$
The first thing to check is whether your $r$ values themselves are statistically significant, i.e. $r_X > r_{significant}$ where

$$ r_{significant} = \frac{1.96}{\sqrt{N-2+1.96^2}}   $$

The error bars for the Pearson's $r$ value are asymmetric and given by 
$$ L_X = r_X - \frac{e^{2F_-}-1}{e^{2F_-}+1} $$
$$ U_X =  \frac{e^{2F_+}-1}{e^{2F_+}+1} - r_X $$


$$ F_{\pm} = \frac{1}{2} \ln \frac{1+r_X}{1-r_X} \pm r_{significant}$$

Comparing two methods
If $error_X$ is some measure of the error, RMSE, MAE, etc, and $error_A > error_B$ then the difference is statistically significant only if 

$$ error_A - error_B > \sqrt {L_A^2 + U_B^2 - 2{r_{AB}}{L_A}{U_B}} $$

where $r_{AB}$ is the Pearson's $r$ value of method $A$ compared to $B$, not to be confused with $r_A$ which compares $A$ to the reference value.  Conversely, if this condition is not satisfied then you cannot say that method $B$ is not more accurate than method $A$ with 95% confidence because the error bars are too large.

Note also that if there is a high degree of correlation between the predictions ($r_{AB} \approx $ 1) and the error bars are similar in size $L_A \approx U_B$ then even small differences in error could be significant.

Usually one can assume that $r_{AB} > 0$ so if $error_A - error_B > \sqrt {L_A^2 + U_B^2}$ or $error_A - error_B > L_A + U_B$ then the difference is statistically significant, but it is better to evaluate $r_{AB}$ to be sure.

The meaning of 95% confidence
Say you compute errors for some property for 50 molecules using method $A$ ($error_A$) and $B$ ($error_B$) and observe that Eq 11 is true.  

Assuming no prior knowledge on the performance of $A$ and $B$, if you repeat this process an additional 40 times using all new molecules each time then in 38 cases (38/40 = 0.95) the errors observed for method $A$ will likely be between $error_A - L_A$ and $error_A + U_A$ and similarly for method $B$. For one of the remaining two cases the error is expected to be larger than this range, while for the other remaining case it is expected to be smaller. Furthermore, in 39 of the 40 cases $error_A$ is likely larger than $error_B$, while $error_A$ is likely smaller than $error_B$ in the remaining case. 

This work is licensed under a Creative Commons Attribution 3.0 Unported License.

Sunday, November 6, 2016

Some useful PyMol commands

Here some PyMol commands I found useful while writing this paper.

Raytracing (pretty pictures)

select br. all within 3 of 63/CA
select br. all within 3 of resi 63
select Ala63, br. all within 3 of resi 63
select br. all within 3 of 2kpp///82/
sele tail,  2kzn///142-147/
sele tail,  2KPP///1-7+91-114/

NMR ensembles set all_states, on 
split_states your_object

Superimposing structures # superimpose protA residues 10-25 and 33-46 to protB residues 22-37 and 41-54:
pair_fit protA///10-25+33-46/CA, protB///22-37+41-54/CA

# superimpose ligA atoms C1, C2, and C4 to ligB atoms C8, C4, and C10, respectively:
pair_fit ligA////C1, ligB////C8, ligA////C2, ligB////C4, ligA////C3, ligB////C10

align cluster_lowe///13-25+36-105+111-141/CA,native///13-25+36-105+111-141/CA

Color using numbers in B-factor column
spectrum b, green_red, selection=n. CA,minimum=0.0, maximum=2
spectrum b, blue_white_red, selection=n. CA
spectrum b, blue_white_red, selection=n. CA,minimum=-1.37, maximum=1.37

spectrum b, blue_red, selection=test////CA
spectrum b, green_red, selection=n. CA,minimum=0.0, maximum=3.6

Label atoms 
label n. CA and i. 44, "(%s%s, %s)" % (resn, resi, b)
label n. CA and i. 33+55, "(%s%s, %s)" % (resn, resi, b)
label n. CA and i. 2, "%s%s, %5.2f" % (resn, resi, b)

Get more digits on the distance measurement
set label_distance_digits, 2

Get the orientation data, which you can paste back in to restore orientation

Change cartoon representation (the first two commands go together)
alter 16:23/, ss='L'

set ribbon_width, 8

create new_obj, cluster_lowe_all2
set cartoon_transparency, 0.5,new_obj

This work is licensed under a Creative Commons Attribution 3.0 Unported License.

Saturday, November 5, 2016

Semi-automatic pKa prediction using QM

Here I outline how I automated the calculations for this paper.  The files and programs can be found in the file that's part of the supplementary material on Figshare.

Generating SMILES
I used PDF to Text to get a text version of Table 3 from this paper and extracted the first column with the names using Excel.  I removed some molecules and the superscripts on some of the names by hand and created a text file with a single name on each line.

I used a python script to convert the names to SMILES strings

python table_3 > table_3.smiles

I used a python script to convert the SMILES string to 2D images

python tabel_3.smiles

I used the Cover Flow option in Finder (Mac) to browse through the images looking for errors. Turned out that several SMILES strings included the HCl which I removed.  If I saw other tautomer possibilities I created those by hand (I need to automate this).  A handful of molecules had carboxyl group which I deprotonated by hand by changing the SMILES string (I need to automate this).

I used a python script to protonate the amines

python table_3.smiles > table_3+.smiles

This program creates all possible single protonation states of all nitrogen atoms (except amides) in the molecule.  I deleted some of the protonation states I didn't want.  For example, for histamine I want to compute the pKa of the amine group but not of the imidazole, so I deleted the line with the histamine SMILES string for the protonated imidazole.  For a few of the molecules I also needed SMILES for doubly protontaes molecules to I used and table_3+.smiles to create a  table_3++.smiles file also and extracted the molecules I needed.  These steps where repeated for the reference molecules as well.

File naming
The each line in the .smiles files is the name of the molecule followed by a SMILES strings, the name is used to construct the filename, and the subsequent workflow makes some assumptions about the filenames.  When I deprotonoate carboxyl group I add a "-" to the filename, e.g. "Phenylalanine-", in the table_3.smmiles and protoam adds a "+" to the filename, e.g.  "Phenylalanine-+".  Molecules in the table_3++.smiles file will have names like.e.g Procaine++. If more than one nitrogen is protonated adds "_1" and "_2", etc., e.g. "Histamine+_1".  When I make tautomers I add "_1" or "_A", etc., e.g. "Cimetidine+_A".  (So in principle you could have, e.g. "name_A_1", but that didn't happen in practise).  Some manual editing is required for everything to come out right.

Creating input files from SMILES
I use a python script to create sdf files from the SMILES strings, e.g.

python table_3.smiles

The script creates 20 different conformations for each SMILES string. The sdf files are named xxx_y.sdf where xxx is the names in the .smiles file and y is an integer between 0 and 19.

I use OpenBabel to convert the .sdf files to input files.  For example for MOPAC

for x in *.sdf; do babel \$x \${x%.*}.mop -xf ../pm6-dh+.header; done

The header file contains the keyword and the charge is specfied as "charge=0". So I need to change the charge by,

sed -i 's/charge=0/charge=1/g' *+_*.mop
sed -i 's/charge=1/charge=2/g' *++*.mop
sed -i 's/charge=1/charge=0/g' *-+*.mop
sed -i 's/charge=0/charge=-1/g' *-_*.mop

I change the method by, for example,

sed -i 's/pm6-dh+/am1/g' *.mop

Computing the pKa values
Af all the jobs have run I extract energies from the output files. For MOPAC out files by

grep "FINAL HEAT" *.out > xx.energies
grep "CURRENT" *.out >> xx.energies

The last line is needed in case MOPAC doesn't converge, in which case I just extract the last energy.

For  GAMESS log files
grep "TOTAL FREE ENERGY IN SOLVENT" *.log > xx.energies

I use a phython script to compute the pKa values from the .energy files

MOPAC pKa values
python xx.energies big_table_3.smiles > xx.pka

GAMESS pKa values
python xx.energies big_table_3.smiles > xx.pka

Here big_table_3.smiles all the .smiles files, including reference molecules, combined into one.

The python script finds the most lowest free energy for each protonation state and the appropriate reference molecule for each ionizable site.  The protonation state is defined by the "-" and "+"s so that the lowest free energy for protonated histamine is the lowest free energy found in the output files named "Histamine+_*".  The pKa is related to the free energy difference between e.g. "Histamine+" and "Histamine" or "Phenylalanine-+" and "Phenylalanine-" or  "Procaine++" and "Procaine+", i.e.

delta_G = energies[name+"+"] - energies[name]

The python script contains SMILES strings for all the reference molecules and the appropriate reference for a titration is the reference molecule with the largest substructure match around the titrating nitrogen.

Some future improvements
I need to automate the tautomer generation and I shouldn't hand-pick the protonation state.  For example, I should consider both possible single protonation states in histamine and have the program automatically use the lowest free energy.  This also means that for, e.g. phenylalanine I should consider both the neutral and zwitterionic protonation state, i.e. "Phenylalanine-+_A" and "Phenylalanine-+_B" and  pick the lowest energy.

This work is licensed under a Creative Commons Attribution 3.0 Unported License.

Generating coordinates for molecules with diverse set of elements

It looks like PM6 in GAMESS is working, but we need to test it for more elements.  Here's how I generated the MOPAC input files.  First I made a list of SMILES strings for for most of the elements, which I then converted to sdf files with  The I used OpenBabel to convert to MOPAC input files.

The implementation is only done for RHF so the molecules need to be closed shell.  So, for example for Sc I used ScCl$_3$ For many of transition metals I couldn't really think of a closed shell molecule, so I didn't include those elements.  For the non-metals the Cactus server automatically adds hydrogen.  Cactus also interprets names like Sc as sulphur, so square brackets are needed.

This work is licensed under a Creative Commons Attribution 3.0 Unported License.

writing unicode csv files from Excel for Mac

update: ... or you can update Excel to the latest version (15.27) and then you get the option to save as CSV UTF-8 (thanks fxcoudert #twitterrules)

Special characters become corrupted when saving Excel files in the csv format.  There is apparently no easy fix for Mac, so here is what I ended up doing.

Save Excel file as UFT-16 Unicode Text.

Open the file in TextEdit.  Edit > Find > Find and Repace (OPTION-CMD-F)
In the search panel press OPTION-TAB and in the replace window type "," and press "all"
Save the extension to csv when you save the file.

You can also do it in vi
 ”:1,$s/(Press CTRL+v) then (Press TAB)/,/g”

This work is licensed under a Creative Commons Attribution 3.0 Unported License.