Predicting most labile CH bond using semiempirical methods

Yesterday, I mentioned on Twitter that I had written some prototype code to predict labile proton using PM3/COSMO.  A few people expressed interest so I put the code on GitHub.

The code is based on some previous work and removes all CH protons/hydrides/H atoms and finds the position with the lowest free energy.  I've just finished the code so I have no idea how well it works or if PM3/COSMO is the best choice.  Also, there are very few comments and some aspects of the setup (paths, etc) are particular to my machine and needs to be changed.

In its current form the method doesn't consider different reference molecules, so the prediction is "pure PM3/COSMO".  The basic code to do it is there but I haven't extended it yet. Also, the code only considers CH groups, so if have other groups (OH, NH, etc) with lower pKa they won't be considered.  This is very much a work in progress but if you have suggestions, test cases, ideas, or any other kind of feedback please let me know.  I hope you find it useful.


This work is licensed under a Creative Commons Attribution 4.0

Planned papers for 2017 - six months in

In January I wrote about the papers I plan to publish and made this list:

Accepted
1. Protein structure refinement using a quantum mechanics-based chemical shielding predictor
2. Prediction of pKa values for drug-like molecules using semiempirical quantum chemical methods

Probable
3. Intermolecular Interactions in the Condensed Phase: Evaluation of Semi-empirical Quantum Mechanical Methods
4. Fast Prediction of the Regioselectivity of Electrophilic Aromatic Substitution Reactions of Heteroaromatic Systems Using Semi-Empirical Quantum Chemical Methods
5. Benchmarking cost vs. accuracy for computation of NMR shielding constants by quantum mechanical methods
6. Improved prediction of chemical shifts using machine learning
7. PM6 for all elements in GAMESS, including PCM interface

Probably not in 2017
8. Protonator: an open source program for the rapid prediction of the dominant protonation states of organic molecules in aqueous solution
9. pKa prediction using semi-empirical methods: difficult cases
10. Prediction of C-H pKa values and homolytic bond strengths using semi-empirical methods
11. High throughput transition state determination using semi-empirical methods

The status is

Published
1. Protein structure refinement using a quantum mechanics-based chemical shielding predictor
2. Prediction of pKa values for drug-like molecules using semiempirical quantum chemical methods

3. Intermolecular Interactions in the Condensed Phase: Evaluation of Semi-empirical Quantum Mechanical Methods

Probable (at least submission)
4. Fast Prediction of the Regioselectivity of Electrophilic Aromatic Substitution Reactions of Heteroaromatic Systems Using Semi-Empirical Quantum Chemical Methods

We had a draft ready to be sent in in mid-April, but decided to include a "few" more types of heteroaromatics and the study has now ballooned to nearly 600 compounds (up from about 150 in the original draft)! The calculations and most of the analysis is done, but the paper basically has to be rewritten, and this really has to be done by my synthetic chemistry co-author, since the study is now much more about the chemistries of these heteroaromatic and much less about the method.  Even though its out of my hands I am still hopeful that we can submit it this year.

9. pKa prediction using semi-empirical methods: difficult cases automation

This paper is 2/3 written and presents a completely automated PM3-based pKa prediction protocol. The method works quite well, but most outliers turn out to be due to high energy conformations. The main remaining issue is to find a conformer-search protocol that consistently gives low-energy conformations. Depending on how much time I have to devote to paper 4 and the proposal mentioned below, I am still hopefull I can get this published this year.

7. PM6 for all elements in GAMESS, including PCM interface

This will basically be a paper on the accuracy of the SMD method using semiempirical methods.  I'm hopefull we will submit this year, but there is still a lot to do.  Among other things, it explains why PM3/COSMO works best for pKa predictions: the solvation energy errors happen to be smallest for this method.

10. Prediction of C-H pKa values and homolytic bond strengths using semi-empirical methods

I am planning to submit a proposal on prediction of pKa,  homolytic bond strengths, etc as predictors of CH activation-sites. The proposals is due September 25th, so much of my "spare" time until then will be spent on getting preliminary results and writing.

This work is licensed under a Creative Commons Attribution 4.0

Predicting pKa values using PM3 - conformer search using implicit solvation

Disclaimer: These are preliminary results and may contain errors

In a previous post I showed that minimising RDKit-generated conformers generated with MMFF led to slightly worse PM3/COSMO pKa predictions. One reason might be that the MMFF minimisation is done in the gas phase.  Anders Christensen reminded me that TINKER can do MMFF/GBSA minimisations so I used TINKER so minimise 20 and 50 RDKit-generated conformers in used these as initial guesses for the PM3/COSMO energy minimisations (20mmtk and 50mmtk

As you can see there is relatively little difference in the overall performance of Xmm and Xmmtk. The error distribution is a bit tighter for the tk variants and 50mmtk does better at finding structures closer to the "global" minimum, but 50nomm is still the best choice for avoiding >1 pKa errors.

So, either GBSA is not a good substitute for COSMO or MMFF is not a good substitute for PM3.  In light of this recent paper my money is on the latter.  So I'll go with 50nomm for now.

This work is licensed under a Creative Commons Attribution 4.0

Computing apparent pKa values using QM energies of isodesmic reactions

A few days ago I presented some preliminary pKa results.  Almost all the molecules in the set have more than one ionisable group. Here's how I compute the apparent pKa values

The microscopic pKa values are computed by
\begin{equation}
\label{eqn:pka}
\mathrm{pK_a}=\mathrm{pK_a^{ref}} + \frac{\Delta G^\circ}{RT\ln (10)}
\end{equation}
where $\Delta G^\circ$ denotes the change in standard free energy for the isodesmic reaction
\begin{equation}
\mathrm{ BH^+ + B_{ref} \rightleftharpoons B + B_{ref}H^+ }
\end{equation}
If there are more than one titrateable sites then several protonation states can contribute to the apparent pKa.  Here I follow the approach described by Bochevarov et al. If $\mathrm{D}$ and $\mathrm{P}$ differ by one proton and there are $M$ deprotonated sites and $N$ protonated sites then
\begin{equation}
\mathrm{K_{app}}  = \frac{([\mathrm{D}_1] + [\mathrm{D}_2] + ... [\mathrm{D}_M])[\mathrm{H}^+]}{[\mathrm{P}_1] + [\mathrm{P}_2] + ... [\mathrm{P}_N]}
\end{equation}
which can be rewritten in terms of microscopic pKa values
\begin{equation}
\mathrm{K_{app}}  = \sum^M_j \frac{1}{\sum^N_i 10^{pK_{ij}} }
\end{equation}
The sum contains contains microscopic pKa values for which more than one protonation state is changed. For example, molecule with three ionizable groups $(\mathrm{B_1H^+B_2H^+B_3H^+})$ will have the following microscopic pKa value
\begin{equation}
\label{eqn:pkapp}
K_{ij} = \frac{ [\mathrm{B_1B_2H^+B_3][H^+]}}{[\mathrm{B_1H^+B_2B_3H^+}]}
\end{equation}
in the expression for the apparent pKa value for going from the doubly to the singly protonated state. However, the error in such a pKa value is considerably higher due to less error cancellation and such pKa values are therefore neglected in the sum.

In the Bochevarov et al. implementation the user defines the titrateable group for which the apparent pKa is computed.  However, in my approach all possible protonation states so the assignment of the apparent pKa value to a particular ionizable group is not immediately obvious.  Inspection of Eq $\ref{eqn:pkapp}$ shows that the largest microscopic pKa values will dominate the sum over $N$, while the smallest of these maximum pKa values will dominate the sum over $M$. Thus, the apparent pKa is assigned to the functional group corresponding to the microscopic pKa
\begin{equation}
\label{eqn:max}
pK^\prime_{ij} = \min_{j}(\{ \max_{i}(\{ pK_{ij} \} ) \} )
\end{equation}
In some cases there are several microscopic pKa that contribute almost equally to the respective sums and for these cases the apparent pKa value cannot meaningfully be assigned to a single ionizable sites.  In my approach include all microscopic pKa values that are within 0.6 pKa units of the respective maximum and minimum values defined in Eq $\ref{eqn:max}$.   I choose 0.6 because that is the site-site interaction energy below which two groups titrate simultaneously.

You can find the code I wrote to do this here. It is not cleaned up yet.



This work is licensed under a Creative Commons Attribution 4.0

Small molecule energy minimisation and conformational search using Tinker

In this post I outlined the (possible) need for conformational analysis using a continuum model. Anders Christensen reminded me that Tinker can do this so I looked in to it. This blogpost and this GitHub repo were particularly helpful.  Here's what I found.

Generating atomic charges
Tinker has MMFF force field parameters except the atomic charges.  These can be generated from an sdf file using sdf2tinkerxyz (which relies on OpenBabel). When you download from SourceForge you get a bin file, which I didn't know what to do with so I downloaded the Linux executable from here instead.

./sdf2tinkerxyz -k default.key < xxx.sdf

creates yyy.xyz (coordinates) and yyy.key (charges), where yyy is the title in the sdf file (which I'll assume = xxx in the following). "default.key" is a file with Tinker keywords that is added to xxx.key. Here's mine

# Solvation Model
SOLVATE GBSA


#Force Field Parameters
PARAMETERS /opt/tinker/tinker/params/mmff.prm

Energy minimisation
To isolate the effect of continuum solvation on the pKa predictions, I want to generate conformers with RDKit and minimise them with MMFF/GBSA using Tinker.  This is done by

/opt/tinker/tinker/bin/minimize xxx.xyz -k xxx.key 0.01 > xxx.tout

"0.01" is the convergence criterium (in kcal/molÅ??). xxx.tout contains the energy information and a xxx.xyz_2 file is generated with the optimized coordinates.  This is in a format that OpenBabel can't handle, so I need to write a converter.  The main challenge is to translate the atom types into names. See the list called lookup in this code.

Conformational Search using SCAN
Tinker also has its own conformational search method. It looks like it's based on eigenvector following but I haven't looked closely at it yet.  This is done by

/opt/tinker/tinker/bin/scan xxx.xyz -k xxx.key 0 10 20 0.00001 > xxx.tout

Here I use the settings used in the DP4 GitHub repo.  "0" is automatic selection of torsional angles, "10" is the number of search directions (modes?), "20" is the energy threshold for local minima,and "0.00001" is the optimisation convergence criterium.  You get a bunch of coordinate files and xxx.tout holds the energy information.

I haven't really played with this option yet. If you have any tips, please leave comment.


This work is licensed under a Creative Commons Attribution 4.0

Predicting pKa values using PM3 - new code and conformer search

Disclaimer: These are preliminary results and may contain errors

The code I used in my previous pKa prediction paper had some limitations that I have now mostly removed: The setup of protonation states is automated and includes acids like acetic acid and phenol, there is a rudimentary tautomer generations, and the reference molecules are chosen a bit more systematically.

I ran the new code on the same set of molecules and saw a few differences.  Some were expected and due to different references and overall protonation states but most major differences were due to different conformations even though I used the same code to generate conformers. This let me to look at the effect of conformer generation.

The original study used 20 MMFF-minimised conformations ("20mm") generated using RDKit as starting points for PM3/COSMO minimisations.  Now I've tried skipping the MMFF minimisation ("nomm"), 50 conformations with and without MMFF minimisation, and a scheme where I MMFF-minimise 50 or 100 conformations and select conformers that have energies within either 10 or 20 kcal/mol of the lowest energy ("XecutY") for each protonation state/tautomer.

The RMSEs are very similar, but 50nomm has the fewest number of cases where the error in the pKa value (ΔpKa) is greater than 1 pH unit. However, the major outlier for 20mm and 20nomm is Sparteine for which 20mm and 20nomm actually find a conformer that is 4.9 kcal/mol lower than for 50nomm, for the protonated state.  

This lead me to look at the energies themselves. The numbers labelled ΔE in the table are computed as follows. For a given molecule I find the lowest energy for the protonated and deprotonated state for each method and identify the method with the lowest energy $E_{min}$. Then I count the number of molecules for which the difference to $E_{min}$ is greater than 1.36 kcal/mol and average that number for protonated and deprotonated to get one number per method.  

So there are 4 molecules for which 50nomm doesn't something close to the "global" minimum for either the protonated or deprotonated state, or both. In principle, I should go on and try 100nomm to see if I can "converge", but that's starting to become pretty expensive and the point is trying to develop a practically useful tool.

The 100ecut20 results suggest that the MM gas phase energetics don't represent the PM3/COSMO energetics all that well.  So it would be interesting to try a MM conformational search with an implicit solvent model. RDKit can't do this, but Macrmodel and NAMD should be able to do this.  It also looks like the next release of OpenMM will have this capability.  

Suggestions and comments most welcome.

This work is licensed under a Creative Commons Attribution 4.0

Finding the function that is its own derivative

I saw this "derivation" some years ago but now I can't find it.  If anyone knows the source please let me know.

Let's say we want to find a function $f(x)$ for which
$$ \frac{d}{dx} f(x) = f^\prime (x) = f(x) $$
Well, how about $f(x) = x$?
$$ f(x) = x \implies f^\prime (x) = 1$$
No, because $f^\prime (x) \ne x$. OK, how about $f(x) = 1+x$?
$$ f(x) = 1+x \implies f^\prime (x) = 1$$
That get's me a little closer.  I get the "$1$" back, but I need something whose derivative will give me back the $x$:
$$ f(x) = 1+x+\tfrac{1}{2}x^2 \implies f^\prime (x) = 1 + x$$
Now I need something whose derivative will me back the $\tfrac{1}{2}x^2$:
$$ f(x) = 1+x+\tfrac{1}{2}x^2 + \tfrac{1}{2 \cdot 3}x^3 \implies f^\prime (x) = 1 + x + \tfrac{1}{2}x^2$$
OK, so if I use infinitely many terms then this function fits the bill
$$ f(x) = 1+x+\tfrac{1}{2}x^2 + \tfrac{1}{2 \cdot 3}x^3 + \cdots \tfrac{1}{n!}x^n \cdots  \implies f^\prime (x) = 1 + x + \tfrac{1}{2}x^2 + \cdots \tfrac{1}{(n-1)!}x^{n-1} \cdots $$
Can I write $f(x)$ in some compact form?  Well, $f(0) = 1$ and $f(1) $ will give me some number, which I'll call $e$.
$$ f(1) = 1+1+\tfrac{1}{2} + \tfrac{1}{2 \cdot 3} + \cdots = e  $$
You only need a few terms before $e$ has converged to the first two decimal places (2.717).

For $x$ = 2 you need a few more terms to get the same precision, but the number (7.382) is roughly 2.717 x 2.717, an agreement that quickly gets better with a few more terms
$$ f(2) = 1+2+\tfrac{1}{2}4 + \tfrac{1}{2 \cdot 3}8 + \cdots = e^2  $$
So
$$ f(x) = e^x = 1+x+\tfrac{1}{2}x^2 + \tfrac{1}{2 \cdot 3}x^3 + \cdots \tfrac{1}{n!}x^n \cdots  $$
and
$$  \frac{d}{dx} e^x = e^x $$


This work is licensed under a Creative Commons Attribution 4.0