Virtual Ligand Strategy in Transition Metal Catalysis Toward Highly Efficient Elucidation of Reaction Mechanisms and Computational Catalyst Design

Wataru Matsuoka, Yu Harabuchi, and Satoshi Maeda (2023)
Highlighted by Jan Jensen

This perspective shows how an old computational tool can be adapted to serve a new purpose. When I started in compchem changing, say, a few F atoms to and H atoms in a molecule often made the difference between waiting a few days and a few weeks for the calculations to finish. People therefore developed pseudo H atoms that could mimic the electronic effect of larger atoms or even entire functional groups. Some of these methods were later adapted to serve as boundary atoms in QM/MM calculations and now they have found a new use in screening for ligands in organometallic catalysts.

The use of pseudoatoms to model such ligands not only speeds up the individual calculations but also maps the chemical space on to just two dimensions, electronic and steric, that allows the space to be searched more efficiently. Once the desired combination of electronics and sterics is found corresponding real ligands are found by another, much faster, screen if commercially available or synthetically accessible ligands.

The authors use this approach to identify two phosphine ligands for a chemoselective Suzuki–Miyaura cross-coupling catalyst, complete with experimental verification.

The downside is that the parameterisation of these "virtual ligands" are a bit involved and very ligand-dependent. But an interesting approach non-the-less.

This work is licensed under a Creative Commons Attribution 4.0 International License.

Ultra-large library docking for discovering new chemotypes

Jiankun Lyu, Sheng Wang, Trent E. Balius, Isha Singh, Anat Levit, Yurii S. Moroz, Matthew J. O’Meara, Tao Che, Enkhjargal Algaa, Kateryna Tolmachova, Andrey A. Tolmachev, Brian K. Shoichet, Bryan L. Roth & John J. Irwin (2019)

Highlighted by Jan Jensen

Figure 3a from the paper. (c) Nature

This paper has already been thoroughly highlighed several places, such as here and here, so I'll just summarise what the main take-home messages are for me.

• The size of the libraries (99 and 138 million) that are screened are truly impressive, especially when you realise that they sampled 280 conformations for each molecule! This required 1.2 calendar days on 1,500 cores.
• The libraries where made from 70,000 commercially available building blocks, which where combined using 130 known reactions. The molecules in the library should therefore be easy to synthesise
• Indeed, for one target they selected 589 molecules for synthesis and successfully made 549, for which they measured affinities.
• The selected molecules spanned the whole range of docking score, which results in a thorough test of the accuracy. As shown in the figure above, the scores can only really be used to weed out the very weak binders.
• As Derek Lowe notes "That definitely argues for setting up these virtual libraries according to expected ease of synthesis, because otherwise you could spend a lot of time making tough compounds that don’t do anything. People have."

Very commendably, the authors have made the libraries available as a public database.



This work is licensed under a Creative Commons Attribution 4.0 International License.

Hunting for organic molecules with artificial intelligence: Molecules optimized for desired excitation energies

Masato Sumita, Xiufeng Yang, Shinsuke Ishihara, Ryo Tamura, and Koji Tsuda (2018)

Highlighted by Jan Jensen

Figure 1 from the paper. Reproduced under the CC-BY-NC-ND license

Sumita and co-workers combine Monte Carlo tree search (MCTS) and a recurrent neural network (RNN) to discover molecules with specific excitation levels.  The general approach is very similar to the one used by Segler, Waller, and co-workers to predict retrosynthetic pathways, that I highlighted last month. 

At the core of the method (called ChemTS) is a RNN trained to generate SMILES string representations of molecules - another approach pioneered by Segler and Waller. Trained on thousands of valid SMILES strings, the RNN predicts that, for example, a likely next character in the SMILES string "c1ccccc" is "1" (to form benzene), just like an RNN trained on thousands of English words would predict that a likely next character in "chemistr" is "y".

Since there is more than one probable choice for each new character the number of possible SMILES strings quickly become unmanageable: even five possible characters for each position in a 20-character SMILES string results in $10^{14}$ possibilities. This is where MCTS is helpful (paraphrased from my previous highlight):

A MCTS starts by evaluating a number of possible SMILES strings randomly and then assigning likelihood scores to the early parts of the string depending on whether the encoded molecule has a desired property or not. The process is then repeated except that the early parts of the SMILES string is chosen based on likelihood scores, which are continuously updated and added to unscored characters. The changing likelihood scores means that the search for new SMILES strings is directed towards the more promising areas of the tree. I have given a short illustration of the process here. The process is repeated for a given number of steps and the SMILES strings with properties closest to the target are selected.

The desired property is a certain value of the molecules lowest excitation level (200, 300, 400, 500, or 600 nm), which is predicted using TDFT at the B3LYP/3-21G* level of theory.  For example, given two days of CPU time on 12 cores, ChemTS generated 646 possible molecules of which 34 has a predicted excitation energy within 20 nm of 200 nm. Two of these molecules where tested experimentally and one molecule did indeed have an excitation energy in the desired range.

This work is licensed under a Creative Commons Attribution 4.0 International License.

Efficient prediction of reaction paths through molecular graph and reaction network analysis

Yeonjoon Kim, Jin Woo Kim, Zeehyo Kim and Woo Youn Kim (2017)

Highlighted by Jan Jensen

Figure 2 from the paper. Reproduced under the CC-BY licence.

The paper presents an interesting approach to generating the most important intermediates given reactants and products. 

1. The reactants and products are mapped to atom connectivity (AC) matrices, where the presence and absence of bonds are represented by 1's and 0's. 

2. Intermediates are then enumerated by changing the matrix elements using simple rules and constraints are employed to keep the number of intermediates manageable: the change in bonding is limited to a maximum of two per steps and only changes to bonds involving "active atoms" (atoms with different bonding in reactants and products) are considered. Identifying active atoms involves mapping reactant atoms to product atoms, using the method proposed by Floudas and co-workers.  The AC matrices are converted to SMILES strings and 3D coordinates to ensure that they correspond to chemically reasonable molecules.

3. A reaction network is then constructed by finding intermediates that best connect reactants and products by computing "chemical distances (CDs)" between AC matrices and identifying intermediates with CDs to reactant and product that are similar to the CD between the reactant and product themselves. The CDs are then used to find the shortest path to reactant and product for each intermediate to yields a minimal reaction network.

4. Finally, TSs connecting the structures in the minimal reaction are found using conventional QM methods.

The method is applied to Claisen ester condensation and Cobalt-catalyzed hydroformylation. The method is implemented in a python program called ACE-Reaction but the availability of this program is unclear.  

This work is licensed under a Creative Commons Attribution 4.0 International License.

Understanding and Breaking Scaling Relations in Single-Site Catalysis: Methane-to-methanol Conversion by Fe(IV)=O

Terry Z. H. Gani and Heather J. Kulik (2017)

Highlighted by Jan Jensen

This is the first study I have come across that locates TS structures as part of a "high-throughput" single-site catalyst design study. Furthermore, the catalyst contains iron, which is not the easiest of elements to work with computationally. 

The study locates 76 and 43 TSs for the oxo formation (TS1) and hydrogen atom transfer (HAT, TS2) steps of the catalytic cycle. These are relatively small numbers compared to high throughput studies of other properties (hence the quotation marks), but they are roughly an order of magnitude larger than the number of TSs found in typical computational study of catalysts. The number is smaller for HAT due to difficulties in locating TSs for this step.

The TSs were located using either NEB implemented in DL-FIND or Q-CHEM where initial guess structures were generated using a locally modified version of molSimplify.

The studies show that there is a good correlation between reaction energy and barrier for the HAT step (R2 = 0.99) but a poor correlation for the oxo formation (R2 = 0.50 - 0.81). The authors conclude "Overall, our work shows that LFERs can be leveraged in single-site catalyst screening only when the coordination geometry is held fixed. Reliance solely on LFERs for single-site catalysis will thus miss rich areas of chemical space accessible through scaffold distortion."

This work is licensed under a Creative Commons Attribution 4.0 International License.

An automated transition state search and its application to diverse types of organic reactions

Leif D. Jacobson, Art D. Bochevarov, Mark A. Watson, Thomas F. Hughes, David Rinaldo, Stephan Ehrlich, Thomas B. Steinbrecher, S. Vaitheeswaran, Dean M. Philipp, Mathew D. Halls, and Richard A. Friesner (2017)

Highlighted by Jan Jensen

Copyright 2017 American Chemical Society

Finding transition states remains one of the most labor intensive pursuits in computational chemistry.  While interpolation methods are becoming increasingly robust, they usually require that the atom order for reactant and product are identical (atom mapping) and can be sensitive the starting conformations and relative orientation in case of bi-molecular reactions.  Furthermore, one still has to check whether the right TS is found and formulate a strategy if it is not.  All these things to do not immediately lend themselves to automation but this paper proposes solutions for all these problems.

In particular the paper offers a very elegant solution for the atom mapping problem: bonds are broken in both reactants and products until the connectivity of the fragments are identical after which the atoms in the fragments can be easily matched. Both the comparison and atom mapping of fragments can be easily done with modern cheminformatics toolkits such as RDKit using canonical smiles and  maximum common substructure searchers (after atom order and charge has been removed).  Cases where this fails due to equivalent atoms (e.g. the hydrogens in a methylene group) can then be dealt with by searching for the solution with the lowest RSMD between reactant and product.

The study focussed on relatively small and rigid molecules and issues due to multiple conformations is left for a future publication.

This work is licensed under a Creative Commons Attribution 4.0 International License.

Automatic chemical design using a data-driven continuous representation of molecules

Rafael Gómez-Bombarelli, David Duvenaud, José Miguel Hernández-Lobato, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, Ryan P. Adams, and Alán Aspuru-Guzik (2016)

Contributed by Jan Jensen

Chemical space is discrete which makes it hard to search with standard techniques such as gradient-based minimisation.  This paper used a standard machine learning tool called an autoencoder to help solve that problem.  One way to think of an autoencoder is as a data-compressor where one neural network is trained to describe a data set such as an image in some compressed representation and another network is trained to recover the image from the compressed format.

The interesting thing in the context of chemical space is that the compressed format can be a continuous function such as a real-valued vector (latent space). (Another use of autoencoders is dimensionality reduction for data visualization, e.g. as an alternative to principal component analysis.)  This latent space is therefore a continuous representation of the chemical space (a set of SMILES strings) that the autoencoder was trained on.  Another neural net can then be trained to map some chemical property, such as logP values, on this latent space and the space can be searched for regions with desired logP values with techniques as simple as interpolation.

One problem with autoencoders is that they are "lossy" which in this case translates to the fact that not all points in latent space can be decoded to a valid molecule (SMILES string) but the failure rate is relatively low for the two proof-of-concept applications in the paper.

This is a very interesting new tool in the hunt for molecules with new properties.


This work is licensed under a Creative Commons Attribution 4.0

Diverse Optimal Molecular Libraries for Organic Light-Emitting Diodes

Chetan Rupakheti, Rachael Al-Saadon, Yuqi Zhang, Aaron M. Virshup, Peng Zhang, Weitao Yang, and David N. Beratan (2016)
Contributed by Jan Jensen

This paper uses the property-optimizing ACSESS (PO-ACSESS) method to create a set of diverse set of candidates for organic light-emitting diodes (OLEDs) with efficient blue emissions.  There are over 10$^{60}$ synthetically feasible, low molecular eight organic molecules so it is impossible to do an exhaustive search of this so-called small molecule universe looking for better blue OLEDs. Yang and Beratan have therefore developed PO-ACSESS, a stochastic genetic algorithm to extract a set of diverse molecules with a given set of properties.

PO-ACSESS starts with a set of seed molecules (in this case organic molecules that are known blue emitters) that are then randomly mutated and mated. A mutation may be any of a set of elementary steps such a change in bond order, ring formation, atom addition/deletion, etc. and mating is done by randomly fragmenting the molecules at a rotateable bond followed by re-combination of fragments from different parents. The new molecules that do not meet the desired stability, synthetic-feasibility, and electronic criteria are removed. To enforce diversity, molecules that are too close in chemical space are also removed and the entire process is repeated starting with these molecules. The chemical space distance between two compounds is defined as the distance between compounds based on their descriptors, such as atomic number, Gasteiger−Marsili partial charge, atomic polarizability, and topological steric index.

In the case of blue OLEDs the electronic criteria are a vertical singlet excitation energy in the 2.4−4.1 eV range (blue region), an oscillator strength of ≥0.4, and a singlet-triplet gap ≤ 0.3 eV. “Initially, the threshold is set to a less stringent value … to ensure that the population does not collapse to zero, because of the fitness constraint.” The electronic properties were calculated using semiempirical methods such as AM1, ZINDO/S, and DFTB.

The design process was done in two stages: First PO-ACSESS was used to identify 195 molecules with the desired vertical singlet excitation energies and oscillator strengths.  I could not find any details on what seed molecules were used nor the number of iterations. These 195 molecules where then used as seeds for further optimization of the singlet-triplet gap. This PO-ACSESS search was run for 50 iterations where ∼90 molecules were generated per iteration that satisfied the required singlet-triplet gap constraint, which was completed in 7 days on a 16-core CPU. I’d be very curious to know how many molecule were actually screened as part of the process. Anyway, ∼60 structures with a computed singlet-triplet gap-value of ∼0.3 eV that are predicted to emit in the blue region and to have reasonably high oscillator strengths. 

This work is licensed under a Creative Commons Attribution 4.0

Navigating molecular space for reaction mechanisms: an efficient, automated procedure

Paul M. Zimmerman (2015)
Contributed by Jan Jensen

This methodology takes a single structure as input automatically finds related intermediates and transition states connecting them to map out entire reaction mechanisms.  The methodology is the combination of two different methods.

A single elementary step search "that takes a single chemical structure and proposes new intermediates that could be one elementary step away". Here an "elementary step" is defined as "chemical rearrangements with no more than two connections breaking and two connections forming simultaneously, while maintaining the upper and lower limits for coordination number of each atom." The energy of these intermediates are then computed using DFT and the ones with sufficiently low energy are kept.

The Growing String Method is then used to find the transition state and reaction path connecting the intermediates to the input structure and reaction paths with sufficiently low barriers are kept.  This method, as implemented in Q-Chem, is completely automated and appears to be very robust.

The procedure is then repeated using the remaining intermediates as input until no new intermediates are formed.  The thermodynamic and kinetic screening helps insure that the number of intermediates do no grow out of control.  Also, for larger systems one can define inactive atoms (e.g. non-reactive ligands) that are not considered when creating new intermediates.


This work is licensed under a Creative Commons Attribution 4.0

Big Data Meets Quantum Chemistry Approximations: The ∆-Machine Learning Approach

Raghunathan Ramakrishnan, Pavlo O. Dral, Matthias Rupp, and O. Anatole von Lilienfeld Journal of Chemical Theory and Computation 2015, ASAP (arXiv)

Contributed by +Jan Jensen 

Figure 1. Two hypothetical property profiles connecting two constitutional isomers of C$_7$H_$_{10}$O$_2$. The Δ-model, estimates the difference between baseline and target line properties (arrow) which differ in level of theory (b → t), geometry ($R_b$ → $R_t$), and property ($E_b$ → $H_t$). Reprinted with permission from J. Chem. Theory Comput. 2015, ASAP. Copyright (2015) American Chemical Society.

The idea behind this method is best explained by a specific example.  The G4MP2 enthalpies [$H_t(R_t)$] of  C$_7$H_$_{10}$O$_2$ isomers are estimated using PM7 electronic energies [$E_t(R_b)$] by 

$$H_t(R_t) \approx \Delta_b^t(R_b) = E_b(R_b)+ \sum_{i=1}^N\alpha_i e^{|R_i-R_b|/\sigma}$$

Here {$\alpha_i$} and $\sigma$ are parameters found by regression using a training set of $N$ molecules and $|R_i-R_b|$ is a measure of similarity between the target molecule and training molecule $i$.  The latter is described in more detail here, but I found it pretty interesting so I am summarizing it here.

A Coulomb matrix ($\mathbf{C}$) is constructed for each molecule
$$
C_{kl}= \begin{cases}
 0.5 Z_k^{2.4} & \text{if }  i=j\\
 Z_kZ_l/r_{kl}& \text{if } i \ne j
\end{cases}
$$
where $r_{kl}$ is the distance between atom $k$ and $l$ and $Z_k$ is the nuclear charge of atom $k$. Then the elements are sorted such that the diagonal elements are in descending order and the similarity is computed by
$$|R_i-R_b| = \sum_{k,l} |C_{kl}^i - C_{kl}^b | $$
Using this approach and a training set of ($N$ =) 1000 molecules the G4MP2 atomization enthalpies of 6095 constitutional isomers of C$_7$H_$_{10}$O$_2$ can be reproduced with a MAE of 3.9 kcal/mol using PM7, compared to an MAE of 6.4 kcal/mol for uncorrected PM7.  Using PBE or B3LYP/6-31G(2df,p) the MAE can be brought below 1 kcal/mol using a 1K training set.

In another interesting application the MAE of RHF/6-31G(d) relative to CCSD(T)/6-31G(d) atomization energies for the same set of molecules can be reduced from 3 to less than 1 kcal/mol using a 1K training set.

This is thus a very interesting approach for obtaining chemical accuracy using methods that are sufficiently fast to study thousands of molecules. The caveat is that about 1000 high level calculations appears to be needed to train the method but perhaps more generally applicable parameter sets can be found using, for example, functional group identification.


This work is licensed under a Creative Commons Attribution 4.0  

Large-scale virtual high-throughput screening for the identification of new battery electrolyte solvents: computing infrastructure and collective properties

Tamara Husch, Nusret Duygu Yilmazer, Andrea Balducci and Martin Korth, Phys. Chem. Chem. Phys., 2015,17, 3394-3401
Contributed by Tobias Schwabe

If you are among those who (like me) follow from time to time what is going on in volunteer computing for Computational Chemistry, you might be well aware of the first projects in this field: QMC@home (see: www.qmcathome.org). It's been a little bit quiet around there lately but now it seems Martin Korth and his team are setting the stage for a new project for the QMC@home community: cleanmobility.now. The foundations for that are laid out in their recent PCCP paper.

Cleanmobility.now aims at supporting the development of better batteries for electric cars. As a first starting point, finding new electrolytes has been chosen as target – certainly a good choice as this idea is quite in vogue right now: e.g. see this JPC/C Feature Article (which, by the way, also shows the usage of the very interesting projected WFT-in-DFT embedding method to get accurate results for large complex systems)[1]. And for a more general overview, you can check some recent reviews about Computational Chemistry in this field [2,3].

Husch et al. make an interesting contribution here because they attack the problem by virtual HTS, one of the few studies where this idea is not “just” applied for drug discovery. Especially, they tackle the problem of collective properties. Their pilot studies show already some promising results. For example, they found nitriles to be potential electrolytes, which have also attracted some interest from experimental side.

References

[1] Taylor A. Barnes, Jakub W. Kaminski, Oleg Borodin, and Thomas F. Miller, III, J. Phys. Chem. C 2015, 119, 3865−3880, DOI: 10.1021/jp510882g

[2] Mahesh Datt Bhatt and Colm O’Dwyer, Phys. Chem. Chem. Phys., 2015, 17, 4799—4844, DOI: 10.1039/c4cp05552g

[3] Martin Korth, in Specialist Periodical Reports: Chemical Modeling: Applications and Theory, ed. M. Springborg and J.-O. Joswig, Royal Society of Chemistry, London, UK, 2014

Materials Cartography: Representing and Mining Materials Space Using Structural and Electronic Fingerprints

Olexandr Isayev, Denis Fourches, Eugene N. Muratov, Corey Oses, Kevin Rasch, Alexander Tropsha, and Stefano Curtarolo.  Chem. Mater.ASAP. (2015) DOI: 10.1021/cm503507h

Contributed by Alán Aspuru-Guzik

What brings a chemist to post about a joint paper of a professor from a school of pharmacy and one from a department of materials science? A great multidisciplinary paper!

This paper is a tour de force for computational materials science. The authors, from the Tropsha and Curtarolo groups have applied state-of-the-art tools from chemoinformatics and machine learning to the challenging problem of materials design. In this paper, led by Olexandr Isayev, the authors employ chemical descriptors based on electronic properties such as the density of states, as well as local properties such as the modified simplex approach to catalog materials and their properties. They apply the methods to the large datasets compiled by the Curtarolo group and find very interesting domains of material space once they apply the Tanimoto similarity metric.

The authors generate networks that they call materials carotgrams, where the nodes are compounds and the connections are the similarities between them. It is nice to see that naturally, regions of similar physicochemical properties emerge from their analysis. 

As with all these methods, the proof is in the pudding, and the challenge to this great collaboration is to propose a material that has not been synthesized yet, and show it has better properties than any other material out there.

Discovering chemistry with an ab initio nanoreactor.

Lee-Ping Wang, Alexey Titov, Robert McGibbon, Fang Liu, Vijay S. Pande and Todd J. Martínez

Nature Chemistry

6, 1044–1048 (2014)

Contributed by Alán Aspuru-Guzik

When is quantum chemistry going to take the primary discovery role, rather than an explanatory one? This one of the questions asked in the recent Nature Chemistry paper by the groups of Vijay Pande and Todd Martínez at Stanford.

With the advent of fast GPGPU quantum chemistry in codes such as Terachem, it is practical to carry out medium-scale quantum dynamics simulations for a considerable amount of time. In this paper, the authors employ a spherical "piston" to compress and decompress high-temperature reaction mixtures, which in turn result in an accelerated process for triggering chemical reactions. Automated algorithms are employed to recognize the nature of the chemical transformations involved. The authors use this approach to identify novel mechanisms for the Urey-Miller origins of life experiment.

In this modern world of Facebook, Twitter, Youtube, Facebook and blogs, a video is worth more than a thousand words. The video below summarizes the paper very well.

This paper is indeed, one of my top 5 favorite papers in the field of computational chemistry of 2014. 

Todd Martínez explaining the future of the nanoreactor in the QUITEL 2014 conference held in the Galápagos Islands in Ecuador. He suggested, in a joking fashion, that the nanoreactor could eventually evolve a flask with molecules into life ab initio. The virtual end product of the evolution seems to be the contributor of this post.

Quantum Chemical Approach to Estimating the Thermodynamics of Metabolic Reactions

Adrian Jinich, Dmitrij Rappoport, Ian Dunn, Benjamin Sanchez-Lengeling, Roberto Olivares-Amaya, Elad Noor, Arren Bar Even & Alán Aspuru-Guzik Sci. Rep. 2014, 4, 7022

Contributed by +Jan Jensen 

'

DOI: 10.1038/srep07022 CC-BY-NC-ND

Computational chemistry is finally in a position to conduct studies on hundreds or even thousands of molecules within a reasonable amount of time.  This has led to benchmark datasets which have been used to develop new semi-empirical methods (e.g. DFT-Dx or HF-3c) as well as high-throughput screening of molecules with desirable properties. 

This paper represents a third kind of study: a concerted effort to fill in missing experimental data needed to quantitatively understand a complex chemical system - in this case cellular metabolism. Standard Gibbs reaction free energies of biochemical reactions are crucial to quantitative modeling of metabolism (metabolic flux analysis) but only a small fraction have been measured experimentally.  

Many of the reactions involve molecule-sizes that can be modeled routinely by electronic structure calculations.  However, there are new challenges compared to, for example, atmospheric chemistry such as the modeling of pH- and solvation-effects and multiply charged ions.  Thus, any conformational search must often be repeated for several different protonation states and multiply charged ions often necessitate the treatment of explicit solvent molecules and, even then, result in relatively solvation energies with relatively large errors.  Also, multiply charged anions may present additional problems for DFT due to incomplete cancellation of the inter-electronic Coulomb repulsion at larger distances. 

This paper represents the first, pioneering, attempt to address these issues with high-throughput in mind and the results are encouraging. For reactions involving non-multiply charged species B3LYP/6-31G* with the COSMO solvation model augmented by explicit water molecules result in MADs less than 5 kcal/mol, while being fast enough to compute more than a hundred reaction energies (involving more than 5000 quantum calculations). 

However, this paper is more a call to action than the definitive answer.  Many new methodological improvements (perhaps such as this one?) are needed in order to increase the accuracy and efficiency needed produce a complete and accurate picture of cellular metabolism. I hope this paper will spur such developments - hopefully in the form of large and open collaborations (perhaps similar to this one?).  

Thanks to Adrian Jinich for alerting me to this paper

This work is licensed under a Creative Commons Attribution 4.0 International License.

Molecularspace.org

J. Hachmann, C. Román-Salgado, K. Trepte, A. Gold-Parker, M.A. Blood-Forsythe, L.R. Seress, R. Olivares-Amaya, A. Aspuru-Guzik, The Harvard Clean Energy Project Database
Contributed by +Jan Jensen

The Harvard Clean Energy project, lead by CCH editor Alán Aspuru-Guzik, has released the results of 150,000,000 DFT calculations on 2,300,000 compounds under the CC-BY license at www.molecularspace.org. The details of the calculations has been described in a 2011 Journal of Physical Chemistry Letters paper and will be described further in an upcoming paper.

The compounds in the database were selected as potential candidates for organic photovoltaics and the database contains HOMO and LUMO orbitals energies, and their difference, computed using various DFT functionals, which, in turn are used to estimate power conversion efficiency, the open circuit voltage, and the short-circuit current density using the standard Scharber model.

The HOMO and LUMO energies and gaps represent averages of several DFT methods and conformations and are calibrated against experimental data.  Details of this calibration are not given the 2011 JCP Letters paper, so they may appear in the forthcoming paper.

According to the website, molecularspace.org will soon be augmented by a designer module that "will allow you to help us design new materials. By following simple chemical rules and using our predictive model, you can help us develop new candidates for solar cell materials."

The calculations were made possible by using the World Community Grid managed by IBM.


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

Computational Design and Selection of Optimal Organic Photovoltaic Materials

Noel M. O’Boyle, Casey M. Campbell and Geoffrey R. Hutchison Journal of Physical Chemistry C 2011, 115, 16200 (Paywall)

Quantum chemistry for high throughput screening
In this work over 90,000 pi-conjugated copolymer were computationally screened for new and efficient organic photo-voltaics. Not being an expert in organic photo-voltaics I highlight this paper as a very interesting example of what I believe is an important emerging trend in the use of quantum chemistry: efficient high through-put screening of molecules for desirable properties.  Computers and software have now reached a point where this is computationally feasible to perform computations on thousands of molecules and the challenge is now to make it practically possible, i.e. to find the right combinations of methods and automate their use.

How does one construct 90,000 geometries?  Cheminformatics meets quantum chemistry
OpenBabel is used to construct the 3D structure of each polymer, starting from its SMILES string, and to find the lowest energy conformation using a weighted rotor-search and the MMFF94 force field.  This lowest energy conformation is then minimized with PM6 and used to compute energies and oscillator strengths of the 15 lowest- energy electronic transitions with ZINDO/S, which forms the basis for estimating the energy conversion efficiency (see the paper for more details).  The required CPU time is  8-10 minutes per polymer on a single core.  To put that number in perspective: using 100 cores, 100,000 polymers can be screened in roughly a week.

Starting from 131 different monomers, all possible (19,701) dimers are made and these dimers are then used to construct the corresponding 58,707 tetramers. (A brief description on how exactly this was scripted would have been a welcome addition to the supplementary materials).  The energy conversion efficiency was computed for all these dimer and tetramers.  These results were used to calibrate a genetic search algorithm that was used to identify hexamers and octamers with high energy conversion efficiency without doing an exhaustive search.

Promising candidates and new strategies
The current state-of-the-art is ca 8% energy conversion efficiency.  This study found 621 polymers (mostly hexamers and octamers) with greater than 9%  and 2 with greater than 11 % energy conversion efficiency.  (As the authors point out these polymers still need to be filtered for solubility, crystal packing, and other factors.)  Just as interestingly new design strategies emerged:

"Our analysis of component monomers, dimers, and the copolymer sequence demonstrates important design rules for copolymer photovoltaics. Most importantly, the conventional picture of combining a strong donor and strong acceptor into an alternating copolymer is found to frequently yield poor energy- level alignment. Instead, our top hexamers and octamers reflect a decreased optical band gap due to high coupling between the two-component monomers, not solely due to particular HOMO or LUMO energies of the monomers themselves."

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