Alex Tropsha: Applications of machine learning to materials and chemical property prediction

Alex TropshaAlex Tropsha, of the University of North Carolina Chapel Hill, UNC Eshelman School of Pharmacy, is benefiting from the explosive growth of materials data. There are 160,000 entries in the Inorganic Crystal Structure Database (ICSD). There are numerous commercial and open experimental databases (NIST, MatWeb, MatBase etc.), and huge databases such as AFLOWLIB, Materials Project, and Harvard Clean Energy. The chemical space of possible materials is huge : about 10100 candidates.16 The US government’s Materials Genome Initiative recognizes the need for new high performance materials. The growth of materials databases and emerging informatics approaches offers the opportunity to transform materials discovery into data- and knowledge-driven rational design.

AFLOW is a globally available database of 1,688,245 material compounds, with over 167,136,255 calculated properties. The optimized geometries, symmetries, band structures, and densities of states available in the AFLOWLIB consortium databases have been converted into two distinct types of fingerprints: Band structure fingerprints (B- fingerprints), and Density of States fingerprints (D-fingerprints).17 The framework is employed to query large databases of materials using similarity concepts, to map the connectivity of materials space (as a materials cartogram) for rapidly identifying regions with unique organizations and properties, and to develop predictive quantitative materials structure−property relationship (QMSPR) models for guiding materials design.

To represent the library of materials as a network (a material cartogram), the researchers considered each material, encoded by its fingerprint, as a node. Edges exist between nodes with similarities above certain thresholds (in this case, Tanimoto similarity and a threshold of 0.7). A materials map from B-fingerprints was made from 15,000 materials from ICSD, using DFT PBE calculations from AFLOWLIB. Four big clusters were observed: insulators, ceramics, and complex oxides; bimetals and polymetals; metallic and nonmetallic combinations; and small band gap semiconductors.

Novel descriptors (property-labeled materials fragments) not requiring prior DFT calculations have also been developed by Voronoi tessellation and neighbors search of crystal structures, followed by infinite periodic graph construction and property labeling, and generation of circular fingerprints.18 Starting from only a crystal structure, regression models can be built to predict band gap energy, and thus electronic properties, or to predict thermo-mechanical properties such as bulk modulus, shear modulus, thermal expansion, heat capacity, and thermal conductivity. All the models are trained based on DFT-computed properties. Heuristic design rules can be extracted.

Material informatics has also been applied to the design of a novel photocathode material for dye-sensitized solar cells (DSSCs).19 By conducting a virtual screening of 50,000 known inorganic compounds, the researchers have identified lead titanate (PbTiO3), as the most promising photocathode material. Notably, lead titanate is significantly different from the traditional base elements or crystal structures used for photocathodes. In experimental validation, the fabricated lead titanate DSSC devices exhibited the best performance in aqueous solution, showing remarkably high fill factors compared to typical photocathode systems. Currently, device performance is low, but it might be improved by designing a new dye.

Next Alex discussed applications of machine learning to designing chemicals with the desired physical and biological properties where compound structure is described only by its SMILES notation, and no other conventional chemical descriptors are used. The new approach developed in his lab is based on concepts from text mining that rely on neural networks to solve the problem of semantic similarity of texts.

The British linguist J. R. Firth is noted for drawing attention to the context-dependent nature of meaning. In particular, he is known for the 1957 quotation: “You shall know a word by the company it keeps”. To define the semantic similarity between two entities, Alex and his colleagues have made use of approaches embedded in Word2Vec, a neural-network-based approach to describe linguistic context of words developed at Google.20 With Word2Vec, a network is trained using each word of a corpus of text and some configurable number of surrounding words. The model can be trained to either predict the surrounding context based on the current word, or to predict the current word from the context. Elena Tutubalina and Alex (manuscript in preparation) have performed drug clustering in semantic similarity space, using webmd.com, patient.info, drugs.com, amazon.com, askapatient.com, and dailystrength.org as sources of user comments, and showed that drugs with similar pharmaceutical action do cluster together in the semantic similarity space.

Alex’s team has also experimented with de novo design of molecules with the desired properties using SMILES in Deep Reinforcement Learning:

Deep Reinforcement Learning Model

Structural bias, physical properties, and biological activity have been used in proof of concept case studies of user-biased molecular design. In summary, Alex cited Confucius who said, “Without knowing the force of words, it is impossible to know more”. Alex quipped “And remember: anything you say can, and will be used … for text mining!”.

References

  1. Walsh, A. Inorganic materials: The quest for new functionality. Nat. Chem. 2015, 7 (4), 274-275.
  2. Isayev, O.; Fourches, D.; Muratov, E. N.; Oses, C.; Rasch, K.; Tropsha, A.; Curtarolo, S. Materials Cartography: Representing and Mining Materials Space Using Structural and Electronic Fingerprints. Chem. Mater. 2015, 27 (3), 735-743.
  3. Isayev, O.; Oses, C.; Toher, C.; Gossett, E.; Curtarolo, S.; Tropsha, A. Universal fragment descriptors for predicting properties of inorganic crystals. Nat. Commun. 2017, 8, 15679.
  4. Moot, T.; Isayev, O.; Call, R. W.; McCullough, S. M.; Zemaitis, M.; Lopez, R.; Cahoon, J. F.; Tropsha, A. Material informatics driven design and experimental validation of lead titanate as an aqueous solar photocathode. Mater. Discovery 2016, 6, 9-16.
  5. Mikolov, T.; Chen, K.; Corrado, G.; Dean, J. Efficient Estimation of Word Representations in Vector Space. 2013, arXiv.org e-Print archive. https://arxiv.org/abs/1301.3781 (accessed September 4, 2017).