December, 2019 Newsletter 9

Summary of ESRS fellows´scientific outcomes

Raquel Rodríguez-Pérez

"During the BigChem project, I have had the opportunity to investigate in academia and industry. In particular, I worked at the University of Bonn and Boehringer Ingelheim in Germany. My research has focused on one of the main goals of pharmaceutical research, which is the discovery of small molecules that bind to a biological target and modulate its activity. The experimental search for new bioactive compounds is time consuming and expensive. However, due to the increasing amounts of compound data that are currently available, there are high expectations associated with machine learning (ML) to prioritize experiments. My research has focused on the prediction of compound activity using ML models. I have predicted experimental screening results across a panel of biological targets using distinct strategies. Moreover, I have generated ML models to classify highly and weakly potent inhibitors for more than a hundred kinases. Accurate ML models can guide compound design and further experiments, but insights on predictions are required in practical applications. Hence, I have also introduced an intuitive approach to rationalize activity predictions from complex ML models. Interpretability enables to further validate the model based on domain knowledge and extract information from the chemical features driving the predictions. I have presented my research at different conferences, including the ACS National Meeting in Orlando, Florida (Spring 2019) and published six papers with my major findings."

Oliver Laufkötter

" My research has focused on diverse aspects of virtual database screening methods. My first project involved developing a novel method for representing compound descriptors/fingerprints. The aim of this project was to improve upon existing methods and to apply this method in machine learning studies focused around drug/lead discovery. Specifically, this method involved building a hybrid fingerprint which combined features of both chemical structure information and bioactivity profiles of compounds, termed BaSH fingerprint (Bio-activity Structural Hybrid). The bio-activity profiles stem from high throughput screening studies where millions of compounds are tested in hundreds of biochemical or cell-based assays. This bio-activity and structural hybrid descriptor was termed BaSH fingerprint. By hybridizing these two descriptor types a synergistic effect could be obtained which expressed an improvement in compound activity predictions in random forest machine learning models. My second research project involved benchmarking various similarity searching techniques and identifying which methods are best suited in virtual screening applications. A number of techniques were compared using a variety of parameter settings, including standard similarity searching, Turbo similarity searching, Tversky based similarity searching, and consensus bit scaling similarity searching. The outcome of the study showed that using methods with parameter settings which devalue features of reference compounds gave the best results, i.e. Tversky with a low alpha parameter or consensus bit scaling with low scaling factor. This led to the identification of a novel similarity searching technique termed negative consensus profile scaling. Two articles have been published based on these works, the publications detailing these studies are listed below. The figure shows the concept behind the first research project using the hybrid compound descriptor."

Dr. Arkadii Lin

Arkaii Lin carried out his PhD work at Boehringer Ingelheim Pharma Co. & KG (Germany) and University of Strasbourg (France). He developed new tools and methodologies to visualize, analyze and model Big Data in chemistry using the Generative Topographic Mapping (GTM) approach. Key applications of those approaches include the diversification of the in-house collection of Boehringer Ingelheim, multi-target virtual screening and the determination of optimal frame set size to describe the entire ChEMBL collection. His work has been published in 3 peer-review articles and presented in eight conferences. He successfully defended his PhD in September 2019 and afterwards, took a PostDoc position in the laboratory of Chemoinformatics in Strasbourg in collaboration with Janssen Global Service.

Dipan Ghosh

Dipan PhD’s work at BIGCHEM has focused on the analysis of promiscuous compounds on chemical high throughput assays. Such potentially unwanted compounds should be identified and eliminated during the initial stages of screening. At HMGU, Dipan analyzed the assay interference problem due to the inhibition of luciferase, an enzyme widely used as a reporter. The work at the Lead Discovery Center centered on the identification on frequent hitters in G-protein coupled receptors (GPCR) assays. He has built robust machine learning models to identify promiscuous compounds in those assays and compared the performance with scaffold-based methods. All the models described were built using the OCHEM platform ( http://ochem.eu ) and are freely available to the public.

Laurianne David

“ The Bigchem program gave me the opportunity to work in scientific research in both the academic world, in Prof. Dr. Jürgen Bajorath’s lab in Bonn, and the industrial world, in Dr. Ola Engkvist’s Molecular AI team in AstraZeneca. I learnt a lot about scientific research and the scientific community. On my PhD topic, which was the analysis of compounds promiscuity, I produced a publication focusing on the identification of compounds that interfere with HTS technologies. I also had the opportunity to work with the AstraZeneca team and Boehringer Ingelheim’s researchers on a review paper about deep-learning in pharma. This PhD was an exciting experience!”

Josep Arús-Pons

" In these three years I have been working alongside Prof Jean-Louis Reymond in the University of Bern and with Dr. Ola Engkvist and Dr. Hongming Chen in AstraZeneca. The main goal of my work has been finding novel ways to explore the chemical space. While in Bern, I focused on creating enumerated molecular databases. Specifically the GDB4c database, which contained an enumeration of all ring systems with up to 4 rings. Further characterization of the database showed that 99% of the structures enumerated were novel, opening up possibilities in both synthetic and medicinal chemistry. On the other hand, in AstraZeneca I focused on exploring the chemical space using molecular deep generative models. These models are trained with a set of molecules (e.g. ChEMBL) and once trained are able to generate billions of molecules that are similar, but not equal, to those in the training set. Specifically, my research was understanding and improving models that use the SMILES syntax to represent molecules and to develop a benchmark to compare them. We were able to show that these models are able to easily generalize huge chemical spaces with very small samples. For instance, 83% of the GDB-13 database (containing 1 billion molecules) was learnt by a model trained with a set of only 0.1% of the database. All the results obtained during the PhD are a step in the direction of finding ways to explore unknown chemical space and thus to aid in the first steps of the drug discovery process. "

Dr. Xuejin Zhang (JJ)

JJ successfully finished her PhD work and defended her PhD thesis on 17 ^th September 2019 at ETH Hönggerberg.

Inspired by how human experts tackle the problem of reactant feasibility prediction for large combinatorial library design, Xuejin Zhang’s PhD’s work at BIGCHEM aimed to address this task by building a recommender. This recommender should be able to prioritize the reactants that bear compatible functional group, with high likelihood of being reacted for a reaction of interest.

Thomas Blaschke

"I investigated, developed, and validated methodological aspects of generating molecules with desirable properties. More specifically, I used different architectures of generative neural networks (NN) and researched how they can be fine-tuned to create novel molecules with certain specified properties. Currently, I authored / co-authored five publications. In the first studies, we introduced a method to tune a sequence-based generative NN for molecular de novo design that can learn to generate structures with specified desirable properties. We demonstrated how the model could execute a range of tasks such as generating analogs to a query structure and generating compounds predicted to be active against a biological target. In another work, we investigated the potential use of different NN autoencoders (AE) for de novo molecular design. An AE is a pair of NNs that maps a molecular structure into a continuous numerical representation, so-called latent space, and is able to reconstruct molecules from the latent space. The study showed that the latent space preserves the chemical similarity principle and thus can be used for the generation of analog structures. Furthermore, we studied what amount of chemical space can be sampled with generative models that are trained only with a very small sample of an enumerated database. We developed a method to assess the quality of the training process at different stages of the training and studied how much of the enumerated database can be reproduced by sampling the generative model at different stages of the training. Finally, we studied the ability of compounds to interact with multiple targets, so-called promiscuity. Various machine learning models were built to distinguish between promiscuous and nonpromiscuous compounds. The different machine learning models yielded overall successful predictions and provided evidence for the presence of structure−promiscuity relationships."

Amol Thakkar

Computer Assisted Synthesis Planning (CASP) has gained considerable interest as of late. We have investigated CASP from two avenues, retrosynthetic route prediction, as an idea generator and reaction prediction as a form of validation, and likelihood estimator. Ultimately, the goal is to unite the two to provide an AI synthesis prediction tool capable of assessing the synthetic accessibility of a given target compound(s), whilst proposing routes with a literature precedent, and an assessment of the likelihood of failure. Such a tool has broad implications, in the form of time savings, efficient idea generation, and rapid assessment of synthesizability, thus reducing the overall failure rate. The retrosynthetic model which we have developed to date has been assessed on a series of internal virtual libraries and can predict the route to synthesized compounds with a ca. 70 % success rate on average. In the future, and as reaction prediction methods develop, the two approaches will be incorporated into one entity, providing a powerful tool for chemists, the automation lab, and de novo design groups to efficiently asses the synthetic accessibility of future potential small molecule therapies.

The PhD is currently ongoing and will continue for a period of six months at AstraZeneca, before returning to the University of Bern for the remaining years. It is hoped that as the tool develops, we are able asses how well predictions translate into the wet lab, and how they help researchers investigate synthesis in previously unexplored regions of chemical space.

Michael Withnall

At BIGCHEM, Michael has been working on the secure sharing of information using machine learning. His work began at HMGU with the Tetko group, focusing on molecular transformations and multitask learning, which resulted in two publications. During his subsequent stay at AstraZeneca, Michael worked on various contemporary deep learning algorithms under Dr. Hongming Chen and Dr. Ola Engkvist's supervision. His work used low-level molecular representations, and image data, to try to maximise information usage in activity and property prediction. This work has resulted in a conference paper and in another publication that is currently under peer review. Michael expects to submit and defend his thesis at the Technical University of Munich in early 2020.

Publications

Presentations of BIGCHEM work

BIGCHEM Newsletters