AIDD Spring School - Advanced Machine Learning for Drug Discovery
09 May 2022 - 18 May 2022
East Campus USI-SUPSI, Lugano-Viganello
AIDD is a European Project which aims at training and preparing a new generation of scientists who have skills both in machine learning and in chemistry and can advance medicinal chemistry. In this context IDSIA USI-SUPSI is proud to host and co-organise the first Summer School on advanced Machine Learning for Drug Discovery.

Programme

Link to the Zoom broadcast (Valid only for this session)

 

  • 11:00 - 12:00 / 13:00 - 13:50 - Sequential decision making, RL and MDPs
    Speaker: Oleg Szehr

    This is a 3 times 45mins introductory crash course in Markov Decision Processes and Reinforcement Learning (RL). The course provides students with an overview of the main theoretical concepts and introduces a basic toolbox of RL algorithms. It enables students to recognize and formalize RL problems and to choose appropriate RL algorithms for their practical work. The first lecture introduces the concept of learning from a reward signal and (bandit-) optimization algorithms for immediate rewards. The second lecture introduces Markov Decision Processes more broadly and presents some classical techniques for their solution (such as Monte Carlo and Temporal Difference methods). The focus of the third lecture lies on neural-network based RL algorithms, including neural bandit models, deep Q-learning and Alpha Go.
     
  • 13:50 - 14:40 - Synthesis planning strategies
    Speaker: Philippe Schwaller

    In organic chemistry, we are currently witnessing a rise in machine learning approaches, which show great potential for improving molecular designs and accelerating the discovery of novel molecules. One of the bottlenecks in the molecular design cycle is the synthesis. Can machine learning overcome this bottleneck and facilitate the synthesis of organic molecules that have never been made before? In my lecture, I will introduce multiple synthesis planning-related tasks and provide an overview of the different contributions (with a focus on Transformer models) that are at the foundation of the digital synthetic chemistry revolution.
     
  • 15:00 - 15:50 (ZOOM) - SELFIES: self-referencing embedded strings
    Speaker: Florian Häse

    Efforts to model molecular systems - whether explicitly based on physical principles or implicitly with large amounts of data - rely on meaningful representations of molecules themselves. While a quantum mechanical approach to derive some electronic properties of a molecule might require the molecule to be described in terms of the spatial arrangement of its nuclear charges, a machine learning approach might benefit from other representations that more efficiently expose the constraints and intrinsic symmetries governing the targeted modeling use case. This talk focuses on a string-based representation of small organic molecules - SELFIES - which was conceived to be robust with respect to single (or multiple) character mutations. Following the concept of a Chomsky-type-2 context-free grammar, SELFIES strings can be related to organic molecules whose atoms satisfy basic chemical constraints arising from orbital hybridization. In this sense, SELFIES constitutes a representation for "syntactically" valid molecules. We will demonstrate how a SELFIES string can be derived for a small organic molecule and will discuss the advantages (and drawbacks) of SELFIES in comparison to other string-based representations. Finally, we will highlight a few selected applications of SELFIES for generative modeling and beyond to spark a discussion on future directions for the representation of chemical systems.

     
  • 16:00 - 16:50 - HPC for drug discovery
    Speakers: Silvano Coletti and Carmine Talarico

    Exscalate4CoV, the private-public consortium supported by the EU’s Horizon 2020 programme for research and innovation, led by Dompé Farmaceutici and representing 18 partners, during past months has intensively used the competences and the resources of the European supercomputing centers installed at, CINECA, Barcelona Supercomputing Center and FZ Juelich, and industrial HPC facility such as ENI HPC5 supercomputer, to simulate the interactions between the proteins of the coronavirus SARS-CoV-2 and the molecules of potential therapeutic drugs, in a race against time to identify a therapy to fight the virus effectively.

    During the project we have generated valuable data: more than 600 active molecules identified so far out of +70000 experimental data generated (+50000 and 20000 within E4C and COVIRAL project respectively), and +27 peer review papers with a global impact factor >154 points in the first year of the project and a final impact factor of 200.

    We deployed precious web platforms to support the global research community with bioinformatics and simulation tools. The most complete (> 60 simulations) and the most informative (>10 µs) set of SARS-COV-2 molecular dynamics simulations were released thanks to the best European HPC resources. Given the ongoing health emergency, we have pushed the best hardware and software technologies to the extreme performing the world’s biggest and the fastest virtual screening simulation ever, looking for novel molecules running more than 1 trillion simulations in one single shoot, along with the deployment of ad hoc virtual screening protocols and X-ray validation of the most relevant findings. We were involved in spreading the knowledge among scientists around the world improving our capabilities to share information, by using cutting edge technologies such as virtual reality. These innovative methods made the ability to understand and evaluate the results straightforward during the project.

    Yet importantly 1 completed clinical trial with 150 patients in 3 EU countries, and another 3 clinical candidates.

 

  • 09:30 - 10:30 - Artificial curiosity
    Speaker: Jürgen Schmidhuber

    For over three decades we have published work about artificial scientists equipped with artificial curiosity and creativity. In this context, I have frequently pointed out that there are two important things in science: Finding answers to given questions, and (B) Coming up with good questions. is arguably just the standard problem of computer science. But how to implement the creative part (B) in artificial systems through reinforcement learning (RL), gradient-based artificial neural networks (NNs), and other machine learning methods? Here I summarise some of our approaches:

    1990: Curiosity through the principle of generative adversarial networks
    1991: Curiosity through NNs that maximise learning progress
    1995: RL to maximise information gain or Bayesian surprise. (2011: Do this optimally)
    1997: Adversarial RL agents design surprising computational experiments
    2006: RL to maximise compression progress like scientists/artists/comedians do
    2011: PowerPlay continually searches for novel well-defined computational problems whose solutions can easily be added to the skill repertoire, taking into account verification time
    2015: Planning and curiosity with spatio-temporal abstractions in NNs
    2022: Ongoing work
     
  • 11:00 - 12:00 - Constructing accurate machine learning force fields for flexible molecules

    Speaker: Leonardo Medrano

    In recent years, the use of machine learning (ML) in computational chemistry has enabled numerous advances previously out of reach due to the computational complexity of traditional electronic-structure methods. One of the most promising applications is the construction of ML-based force fields (FFs), with the aim to narrow the gap between the accuracy of ab initio methods and the efficiency of classical FFs. This has provided a tool for computing several physicochemical properties that would require millions of CPU years otherwise. The key idea is to learn the statistical relation between chemical structure and potential energy without relying on a preconceived notion of fixed chemical bonds or knowledge about the relevant interactions. However, many successful applications of MLFFs have been restricted to small-and medium-sized molecules. This lecture gives an overview of the core concepts underlying ML-FFs and discusses the challenges for the generation of MLFFs of more flexible and complex molecular systems.
     
  • 13:00 - 18:00 - Project check
  • 09:30 - 12:00 - Project check
     
  • 13:00 - 13:50 - Students' groups meetings: planning of project discussions in groups  

  • 13:50 - 14:40 - Few and zero-shot learning in drug discovery
    Speaker: Günter Klambauer 

    Building models for molecular properties and activities based on very few measurements is a central problem in drug discovery (DD). Almost all drug discovery projects start with no or few known active molecules and face the problem of selecting promising molecules for screening. Therefore, zero- and few-shot learning methods have been introduced to computer-aided drug design, which have the potential to improve this critical phase of the drug discovery process. In this talk, an overview of zero- and few-shot methods in DD is given and a framework in which zero- and few-shot learning methods can be explained and related to each other is suggested.


  • 15:00 - 15:50 - Experimental computational work
    Speaker: Mike Preuss

    Many developments in modern computing come with insufficient theory, thus most questions on how to apply what to which problem have to be resolved purely experimentally. However, methodology here is not very strong, and usually not explained to newcomers. In this talk, I give an overview over common mistakes in applying machine learning/AI algorithms and also hint at the foundations of experimentation as they are known in other sciences. These are largely applicable to experimental computational work as well and lead to more informative, more reliable and more replicable experimental results.

     
  • 16:00 - 16:50 - Magic rings: navigation in the ring chemical space guided by the bioactive ring
    Speaker: Peter Ertl

    The large majority of bioactive molecules contain a more or less complex ring system as a central structural element. This central core determines the basic molecule shape, keeps substituents in their proper positions, and often also contributes to the biological activity itself. In this study the ring systems extracted from one billion molecules are processed and differences between rings from bioactive molecules and common synthetic molecules are analyzed. The bioactive rings seem to be distributed throughout the large portion of chemical space, but not uniformly; one can see several more dense regions, where the bioactive rings often appear in small clusters, as well as empty areas. A web tool offering an interactive navigation in the ring chemical space and supporting identification of bioisosteric ring analogs available at https://bit.ly/magicrings is also described.
  • 09:30 - 10:-30 - Deep geometric learning
    Speaker: Silvio Giancola

    Geometric Deep Learning is a niche domain in Artificial Intelligence, that generalizes neural networks from Euclidean to non-Euclidean domains. In his talk, Silvio will review the literature in Geometric Deep Learning applied to 3D Computer Vision, highlighting the latest findings published from his research group. The talk will highlight the paradigms and trends in 3D computer vision, present the variety of representations for 3D shapes and scenes, spotlight the challenges in processing such diverse non-Euclidian data, and review the current literature in 3D computer vision tasks.
     
  • 11:00 - 12:00 - Artificial intelligence and the chemical space
    Speaker: Jean-Louis Reymond

    Since the advent of organic chemistry as a basis for drug development, millions of organic molecules of various sizes and shapes have been discovered or synthesized and tested for various properties, and thousands of them have become clinical drugs. Can we understand this data and use it to discover new drugs addressing unmet medical needs? To contribute to this endeavor, we develop computational tools to enumerate, visualize and search the vast chemical space of drug-like molecules and use our tools in applied projects involving organic synthesis and bioassays. I will discuss the impact of artificial intelligence for enumerating chemical space, selecting new molecules, assigning their targets and plan their syntheses, with experimental examples in small molecule drugs and peptides.

  • 13:00 - 13:50 - Comparing and clustering synthetic route prediction
    Speaker: Samuel Genheden

    The list of available software and published algorithms for multi-step retrosynthesis has increased substantially recently due to the rise of deep learning and artificial intelligence. Such methods typically produce a large number of synthetic routes that can be assessed by chemists. However, it is a challenge to analyze a large number of routes, and to this end we developed a method to compute the similarity between two routes, which can be used as the basis for clustering the predictions into a manageable set of groups. The route similarity and route clustering techniques can also be used to benchmark two competing route prediction methods. We have developed a benchmark framework called PaRoutes that provide experimental reference routes and metrics to evaluate how well a route prediction experiment reproduce the reference routes. This will be an important tool in comparing novel developments, understand the limitations of the state-of-the-art, and eventually pave the way for methods able to affect molecular design efforts. 
     
  • 13:50 - 15:50 - Explainable AIU: interpreting, explaining and visualising deep learning
    Spekaers: Alessandro Antonucci and Alessandro Facchini

    Fuelled by progress in computer power and the large availability of data, models generated by machine learning (ML) algorithms are gaining wide currency in scientific research. In some fields, their performance revolutionises the traditional approach to scientific inquiry based on theories, models, and experiments and promotes a transition towards a data-centric science. Unfortunately, ML models suffer from the problem of being epistemically opaque, which roughly means that their format, structure, and complexity prevent human users from understanding their functioning and behaviour on various levels, and therefore from relying on them in critical situations.

    The recent research program of eXplainable Artificial Intelligence (XAI) develops methods of countering the various manifestations of epistemic opacity displayed by Machine Learning (ML) systems. The aim of these methods is to make the complex algorithms and the often conceptually alien and intractable decision making of ML systems more humanly understandable. In the XAI literature, various types of opacity are identified at various levels of AI models, and various types of solutions are presented for a wide range of application domains. In addition to high-stake areas such as legal and medical systems, growing attention is being paid to the questions of how specific types of opacity particularly affect the use of ML models in scientific research, and how XAI and related methods might contribute to addressing these problems. The particular issue with ML approaches in science lies in the prima facie mismatch between the norms of scientific explanation and understanding through experiment, model building and theory construction on the one hand and the putatively association-based and theory-free predictions of AI on the other.

    The lecture is subdivided in two parts. In the first part (~30min), opacity is introduced as a concept whose meaning depends on the context of application, and on the purposes and characteristics of its users. In particular, we introduce the distinction between access and link opacity. Discussing the latter, we will highlight its implications for the fundamental purposes of scientific inquiry, such as explanation, understanding, and objectivity, but also how XAI methods might be employed to formulate theoretical hypotheses regarding the studied natural phenomenon. The second part (~60min) is intended as a critical discussion of the most popular XAI tools proposed so far. This is achieved by first presenting a number of (somehow) “transparent” models already adopted in classical ML. Popular explanation tools such as uncertainty and saliency maps as well as textual and numerical (e.g., Shapley’s) explanations are indeed presented. Finally, we present a number of XAI tools (such su LIME, LORE, DeepExplain, and SHAP), together with a number of demonstrative applications related to different domains.

  • 16:00 -16:50 - Project discussion by students
  • 09:30 - 10:30 / 11:00 - 12:00 - Graph neural networks at the service of molecular simulations
    Speaker: Vittorio Limongelli

    Molecular binding interaction like drug/protein is one of the fundamental processes in biology. Elucidating its structural and energetic aspects might help in understanding cell functioning and develop ad hoc exogenous control (e.g., drug design) [1]. However, studying such processes is often elusive to both experimental and theoretical techniques due to their limiting size- and time-scale. In the recent years, significant advance has been marked by developing coarse-grained and enhanced sampling simulation techniques [2,3], however the sampling capability still represents a limiting factor. In the present talk, we are going to overview the theory of drug/protein interaction and see how algorithms loosely referred to with the name of machine learning (ML) might be employed in the field of atomistic simulations. In particular, I introduce an innovative transfer learning methodology that is able to learn the free-energy of a given molecular system - obtained from accurate calculations - and transfer such information on a previously unseen molecular system of different size - i.e., a system with a significantly larger number of atoms and degrees of freedom that cannot be easily characterized by free-energy calculations. The proposed methodology relies on utilizing (i) a novel hypergraph representation of molecules, encoding all the relevant physico/ chemical properties for characterizing the potential energy of a conformation; and (ii) novel message passing and pooling layers for processing and making predictions on such hypergraph- structured data. Despite the complexity of the problem, our results show a remarkable AUC of 0.92 for transfer learning from tri-alanine to the deca-alanine system in the classification of high and low energy states. Moreover, we show that the very same transfer learning approach can be used to group, in an unsupervised way, various secondary structures of deca-alanine in clusters having similar free-energy values. This result is very notable since deca-alanine assumes secondary structures - i.e. states - which are not present neither in alanine nor in tri-alanine used as training set.
    A tutorial session follows the lecture in which we present a graph neural network able to reproduce the inter-atomic potential with a comparable accuracy and potentially improved performance with respect to a molecular dynamics simulator.

    References:
    Limongelli V. Wiley Interdiscip. Rev. Comput. Mol. Sci. 10, e1455 (2020)Raniolo S, Limongelli V. Nat. Proc. 15, 2837-2866 (2020)
    Souza et al. Nat. Commun. 11: 3714 (2020)

  • 13:00 - 13:50 (ZOOM) - Structure based drug discovery
    Speaker: Michael Sattler

    The talk will cover an introduction to integrative structural biology (combining NMR, SAXS/SANS, X-ray crystallography and cryo-EM) to elucidate details of the structure and dynamics of biomolecular complexes (protein-protein, protein-RNA), also considering opportunities and caveats of AlphaFold2. Recent examples will be presented that highlight the role of dynamic conformations and transient interactions. Approaches and examples will be presented for structure- and fragment-based drug discovery and the role of computational methods and opportunities using machine learning approaches in targeting protein-protein interactions and RNAs.
     
  • References:
    Napolitano, V. et al. Acriflavine, a clinically approved drug, inhibits SARS-CoV-2 and other betacoronaviruses. Cell Chem. Biol. 1–24 (2022) doi:10.1016/j.chembiol.2021.11.006.
    Lopez, A. et al. (2021) Client binding shifts the populations of dynamic Hsp90 conformations through an allosteric network, Science Advances, 7(51) eabl7295
    Softley, C. A., Bostock, M. J., Popowicz, G. M. & Sattler, M. Paramagnetic NMR in drug discovery. J. Biomol. NMR 74, 287–309 (2020).
    Jagtap, P. K. A.; et al Sattler, M. Identification of Phenothiazine Derivatives as UHM-Binding Inhibitors of Early Spliceosome Assembly. Nat Commun 2020, 11 (1), 5621. https://doi.org/10.1038/s41467-020-19514-1.
    Jagtap, P. K. A.; Asami, S.; Sippel, C.; Kaila, V. R. I. I.; Hausch, F.; Sattler, M. Selective Inhibitors of FKBP51 Employ Conformational Selection of Dynamic Invisible States. Angew. Chemie - Int. Ed. 2019, 58 (28), 9429–9433. https://doi.org/10.1002/anie.201902994.
    Kooshapur, H. et al. Structural basis for terminal loop recognition and stimulation of pri-miRNA-18a processing by hnRNP A1, Nature Commun 9, 2479 (2018)
    Jagtap, P. K. A.; Garg, D.; Kapp, T. G.; Will, C. L.; Demmer, O.; Lührmann, R.; Kessler, H.; Sattler, M. Rational Design of Cyclic Peptide Inhibitors of U2AF Homology Motif (UHM) Domains to Modulate Pre-MRNA Splicing. J. Med. Chem. 2016, 59 (22), 10190–10197. https://doi.org/10.1021/acs.jmedchem.6b01118.

  • 13:50 - 14:40 - AI formula generator
    Speaker: Guillaume Godin

    Recently, Machine learning models are used to generate arts abstractive objects including music, image and painting creation. This generative trend also touched our senses with attempt in Whisky or Beer.
    Over last years, we have created methods to make new recipe formula composed of synthetic molecules and natural ingredients for perfumery and aroma.
    In this talk, we will show how to acheive robust tools to assist perfumers and flavorist experts to optimize knowledge and increase creativity. Our research won the last Swiss Digital innovation award 2021.
     
  • 15:00 - 15:50 (ZOOM) - Conformal prediction for the design problem1
    Speaker: Clara Wong-Fannjiang

    In many real-world deployments of machine learning, we use a prediction algorithm to choose what data to test next. For example, a data-driven approach for designing proteins is to train a regression model that predicts some real-valued property of a protein sequence, then use it to propose new sequences believed to exhibit higher property values than observed in the training data. Since validating designed sequences in the wet lab is typically costly, it is important to quantify the uncertainty in the model's predictions. However, this is challenging because of a characteristic type of distribution shift between the training and test data in the design setting---one in which the training and test data are statistically dependent, as the latter is chosen based on the former. Consequently, the model's error on the test data---that is, the designed sequences---has an unknown and possibly complex relationship with its error on the training data. We introduce a method to quantify predictive uncertainty in such settings. We do so by constructing confidence sets for predictions that account for the dependence between the training and test data. The confidence sets we construct have finite-sample guarantees that hold for any prediction algorithm, even when a trained model chooses the test-time input distribution. As a motivating use case, we demonstrate how our method can quantify uncertainty for the predicted property values of designed proteins, and can therefore be used to select design algorithms that achieve an acceptable trade-off between high predictions and low predictive uncertainty.
  • 09:30 - 10:30 / 11:00 - 12:00 - Gaussian processes and sequential design of experiments
    Speaker: Dario Azzimonti 

    Gaussian processes (GPs) are a powerful, non parametric, Bayesian method for regression and classification tasks. The Bayesian nature of GPs allows for a natural quantification of uncertainties and for sequential updates of the model. Moreover, GPs have a flexible way to incorporate prior information through the definition of kernel functions. In this talk we introduce GPs for regression tasks, we review the relationships between the kernel function and the prior distribution and highlight their properties. We then focus on an important application of GPs: sequential design of experiments, and, in particular, Bayesian optimization. GPs can be used to approximate expensive to evaluate ("black-box") functions. We can improve this approximation by querying those black-box functions at appropriate locations and updating the model. If we are interested in finding the optimum of such functions, we can use Bayesian optimization techniques. We give here an overview of Bayesian optimization and sequential design of experiments by reviewing the main steps and the most important acquisition functions along with their advantages and drawbacks.

  • 13:00 - 13:50 - Bayesian inference
    Speaker: Adam Arany

    The term ""Bayesian"" have multiple related meanings in machine learning circles. In the most basic level it refers to an interpretation of the concept of probability. It is also used to refer to the Bayes' rule and methods based on it (e.g. Bayesian networks). Furthermore, It is used to indicate the practice of not distinguishing parameters and random variable in a model. All of these properties are intimately connected.
    In my talk I will give a brief introduction on the Bayesian interpretation of probabilities. Than, I will concentrate on the practical aspects of the framework: how we represent probabilistic models and how we update our belief about them. What we do in practical situation with finite computational resources? I will touch on Monte Carlo and Variational Bayes methods.
     
  • 13:50 - 14:40 (ZOOM) - Cell painting assay, data analysis and reporting, and its application for identifying biological activity in new chemical matter
    Speaker: Axel Pahl

    The Cell Painting Assay (CPA) is a cell-based morphological assay that was originally developed by the group of Anne Carpenter at the Broad institute. [1]
    Its main advantage over other morphological assays is its unbiased nature. Analyzing images of multiple stains that are selective for different compartments of the cells, the Open Source software CellProfiler [2] (also developed by the Broad) generates profiles of several hundreds of features that describe the cell's morphology.
    When cells are treated with test compounds with unknown biological activity, their feature profiles can be compared to those of reference compounds with known activity. This allows the detection and identification of a wide range of activities without requiring a prior target hypothesis.
    CPA has been established as a routine screening assay at our institute to identify activity in new chemical matter. We created a custom downstream analysis pipeline to calculate and compare the feature profiles and generate result reports. In addition, a suite of web tools enables interactive analysis of the data by the users.
    The talk gives an introduction to the CPA, our analysis workflow, example results and limitations.

    [1] Bray, M.-A.; Singh, S.; Han, H.; Davis, C. T.; Borgeson, B.; Hartland, C.; Kost-Alimova, M.; Gustafsdottir, S. M.; Gibson, C. C.; Carpenter, A. E. Cell Painting, a High-Content Image-Based Assay for Morphological Profiling Using Multiplexed Fluorescent Dyes. Nature Protocols 2016, 11 (9), 1757–1774. https://doi.org/10.1038/nprot.2016.105. [2] https://cellprofiler.org/
     
  • 15:00 - 15:50 - Overview of toxicity prediction methods
    Spekaer: Emilio Benfenati

    The in silico approaches are nowadays facing a new level of maturity, since they are no more research tools, but also practical tools which are used by authorities and industry, for several purposes. This requires extension of the purposes. The tools should be documented, interpretable, and help reasoning. They should be close to the user’s need, which also means to adapt to the user’s language, and make explicit reference to the regulatory thresholds, for instance. It should be possible to link different tools, addressing different purposes for authorities and industry. Some of these needs are and have to be the same. In particular, tools for adverse properties have to be applied by both industry and regulators. These should be the same tools, since the same conclusions should be obtained by all users. Thus, here we should prefer tools of easy access and good documentation. These tools should address risk assessment, thus ideally should cope with both exposure and hazard. On the other hand, there should be tools used for prediction of functional use, and this refers uniquely to industrial uses. This kind of tools should be linked to the tools for adverse properties, in order to optimize the desired properties through a multi-task perspective.

    These are formidable challenges. Collaborations from multiple players is fundamental. Here we will introduce some examples of recent in silico tools. We will discuss VERMEER, ToxEraser and JANUS software programs, and relate them to the current policy and industry needs.
     
  • 16:00 - 16:50 - Project discussion by students
  • 09:30 - 10:30 - IPR introduction and practical session
    Speaker: Sigrid Scheek

    The two-part seminar will give an introduction on intellectual property (IP) and outline various ways how IP can be secured for commercialization. It will give insights in the requirements and procedures of patenting and will explain the legal rights conferred by a patent. A comparison to copyright protection and alternative licensing strategies will be presented. The impact and benefit of IP protection, especially in a highly innovative and knowledge-based environment such as academia, will be discussed based on hands-on examples. It will be demonstrated how the wealth of additional information that is available in the patent literature can be accessed. The objective of this seminar is to create awareness of intellectual property issues among researchers and to give tips how to deal with IP related issues in scientific everyday life.
     
  • 11:00 - 12:00 (ZOOM) - Molecular graph neural network explainability
    Speaker: Floriane Montanari

    Machine learning models are routinely used in drug discovery projects to predict pharmacokinetic properties and biological activities of small molecules. Usually, such models are built on molecular fingerprints or directly from the molecular graph. The predictions that are obtained might or might not be accompanied by a reliability score.

    Over the years, different techniques have been developed to bring interpretability to such models. They aim at answering the question: “why is the model predicting this value for my input molecule?”. The rationale behind is that medicinal chemists have their own intuition about molecular interactions formed by small molecules and might want to fact-check some predictions made by models. Additionally, if interpretability can be boiled down to substructures and particular atoms, it can guide chemists towards beneficial modifications of the chemical structures.
  • In this talk, we will show different ways of bringing explainability to machine learning models built on small molecules, from SMILES atom substitutions to architectural constraints in graph neural networks.

  • 13:00 - 13:50 - Equivariant (G)NNs
    Speaker: Marco Bertolini

    In this lecture I will introduce the main concepts necessary to define a neural network architecture which is equivariant with respect to translations, rotations, permutations. I will start by reviewing some tools from group and representations theory, before delving into how these are applied in the context of deep learning. I will present some applications of Equivariant Neural Networks to interesting problems relevant for drug discovery use-cases: conformer and electron density prediction, as well as unsupervised invariant representation learning of small molecule descriptors.
     
  • 13:50 - 14:40 - IPR. Practical session
    Speaker: Sigrid Scheek
     
  • 15:00 - 15:50 - A quick guide to Markov Decision Processes and Reinforcement Learning (3)
    Speaker: Oleg Szehr

    This is a 3 times 45mins introductory crash course in Markov Decision Processes and Reinforcement Learning (RL). The course provides students with an overview of the main theoretical concepts and introduces a basic toolbox of RL algorithms. It enables students to recognize and formalize RL problems and to choose appropriate RL algorithms for their practical work. The first lecture introduces the concept of learning from a reward signal and (bandit-) optimization algorithms for immediate rewards. The second lecture introduces Markov Decision Processes more broadly and presents some classical techniques for their solution (such as Monte Carlo and Temporal Difference methods). The focus of the third lecture lies on neural-network based RL algorithms, including neural bandit models, deep Q-learning and Alpha Go.
     
  • 16:00 - 16:50 - Project discussion by students
  • 09:30 - 10:30 - Presentation of the "VIRTUOUS" project
    Speakers: Dario Piga and Gianvito Grasso

    VIRTUOUS (Virtual tongue to predIct the oRganoleptic profile of mediterranean IngredienTs and their effect on hUman hOmeostasis by means of an integrated compUtational multiphysicS platform) is an MSCA Project which aims at realizing  an “artificial tongue”  to predict the organoleptic profile of foods starting from their molecular composition. The algorithm for taste prediction integrates competences on molecular modelling, bio-informatics, cloud computing and  machine learning. Potential application of VIRTUOUS is in the food industry, modern precision agriculture, and as a link to neuroscience for taste determination. In the talk, we will present the main ideas behind VIRTUOUS, along with results achieved in the prediction of sweet and bitter molecules.
     
  • 11:00 - 12:00 - Impulse talks and brainstorming on "challenges in ML

 The programme is still subject to late changes.

Contacts
st.wwwsupsi@supsi.ch