Friday, 10th March, 2017, 11:00-12:00
Speaker: Juan Fernández Recio, Research Director, BSC.
Room Ramón y Cajal Room, PRBB Building
Friday, 20th January, 2017, 13.00 - 14.00
Speaker: Giorgio Colombo, Instituto di Chimica del Riconoscimento Molecolare, CNR, Italia,
Room Xipre Room (Seminar 173.06-183.01), PRBB Building
Thursday, 15th September, 2016, 12.30 - 13.30
Drug design lags far behind other engineering disciplines in lacking predictive, quantitative models that allow small-molecule therapeutics to be designed, rather than fortuitously discovered. While many challenges exist to building these models, our laboratory uses cycles of computational predictions coupled to experimental measurements to rapidly generate data that can be used to improve rigorous, quantitative approaches to small molecule design based on alchemical free energy calculations. In this talk, we describe how this process can be done cheaply in an automated manner by inverting the drug discovery problem, and describe our first few steps toward this goal in the design of selective kinase inhibitors.
Speaker: John Chodera, Assistant Faculty Member, Computational Biology Memorial Sloan-Kettering Cancer Center, NYC, USA
Room Xipre Room (173.06-183.01), PRBB Building
Friday, 29th April, 2016, 11.00 - 12.00
After a brief introduction to deep architectures and their typical supervised and unsupervised training approaches, the talk focuses on incremental strategies (at the base of natural learning). We will present our experience on incremental training of both CNN (Convolutional Neural Networks) and HTM (Hierarchical Temporal Memory). In particular a recently proposed semi-supervised tuning strategy (exploiting time coherence) proved to be very effective in conjunction with HTM, sometimes approaching supervised training accuracy.
Speaker: Davide Maltoni, University of Bologna (Dept. of Computer Science and Engineering - DISI)
Room Xipre Seminar (173.06)
Friday, 26th February, 2016, 11.00-12.00
The goal of Artificial Intelligence (AI) is to build machines that can display complex behavior, such as for example reaching human level performance in some tasks. The Machine Learning approach to achieve this goal is to provide the machine with a powerful mathematical framework and data from which complex behaviors can be learned. In this talk, I will introduce deep neural networks and reinforcement learning, which are considered two of the most promising mathematical frameworks solving difficult tasks in many different domains, and discuss their strengths and challenges.
Speaker: Dr. Silvia Chiappa, Senior Researcher at Google Deep Mind, UK.
Room Charles Darwin Room
Thursday, 5th November, 2015, 12.00-13.00
The era of personalised medicine offers at once a huge opportunity and a major challenge to computational science. The potential impact centres around our ability to marshall substantial quantities of patient data and to use them to perform predictive, mechanistic modelling and simulation in order to deliver therapies and to enhance clinical decision making, on time scales which are far shorter than those usually considered in the context of academic research and development activities. Secure access to personal data, as well as to powerful computational resources, is essential. I shall provide a couple of examples which illustrate the current state of the art. One addresses clinical decision support in the context of blood flow within neurovascular pathologies; the other is concerned with patient specific drug discovery and treatment. We shall discuss the underlying e-infrastructure requirements, including data, compute and networks, and reflect on the potential for cloud and other forms of e-infrastructure provision to meet the anticipated future demand for resources.
Speaker: Peter V. Coveney, Department of Chemistry, UCL.
Room Charles Darwin Room
Wednesday, 3rd December, 2014, 12.00-13.00
In biological systems nearly all processes are governed by interactions between protein molecules, which have evolved to perform an innumerable variety of functions. The ability of some proteins, such as antibodies, to interact with high affinity and specificity is being increasingly exploited for therapeutic and diagnostic applications. Yet, using current methods it is laborious and often difficult to generate antibodies against specific epitopes within a protein, in particular within disordered regions. Likewise, the successful development of antibody-based drugs is often hindered by their relatively poor solubility, which leads to aggregation at the high concentrations necessary for effective storage and delivery.
I will present two computational approaches to rationally modify interactions between protein molecules, including antibodies. The aim of the first is to hamper aberrant interactions by predicting mutations that improve the solubility, while retaining native state and activity. Its application to a single-domain antibody demonstrates that solubility changes upon mutation are estimated with great accuracy, thus offering a cost-effective strategy for the production of soluble proteins. The second consists in the rational design of protein-protein interactions by engineering a scaffold to bind to virtually any target disordered epitope in a protein. We validate this method by designing five single domain antibodies to bind different epitopes within three disease-related intrinsically disordered proteins (α-synculein, Aβ and IAPP). The results show that all antibodies bind to their target with good affinity and specificity. As an example of an application we carried out further experiments on one of these antibodies to show that it inhibits the aggregation of α-synuclein at low substoichiometric concentrations, and that binding indeed occurs at the selected epitope.
Speaker: Pietro Sormanni, PhD Student, Department of Chemistry, University of Cambridge, Cambridge
Room Aula room (470.03 – 4th floor)
Friday, 5th September, 2014, 11.00-12.00
Methods like phenotypic screening, next generation sequencing and high content screening have become standard in drug discovery. All these methods yield a huge amount of data, which need to be processed and analyzed in order to be able to extract biologically relevant conclusions out of them. This novel, richer, more complicated data landscape means we need state-of-the-art in silico approaches to increase the probability of a lead compound to be disease relevant. This requires data analytics approaches informing on relevant assays, compound subset design to probe the biology, visualisation of complex biological data, and target/MOA hypotheses generation.
Speaker: Elisabet Gregori Puigjané; In silico Lead Discovery Group at the Novartis Institutes of Biomedical Research
Room Xipre (seminar 173.06-183.01), PRBB.
Friday, 27th June, 2014, 11.00-12.00
Proteins are molecules involved in essentially all the complex biochemical reactions that take place in living organisms. In order to perform their functions they undergo conformational fluctuations on timescales ranging from nanoseconds to milliseconds and beyond. It is, therefore, important to develop methods capable of characterizing these motions. Nuclear magnetic resonance (NMR) spectroscopy is a powerful technique that enables the determination of the structures and dynamics of proteins at atomic resolution. Since NMR measurements produce values of observables resulting from time and ensemble averages, their interpretation is facilitated by considering ensembles of structures. The determination of an ensemble of conformations from experimental information about just average values is seem to be an ill-defined problem. In the seminar I will show that by using the Principle of Maximum Entropy (PME) it is possible to chose a special distribution (i.e., an ensemble of structures) among all those that are consistent with the experimentally-determined average values by imposing the average values themselves as thermodynamic constraints. This particular maximum entropy distribution provides an accurate representation of the unknown Boltzmann distribution of the system. The problem of determining structural ensembles can thus be solved unambiguously without making any additional assumption apart from the requirement that the experimental data should be consistent with it in the sense of the maximum entropy principle. To implement the maximum entropy principle in a computationally efficient manner, as we demonstrate in this paper in the case of NOE data, it is possible to use experimental measurements as replica-averaged structural restraints in molecular dynamics simulations.
Speaker: Andrea Cavalli, Institute for Research in Biomedicine, Department of Chemistry, University of Cambridge.
Room Xipre (seminar 173.06), PRBB.
Monday, 17th February, 2014, 11:00
Speaker: Jose Duca, Head of Computer-Aided Drug Discovery at Novartis
Room Marie Curie Room