Paolo Missier 

RNA secondary structure: a key for molecular function

In recent years, data are being collected on long non coding RNA (lncRNA) genes, that more than double the Ensembl human gene set. Such RNAs are actively transcribed and processed in cells, but most of them are not yet associated to a known function. New tools for RNA analysis are badly needed, since sequence comparison is a fundamental tool for functional annotation, but often fails when comparing RNA sequence sharing less than 60% sequence identity. It is possible that important functional information is encoded into RNA strucure, but experimental data on RNA 3D structure is very sparse and poor. RNA secondary structure has pros and cons: it can be predicted by computational methods, and therefore can be calculated for all RNA molecules of interest, but the reliability of the prediction is not yet entirely satisfactory. We defined an alphabet for the description of RNA secondary structure, and therefore we can translate a secondary structure into a sequence of characters. Thanks to this alphabet, we could compute a substitution matrix of secondary structure elements, with the rates of variation of structural elements in functionally related RNAs. The alphabet and the matrix were used for building tools for the global and local comparison of RNAs with very low computational costs.

Computational Approaches for MDx Applications: From Academia to Industry

The development of computational approaches for predictive modeling of diseases or drug resistance predictions has opened a new era in precision medicine. Clinical decision-support-systems have been designed for assistance in molecular diagnostics (MDx) or companion diagnostics (CDx) to enhance therapeutic success. These systems are typically based on statistical or machine learning models that were build based on clinical data. The main pitfall of computational models for precision medicine developed in academia is however that most of the software is developed by individuals on a one-person-one-project basis. Thus, these researchers develop software in a prototype-centered manner, meaning that they develop software for quick publishable results and that they neither care on regulatory aspects regarding software development processes nor on the documentation and the maintenance for MDx or CDx software. One important aspect of producing reliable computational models is however the evaluation in clinical trials, which is a necessary, but not sufficient condition towards application in MDx or CDx scenarios.

Main aspects of the talk:

- Statistical and machine learning models for precision medicine

- Clinical evaluation and clinical trials

- Regulatory aspects of software for MDx and CDx applications

“Walking pathways”, cancer and evolution

If our genes are so similar, what really makes a human different fromE. coli? The answer lies in the difference in gene regulation used. In this presentation I will discuss the evolutional advantages of high plasticity of gene regulatory networks that is characteristic of multicellular eukaryotic organisms. At the same time these advantages comes with the price - terrible diseases such as cancer. Non-reversible structural changes of the regulatory networks due to an epigenetic “evolution” of genome regulatory regions provide the basis for realization of normal development programs. On the other side they may cause transformations switching the normal state to a disease state. We call such structural network changes as “walking pathways”. The analysis of this phenomenon helps us to understand the mechanisms of molecular switches (e.g. between programs of cell death and programs of cell survival) and to identify prospective drug targets of cancer. Such structural plasticity of regulatory networks observed in genomes of higher eukaryotes, in my view, is the result of an evolutional "aramorphose" towards emergence of completely new mechanism of evolution of multi-cellular organisms on the basis of combinatorial regulatory code.

Comparative analysis of deep learning architectures for multiomic inference 

The aims of the talk is to describe the design and implementation of different types of neural network architectures to perform inference on gene expression and methylation data. The models were chosen such that each of them explores different properties of the epigenetic data in order to overcome the problem of having sparse and imbalanced datasets. All of these models were subsequently compared in order to assess their pattern recognition abilities.

The talk will contain elements of tutorial so also beginners willing to practice on deep learning could benefit by bringing their laptops.

Effects of mutation and prediction of pathogenicity and phenotype 

High-throughput sequencing platforms are increasingly used to screen patients with genetic disease for pathogenic mutations, but prediction of the effects of mutations remains challenging.

We have developed SAAPdap (Single Amino Acid Polymorphism Data Analysis Pipeline) which uses a set of rule-based analyses for predicting the likely local effects of mutations. These analyses are then used with SAAPpred (Single Amino Acid Polymorphism Predictor) which uses a random forest to predict whether a mutation is likely to be pathogenic. The method gives a fully crossvalidated MCC=0.692 and a partially cross-validated MCC=0.944 (where the same protein is allowed in training and test sets, but not the same mutation). This considerably outperforms well known methods such as MutationAssessor, SIFT and PolyPhen2 (MCC between 0.452 and 0.572).

We have also extended the method to create SAAPpred-myh7 which is able to distinguish between the two major clinical phenotypes (hypertrophic cardiomyopathy, HCM and dilated cardiomyopathy, DCM) associated with mutations in the beta-myosin heavy chain (MYH7) gene product (Myosin-7). Despite having a small and unbalanced dataset, we achieve an MCC=0.53 and a post hoc removal of machine learning models that performed particularly badly, further increased the performance (MCC=0.61). Thus our method for performing the difficult task of differential phenotype prediction is competitive with other methods simply for performing pathogenicity prediction.

Preserving the currency of genomics outcomes over time through selective re-computation: model and initial findings

Complex and computationally expensive processes are common in many areas of bioinformatics, e.g. in genomics and metagenomics.
The outcomes of such processes are time-sensitive, as they depend on algorithms, tools, and reference databases which all evolve over time, often independently of one another.  This suggests that some of the processes may need to be repeatedly re-computed in response to these changes. However, these computations can be expensive, consuming tens of CPU hours each, and not all past cases will be affected by all changes. 
In the ReComp project (http://recomp.org.uk), we have started to investigate methods for optimising re-computation of common genomics processes in response to changes in the underlying reference data. We have chosen a metadata analytics approach: for each execution, e.g. of a variant calling pipeline, we record the provenance of its outcome, detailed cost (time, cloud resources), and details of the process structure, for instance a workflow, into an ever-growing history database. When changes are detected, this meta-database is analysed to determine the expected impact of a change on a population of past cases, as well as the minimal sub-workflow that needs to be re-enacted on each of them. In this talk I will present the ReComp model and our initial findings using a simple variant interpretation process, implemented on the eScience Central workflow manager, as a testbed.