Master Systems Biology selected students presentations
Tuesday 07 July
Dr. Julia Massimelli Sewall | Director of master's programmes Sciences
||Predicting the effect of single-point missense variants in the binding site on protein-ligand binding affinities: A machine learning approach
Ammar Ammar | Systems Biology student
||The visual cortex as a network of phase-oscillators
Kris Evers | Systems Biology student
||Assessing outcomes and risk factors of atrial fibrillation through a lifetime population-level markov model
Cristian Barrios Espinosa | Systems Biology student
||Towards understanding complex behaviour in goal-oriented systems
Raphael Stolpe | Systems Biology student
Predicting the effect of single-point missense variants in the binding site on protein-ligand binding affinities: A machine learning approach
A key concept in drug design is how natural variants, especially the ones occurring in the binding sites of drug targets, affect the inter-individual drug responses and efficacies by altering binding affinities. These effects have been reported in literature on very limited and small datasets that are not suitable for machine learning prediction models. Ideally, a large dataset of binding affinity changes due to binding site single-nucleotide polymorphisms (SNPs) is needed to build a machine learning (ML) model predicting these effects.
However, to the best of our knowledge, such a dataset did not exist yet. To solve this, a reference dataset of ligands binding affinities to proteins with all their reported binding site variants was constructed using a molecular docking approach. A numerical vector representation of protein, binding pocket, mutation, and ligand information was encoded using a total of 256 extracted features to describe the protein-ligand pair. Using this dataset, two designs of machine learning regression models were trained and evaluated on the chosen features to predict the binding affinity, and six different scenarios to split training and test data were evaluated. The models trained on datasets based on ligand molecular weight split reported the best performance. The best performance model reported an RMSE value of 0.57 kcal/mol-1 on an independent test set with an R-squared value of 0.86 and a prediction speed of 29 bound compounds affinities per second. We report an improvement in the protein-ligand binding affinity prediction performance of the ML models compared to several published models. The obtained models can be used in early-stage drug discovery to rapidly and accurately obtain a better overview of the ligand binding affinity variability across genetic variants. This may benefit applicability of medicine across ethnic groups as well as in personalized medicine.
The Visual Cortex as a Network of Phase-Oscillators
The human brain is capable of extracting information from visual scenes. Mathematical models of the visual system can be used to link biological structure to function. Neurons in the early visual cortex are sensitive to oriented stimuli, and local synaptic connections are biased towards neurons with collinear receptive fields. This suggests a role for lateral connections in the early visual cortex in contour integration. Human psychophysical experiments (Field et al., 1993) suggest that a local association field is expressed by the visual system. This local association field suggests a relatively high association between collinear and proximal oriented stimuli.
The higher association between collinear stimuli allows for grouping of smooth contours. In this project the local association field is implemented in a phase-oscillator network model assuming binding-by-synchrony. The phase-oscillators represent neuron populations of the early visual cortex, the lateral connections between the oscillators are based on the local association field theory, and grouping of contours is assumed to take place through synchrony between phase-oscillators. The frequency- and phase code can be used by the brain to interpret a visual scene. Experiments are performed by simulating the phase update in the Kuramoto model. The first experiment confirms that the model is capable of performing contour integration through synchrony. New experiments are designed to generate testable predictions and hypotheses. These predictions, hypotheses and future experiments and extensions of the model are discussed.
Assessing Outcomes and Risk Factors Of Atrial Fibrillation Through A Lifetime Population-Level Markov Model
Cristian Alberto Barrios Espinosa
Atrial fibrillation (AF) is characterized by abnormal electrical and mechanical activity of the atria and presents a major global health burden. Computational models are increasingly being used to study AF and guide its treatment. However, current computational models and epidemiological studies are unable to bridge the gap of knowledge between mechanisms of AF pathophysiology and long-term clinical outcomes.
Aim: To develop a novel computational model to bridge this knowledge gap by simulating a virtual population of patients over a lifetime period. Methods: A Markov model with 5 states for combinations of AF, sleep-disordered breathing (SDB) and death was developed. The stochastic transitions between states are controlled by clinical risk factors (age, sex, SDB) and mechanisms of AF pathophysiology (shortening of effective refractory period and slowing of conduction velocity). Results: The model was calibrated to reproduce real-world data with respect to: 1) epidemiology of AF and SDB, 2) mortality in the general population and in AF patients, 3) patterns of AF paroxysms, and 4) relation between clinical subtypes of AF and atrial fibrosis. Finally, the calibrated model was used to analyze the potential therapeutic effects on AF of different types of risk-factor management for SDB. Conclusion: We successfully developed a model that connects mechanistic and epidemiological research in AF and show how this model can be used to perform a ‘virtual clinical trial’ to generate new hypotheses that can be used to guide new clinical research and ultimately lead to better AF management. Impact: This novel patient-level model will be used to study the effect of new policies on the burden caused by AF.
Towards understanding complex behaviour in goal-oriented systems
Understanding complex, goal-oriented behaviour remains an ongoing challenge in neuroscience. Meaningful progress has been made through intelligently designed experiments. Visual and sensory information processing as well as action generation were found to be a function of the frontoparietal network spanning the visual, sensory and motor cortices. Computational neuroscientists successfully devised models to explain neural phenomena encountered in the frontoparietal network. However, unlike other natural sciences computational neuroscience must not only account for neural phenomena, but also address information processing in the brain. Therefore, an ideal model must be capable of generating complex, goal-oriented behaviour while also accounting for neural phenomena. Dynamical systems have been widely applied to model neural dynamics, but also to generate control policies in robotics. This work proposes to leverage computing with dynamical systems to understand neural dynamics and complex behaviour expressed as control policies in goal-oriented systems.
To that end, this work utilises recent advances in deep reinforcement learning. Agents are trained to solve a variety of reinforcement learning problems including control of an anthropomorphic hand. Afterwards, dynamics recorded while the agent performed the task are critically assessed. Attractor networks and central pattern generators are found as general computational strategies to solve for final state goals and periodic tasks (e.g. walking), respectively. These findings demonstrate the capability of computing with dynamical systems to account for neural dynamics and information processing. Computing with dynamical systems when applied to more complex tasks could pave the way to generating ever more sophisticated hypotheses about complex behaviour in artificial and biological systems.