The amazing results of DeepMind's AlphaFold2 in the last CASP experiment  caused a huge stir in both the AI and biology fields, and this was of 
course widely reported in the general media. The claim is that the  protein folding problem has finally been solved, but has it really? Not 
to spoil the ending, but of course not. In this talk I will not be  talking (much) about AlphaFold2 itself, but instead what inspiration we 
can take from it about future directions we might want to take in protein structure bioinformatics research using modern AI techniques. 
Along the way, I'll illustrate my thoughts with some recent and current  machine-learning-based projects from my own lab in the area of protein 
structure and folding.
 

The last years have seen an immense increase in high-throughput and high-resolution technologies for experimental observation as well as
high-performance techniques to simulate molecular systems at a microscopic level, resulting in vast and ever-increasing amounts of high-dimensional data.
However, experiments provide only a partial view of macromolecular processes and are limited in their temporal and spatial resolution. On the other hand,
atomistic simulations are still not able to sample the conformation space of large complexes, thus leaving significant gaps in our ability to study
molecular processes at a biologically relevant scale. We present our efforts to bridge these gaps, by exploiting the available data and using state-of-the-art
machine-learning methods to design optimal coarse models for complex macromolecular systems. We show that it is possible to define simplified
molecular models to reproduce the essential information contained both in microscopic simulation and experimental measurements.

Geometrical questions commonly arise in clinical practice: for example, what is the optimal shape for a particular medical device? or what shapes of anatomical structures are indicative of pathological events? In this talk we explore two disparate clinical applications of geometrical underpinning: (A) how to design the optimal device for kidney stone removal surgery? and (B) what blood vessel shapes are associated with biomechanical failure? (A) Flexible ureteroscopy is a minimally invasive treatment for the removal of kidney stones by irrigating dust-like stone fragments with a saline solution. Finding the optimal ureteroscope tip shape for efficient flushing of stone fragments is a pertinent but complex question. We represent the renal pelvis (the main hollow cavity within the kidney) as a 2D cavity and employ adjoint-based shape optimisation to identify tip geometries that shrink the size of recirculation zones thereby reducing stone washout times. (B) The aorta is the largest blood vessel in the body, with an archetypal arched “candy-cane” shape and is responsible for transporting blood from the heart to the rest of the body. Aortic dissection, in which the inner layer of the aorta tears, can lead to frank rupture and is often rapidly fatal. Accurate clinical assessment of dissection risk from a CT scan of a patient’s thorax is paramount to patient survival. We apply statistical shape analysis, coupled with hemodynamic simulations, to identify pathological shape features of the aortic arch and to elucidate mechanistic underpinnings of aortic dissection.

The circadian clock generates ~24h rhythms everyday via a transcriptional-translational negative feedback loop. Although this involves the daily entry of repressor molecules into the nucleus after random diffusion through a crowded cytoplasm, the period remains extremely consistent. In this talk, I will describe how we identified a key molecular mechanism for such robustness of the circadian clock against spatio-temporal noise by analyzing spatio-temporal timeseries data of clock molecules. Furthermore, I will illustrate a systemic modeling approach that can identify hidden molecular interactions from oscillatory timeseries with an example of a circadian clock and tumorigenesis system.  Finally, I will talk about a fundamental question underlying the model-based time-series analysis: “Can we always fit a model to given timeseries data as long as the number of parameters is large?”. That is, is Von Neumann's quote “With four parameters I can fit an elephant, and with five I can make him wiggle his trunk” true?

Riemannian optimization is a powerful and active area of research that studies the optimization of functions defined on manifolds with structure. A class of functions of interest is the set of geodesically convex functions, which are functions that are convex when restricted to every geodesic. In this talk, we will present an accelerated first-order method, nearly achieving the same rates as accelerated gradient descent in the Euclidean space, for the optimization of smooth and g-convex or strongly g-convex functions defined on the hyperbolic space or a subset of the sphere. We will talk about accelerated optimization of another non-convex problem, defined in the Euclidean space, that we solve as a proxy. Additionally, for any Riemannian manifold of bounded sectional curvature, we will present reductions from optimization methods for smooth and g-convex functions to methods for smooth and strongly g-convex functions and vice versa.

This talk is based on the paper https://arxiv.org/abs/2012.03618.

-

A link for this talk will be sent to our mailing list a day or two in advance.  If you are not on the list and wish to be sent a link, please contact @email.