Category: Past Talks

Laser phase-contrast Cryo-EM and associated computational opportunities

A phase plate can provide optimum image contrast for weak-phase objects in transmission  electron microscopy, but approaches toward realizing a phase plate have suffered from  instabilities. We have developed a phase plate that is based on coherently phase-shifting the  electron wave function by a laser beam, which is built up to a record-high intensity of ~400  GW/cm^2 by resonance in a Fabry-Perot cavity. We have demonstrated contrast enhancement  with the laser phase plate (LPP) and shown the long-term stability of the device, as well as  generated a high-resolution map of 20S proteasome particles using a standard single-particle  cryo-electron microscopy (Cryo-EM) workflow. 

This talk will focus on our recent work to move beyond proof-of-concept, as well as the  computational opportunities that lie ahead. To demonstrate the benefits of the LPP to Cryo-EM  as well as cryo-electron tomography, we will soon begin working with a state-of-the-art  microscope equipped with a spherical aberration corrector, gun monochromator, and post column energy filter. We will explore improvements to the phase plate design and pursue new  strategies for image acquisition and processing, such as high-resolution two-dimensional  template matching.

Recovering small molecular structures using Cryo-EM

Any current Cryo-EM algorithmic pipeline entails recovering the 3-D structure after particle picking. However, the signal-to-noise ratio of the data, and thus the reliability of particle picking, drops with the molecular mass of the specimens. Accordingly, it is commonly believed that Cryo-EM cannot be used to map molecules with a molecular mass below a certain threshold. Challenging this misconception, I will argue that finding the particle picking is not a prerequisite for structure determination and thus small molecules are, at least in principle, within reach of Cryo-EM. Then, I will introduce two computational frameworks to bypass particle picking and show numerical results.

Bias, variance and map validation in Single Particle Analysis by CryoEM

Single Particle Analysis by CryoEM has established as a mature technique to elucidate the three-dimensional structure of biological macromolecules. Once the experimental data is acquired, reconstructing the macromolecule’s map involves the estimation of millions of parameters. Due to the low Signal-to-Noise Ratio, this estimation is prone to mistakes. Depending on the error size, these mistakes may result into small perturbations in the reconstructed map (variance) or large artifacts (bias). In this talk we will discuss different kinds of mistakes that can be committed and how to identify and tackle them.

Revisiting Orientation Estimation in Cryo-EM Volume Refinement

In Cryo-EM 3D map refinement, popular software packages jointly reconstruct the 3D map while estimating orientations. The orientation estimation can roughly be categorised in two type of approaches: marginalisation and optimisation. While the former tends to be more robust to noise, the latter has better consistency with respect to the data. So far it appears difficult to obtain both data-consistency and noise-robustness in a single method.

In this talk we will revisit the orientation estimation process. In particular, we develop an alternative that can be interpreted as being “in between marginalisation and optimisation” and argue that this new method is both robust to noise and data-consistent. Additionally, the framework of lifting-based global optimisation on manifolds allows analysis of the proposed methods that lead to several practical theoretical guarantees. Both theoretical results and performance on simulated data will be tested in numerical experiments.

Validating and Building Protein Structure Models for cryo-EM Maps Using Deep Learning

Cryo-electron microscopy (cryo-EM) has become one of the main experimental methods for determining protein structures. Protein structure modeling from cryo-EM is in general more difficult than X-ray crystallography since the resolution of maps is often not high enough to specify atom positions. We have been developing a series of computational methods for modeling protein structures from cryo-EM maps. For maps at medium resolution, deep learning can provide useful structure information for structure validation and modeling. Particularly, we have recently developed a protein model quality assessment score, DAQ, which compares local density patterns captured by deep learning with amino acid positions in a model, and detect potential errors in the model. In a large-scale analysis of protein models from cryo-EM, we found that not a small number of models may have some errors. All the tools we developed are available at https://kiharalab.org/emsuites/.

Machine learning structure prediction and cryoEM map interpretation

The machine-learning structure prediction tools AlphaFold and RoseTTAfold have dramatically simplified low-resolution cryoEM map interpretation.  Following this, we describe three tools our lab has developed in conjunction with these approaches.  First, we describe novel machine-learning methodology for detecting errors in models built against low-resolution cryoEM density.  Next, we describe tools for modelling ligands into low-resolution cryoEM density.  Finally, we describe our efforts at adding to RoseTTAFold the ability to predict the structure of protein/nucleic-acid complexes.

Routes to Extract Biological Function from single-particle cryoEM 

Recent progress in data analysis of single-particle cryogenic Electron Microscopy (cryoEM); such as Geometric Machine Learning algorithms, allows one to retrieve the conformational spectrum of heterogeneous molecular ensembles.  Until recently, most efforts have been attempted under equilibrium conditions [1,2], where the conformational spectrum is time-independent, and thus directly yields the free-energy landscape. However, most biological systems operate far from equilibrium to sustain the processes that constitute life.

Under nonequilibrium conditions, the functional pathways are time-dependent, and the evolution of the conformational spectrum can be described by a Fokker-Planck equation, however with an unknown operator. State-of-the-art developments in Machine Learning, the so-called Physics-Informed Neural Networks (PINN) [3], allows one to retrieve the underlying Fokker-Planck operator from sparse observations alone [4], providing a complete physics-based description of the nonequilibrium process. Moreover, time-resolved cryoEM experiments have recently become practical [5]. Together, this enables us to combine the advantages of time-resolved serial crystallography (‘nonequilibrium processes’) with the advantages of single-particle methods (‘avoids averaging over unlike particles’).

Based on these opportunities, we present a conceptional and algorithmic framework to extract functional pathways from nonequilibrium from a collection of time-resolved single-particle images. This constitutes an unexplored route in studying biological function and structural dynamics under nonequilibrium conditions.

References

1. Dashti, A. et al. Trajectories of the ribosome as a Brownian nanomachine. Proc. Natl. Acad.  Sci. U. S. A. 111, 17492–7 (2014).

2. Dashti, A. et al. Retrieving functional pathways of biomolecules from single-particle snapshots. Nat. Commun. 11, 4734 (2020).

3. Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, (2019).

4. Chen, X., Yang, L., Duan, J. & Karniadakis, G. E. Solving inverse stochastic problems from discrete particle observations using the Fokker-Planck equation and physics-informed neural networks. SIAM Journal on Scientific Computing 43, (2021).

5. Dandey, V. et al. Time-resolved cryo-EM using Spotiton, Nature Meth., 17, 897 (2020).