Category: Past Talks

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).

The power spectrum of proteins at high frequencies is remarkably well described by the flat Wilson statistics. Wilson statistics therefore plays a significant role in X-ray crystallography and more recently in cryo-EM. Specifically, modern computational methods for three-dimensional map sharpening and atomic modelling of macromolecules by single particle cryo-EM are based on Wilson statistics. In this talk we provide the first rigorous mathematical derivation of Wilson statistics. The derivation pinpoints the regime of validity of Wilson statistics in terms of the size of the macromolecule. Moreover, the analysis naturally leads to generalizations of the statistics to covariance and higher order spectra. These in turn provide theoretical foundation for assumptions underlying the widespread Bayesian inference framework for three-dimensional refinement and for explaining the limitations of autocorrelation based methods in cryo-EM.

Correlated cryo-light and cryo-electron microscopy (cryo-CLEM) has become an increasingly popular method for combining the resolving power of cryo-EM with the specificity of fluorescence. Although cryo-fluorescence microscopy suffers from optical limitations, it is a powerful way to target the resolving power of cryo-EM toward proteins of interest in heterogeneous cellular environments. Two large areas of methods developments are ongoing in both hardware and software: improving the correlation precision and accuracy to facilitate single-protein targeting, and expanding the cryo-fluorescence toolkit to include room-temperature approaches capable of targeting functions or specific protein conformations. This webinar will begin with a theoretical overview of current obstacles and considerations for software developers seeking to improve correlation precision and accuracy, and conclude with our recent efforts to improve cryo-fluorescence data quality, including characterizing the effect of cryo temperatures on fluorophores and establishing function-based localization in cryo-CLEM.

Although cryo-EM has been drastically improved to generate high-resolution three-dimensional (3D) maps that contain detailed structural information about macromolecules, the computational methods for using the data to automatically build structure models are lagging far behind. The traditional cryo-EM model building approach is template-based homology modeling. Manual de novo modeling is very time-consuming when no template model is found in the database. In recent years, de novo cryo-EM modeling using machine learning (ML) and deep learning (DL) has ranked among the top-performing methods in macromolecular structure modeling. Deep-learning-based de novo cryo-EM modeling is an important application of artificial intelligence, with impressive results and great potential for the next generation of molecular biomedicine. Accordingly, I will talk about the representative ML/DL-based de novo cryo-EM modeling methods that we developed.