AutoSmarTrace: Automated Chain Tracing and Flexibility Analysis of Biological Filaments

Abstract

Single-molecule imaging is widely used to determine statistical distributions of molecular properties. One such characteristic is the bending flexibility of biological filaments, which can be parameterized via the persistence length. Quantitative extraction of persistence length from images of individual filaments requires both the ability to trace the backbone of the chains in the images and sufficient chain statistics to accurately assess the persistence length. Chain tracing can be a tedious task, performed manually or using algorithms that require user input and/or supervision. Such interventions have the potential to introduce user-dependent bias into the chain selection and tracing. Here, we introduce a fully automated algorithm for chain tracing and determination of persistence lengths. Dubbed “AutoSmarTrace”, the algorithm is built off a neural network, trained via machine learning to identify filaments within images recorded using atomic force microscopy (AFM). We validate the performance of AutoSmarTrace on simulated images with widely varying levels of noise, demonstrating its ability to return persistence lengths in agreement with the ground truth. Persistence lengths returned from analysis of experimental images of collagen and DNA agree with previous values obtained from these images with different chain-tracing approaches. While trained on AFM-like images, the algorithm also shows promise to identify chains in other single-molecule imaging approaches, such as rotary shadowing electron microscopy and fluorescence imaging.Statement of SignificanceMachine learning presents powerful capabilities to the analysis of large data sets. Here, we apply this approach to the determination of bending flexibility – described through persistence length – from single-molecule images of biological filaments. We present AutoSmarTrace, a tool for automated tracing and analysis of chain flexibility. Built on a neural network trained via machine learning, we show that AutoSmarTrace can determine persistence lengths from AFM images of a variety of biological macromolecules including collagen and DNA. While trained on AFM-like images, the algorithm works well to identify filaments in other types of images. This technique can free researchers from tedious tracing of chains in images, removing user bias and standardizing determination of chain mechanical parameters from single-molecule conformational images.

Sequence-dependent mechanics of collagen reflect its structural and functional organization

“…Although imaging in liquid conditions is possible with AFM, here, we deposited the collagen from the specified solution conditions and dried prior to imaging. This follows previous work that used AFM to image and quantify the flexibility of collagen types I, II and III (10,31,43), to observe interactions between the NC1 domains of collagen IV (44,45), and to observe binding sites of laminin on collagen IV (46). For the conditions used in this study (>100 mM ionic strength deposition prior to drying), numerous statistical measures demonstrate the collagen to be equilibrated on the mica surface (10).…”

“…SmarTrace, a custom-built MATLAB code, was used to determine the bending flexibility of collagen chains (10). Persistence length determination by SmarTrace has been extensively validated (10,31). Chains were traced to subpixel resolution starting from the C-terminus (NC1-chain boundary for collagen IV; nonglobular end for collagen III).…”

Studying the Dynamics and Inhibition of Vimentin Assembly by Small-Angle X-ray Scattering

“…The network training is based on test data sets with structures of varying persistence lengths. The authors are confident to be able to employ their algorithm to fluorescent images [116]. Although single filaments are nicely detected, entangled, crossed or looped filaments are still difficult to trace properly.…”

“…Schneider et al recently employed neural networks and spline fitting for the detection of semiflexible polymers in AFM images [116]. The network training is based on test data sets with structures of varying persistence lengths.…”

“…A different approach is employed by Schneider et al, as connected structures are separated by removing the intersection points. The remaining parts are then detected [116]. A third approach is employed by Kapoor et al [115]: By representing the traced filament with a polynomial function direction of the individual filament is preserved, thus enabling the tracing of microtubules through intersections [115].…”