Cryo-EM particle sorting drops the human picker: an unsupervised autoencoder puts view angle and structure in one latent space
Synopsis
Cellular cryo-electron tomograms mix many different proteins, and separating them has required template matching or experts hand-picking "good" 2D classes. This unsupervised autoencoder uses pose augmentation and a winner-take-all reconstruction loss to map particle images into a low-dimensional latent space where particles of the same structure cluster together. On a β-galactosidase dataset, 33,097 automatically selected particles reconstructed to 2.55 Å versus 2.60 Å with all particles; in real tomograms of Chlamydomonas chloroplast thylakoid membranes, the model separated four protein-density classes that conventional 2D and 3D classification did not recover.
Interpretation
Cryo-EM and cryo-ET particle images can now be grouped by structure without manual annotation or prior structures of the target proteins, with different projection views of one protein landing at neighboring positions in latent space. Earlier routes either relied on template matching, which needs an existing structure of the target and assumes the in situ structure resembles the in vitro one, or on iterative 2D/3D classification in which experts select the "good" classes, potentially introducing bias. On 97,163 particle images automatically selected from a public β-galactosidase dataset, the model separated protein images from non-protein features such as carbon edges and ice contamination; clustering then selected 33,097 "good" particles, which reconstructed to 2.55 Å versus 2.60 Å when all particles were used.
The core is an autoencoder: an encoder compresses each particle image into a latent vector and a decoder generates a denoised 2D structural template, and training does not require the model to store the particle's rotation angle internally. Earlier unsupervised models of this kind encoded in-plane pose inside the encoder, which made training convergence harder; here multiple rotated copies of each image are passed in and only the copy with the smallest reconstruction error is used for the update. Training runs in three stages: stage 1 uses images down-sampled by a factor of 2 and a coarse 30-degree grid, the later two stages use full-resolution images and a fine ±50-degree, 10-degree-step grid around the current estimate, and stage 3 adds a trainable booster decoder; classification accuracy was 98.6% on an in silico β-galactosidase/TRPV1 mixture and 96.7% on the more structurally similar β-galactosidase/PKM2 mixture.
On real in situ data, the model divided particles on Chlamydomonas chloroplast thylakoid membranes into four groups whose reconstructed 3D densities broadly match the four known photosynthesis membrane proteins. When the same projected 2D images went through cryoSPARC 2D classification and the corresponding 3D subtomograms through RELION 3D classification, both conventional approaches resolved single versus stacked membranes but did not recover the densities of the protein complexes expected on the thylakoid membranes. The experiment used the public EMPIAR-11830 dataset, with particles picked semi-automatically by an established protocol and particles on the inner and outer chloroplast membranes manually excluded; in latent space the proteins did not form clearly separable clusters, the space between them was occupied by bare-membrane and highly tilted "bad" particles, and the four-way split was made by hand from the structural features of decoded templates.
Perspective
The demonstrated scope covers particle curation on single-particle cryo-EM data, classification on in silico mixtures, and a four-way split on one real cellular tomogram, the Chlamydomonas chloroplast thylakoid membrane dataset; that last case is a relatively simple system known to contain mainly four photosynthetic membrane proteins, with particles picked semi-automatically by an established protocol and inner and outer membrane particles manually excluded.
Cluster separation in latent space can be corrupted by non-particle images that connect view-angle clusters, and bare membranes and highly tilted particles occupy the transition regions between protein clusters, reducing separability; the authors note that an extra preprocessing round could remove "bad" particles, but deciding which particles are bad is subjective and may add human bias. The method currently applies only to extracted particles and assumes one protein per particle image, an assumption that fails in systems with widely varying protein sizes; the model has also been tested only on mixtures of 2 to 4 proteins.
