Can we see molecules in 3D without guessing?

Imagine a giant puzzle made of thousands of invisible pieces. Each piece is a molecule floating on an ultra‑thin glass sheet. Cryo‑EM captures frozen images of these pieces and rebuilds them into a 3D model. The challenge is finding the right pieces amid lots of noise. Researchers combined three key steps into one workflow. First, CryoTransformer finds pieces loosely, like a hunter marking everything that looks promising. Then MicrographCleaner removes impurities, cleaning the field. CryoSift scores each 2D class continuously, picking the best candidates. The pipeline doesn’t stop there. After selection, surviving pieces go back to the picker for fine‑tuning. This creates a feedback loop that constantly improves results. Instead of optimizing each step separately, the system is trained to maximize final 3D reconstruction quality. Tests show this integrated approach beats traditional pickers in resolution. The study also shows the best 2D F1 score doesn’t always give the best 3D map, so particle selection should be judged by the map it produces, not isolated metrics.

Cryogenic electron microscopy (cryo‑EM) captures proteins and macromolecular assemblies in their native state by rapidly freezing them, preserving structural details. The final 3D reconstruction quality hinges on extracting a clean particle stack from noisy micrographs. Traditionally, this process is split into three sub‑tasks: particle picking, contamination removal, and 2D class selection. Each sub‑task is trained and evaluated independently, without considering the downstream reconstruction goal. The new method merges these three sub‑tasks into a single reconstruction‑aware pipeline. CryoTransformer employs a transformer‑based model for permissive particle picking, flagging a broad set of potential candidates. MicrographCleaner uses convolutional neural networks to generate masks that exclude contaminants such as debris or other structures, enhancing signal‑to‑noise ratio. CryoSift introduces a continuous quality score for 2D classes, derived from an autoencoder that learns a compact, discriminative representation of particle images. After selection, surviving particles are fed back to the picker for fine‑tuning, creating an iterative feedback loop that continuously refines both picking and cleaning. Benchmark tests demonstrate that the integrated pipeline achieves higher 3D resolution than any individual component used alone. Moreover, results show that the best 2D F1 score does not always translate into the best 3D map, highlighting the need to evaluate particle selection based on the final reconstruction rather than intermediate metrics. This reconstruction‑aware approach could speed up drug discovery and deepen our understanding of cellular machinery.