Lab Newsletter — September 28, 2026: From Blob to Blueprint

Last week we predicted protein structures from sequence and mapped the interactome. But prediction is a hypothesis; experiment is ground truth — and the great experimental engine of modern structural biology is cryo-electron microscopy, which can now image molecular machines at near-atomic detail. The catch: cryo-EM produces a 3D density map — a blurry cloud of electron density — and turning that blob into an atomic model traditionally meant months of expert handwork in 3D graphics software. Today’s digest is about the deep learning automating that whole journey, from raw micrograph to finished blueprint.

🔎 Step 1: find the particles

It starts with a needle-in-a-haystack problem. Bepler et al. (Nature Methods, 2019) tackled it with Topaz, noting that “identifying a sufficient number of particles for analysis can take months of manual effort” and that older tools “find many false positives.” Topaz is “an efficient and accurate particle-picking pipeline using neural networks trained with a general-purpose positive-unlabeled learning method,” which “retrieves many more real particles … while maintaining low false-positive rates” — even “small, non-globular and asymmetric particles.” Learning from a few sparse labels, no negatives required: the automation begins at the very first step.

🧽 Step 2: sharpen the map

The reconstructed map is noisy and loses detail at high frequencies. Sanchez-Garcia et al. (Communications Biology, 2021) built DeepEMhancer because maps “generally need to be post-processed to improve their interpretability,” and classic global-sharpening “ignore[s] the heterogeneity in the map local quality.” Trained “on a dataset of pairs of experimental maps and maps sharpened using their respective atomic models,” it learned “masking-like and sharpening-like operations in a single step” — cleaning up the map so the next steps have something crisp to read (they demonstrated it on the SARS-CoV-2 RNA polymerase).

🧭 Step 3: read the fold, even when it’s blurry

Not every map reaches atomic resolution. Maddhuri Venkata Subramaniya et al. (Nature Methods, 2019) built Emap2sec for exactly the hard middle ground, using “a three-dimensional deep convolutional neural network to assign secondary structure to each grid point in an EM map” at “resolutions of between 5 and 10 Å.” Where individual atoms aren’t visible, it still recovers the α-helices and β-sheets — “substantially better performance than existing methods” — sketching the fold from a cloud.

✏️ Step 4: trace the backbone

At good resolution, you want the chain itself, automatically. Pfab et al. (PNAS, 2021) delivered DeepTracer, “a fully automated deep learning-based method for fast de novo multichain protein complex structure determination from high-resolution cryo-EM maps.” On 476 experimental maps, “residue coverage increased by over 30% … and the rmsd value improved from 1.29 Å to 1.18 Å” versus a state-of-the-art method — and it proved its worth fast on “coronavirus-related cryo-EM maps,” modeling several with no deposited structure at all.

🏗️ Step 5: build and identify the atomic model

The capstone arrived in force. Jamali et al. (Nature, 2024, from the Scheres group at the MRC LMB) built ModelAngelo, which “combines information from the cryo-EM map with information from protein sequence and structure in a single graph neural network” to build atomic models “of similar quality to those generated by human experts.” The showstopper: by feeding “predicted amino acid probabilities for each residue in hidden Markov model sequence searches,” ModelAngelo “outperforms human experts in the identification of proteins with unknown sequences” — automating not just building the model but figuring out what protein it even is.

🧬 Step 6: don’t forget the nucleic acids

Proteins aren’t the whole story — many machines are protein–DNA/RNA complexes. Giri & Kihara (Nature Methods, 2023) filled the gap with CryoREAD, noting “computational methods for nucleic acid structure modeling are relatively scarce.” It “identifies phosphate, sugar and base positions in a cryo-EM map using deep learning, which are traced and modeled into a three-dimensional structure,” building “substantially more accurate models than existing methods” from 2.0 to 5.0 Å — again validated on SARS-CoV-2 complexes.

🧫 Why it’s our kind of problem

Read across the six and it’s AI closing the loop on experimental structure — the complement to AlphaFold’s prediction. Prediction offers a hypothesis; map interpretation delivers the measured truth, and now does it faster and more objectively (ModelAngelo even beats experts at identifying proteins). The deeper pattern is the lab’s own thesis: turn expert-and-labour bottlenecks into learned, automated, reusable tools — the same bet behind BioEngine and the BioImage Model Zoo, here aimed at structural biology’s most tedious step. And structure is load-bearing for everything downstream the lab tracks — function, interactions, drug binding, and the mechanistic layer of the virtual cell. Cryo-EM gave us the blob; AI is learning to hand back the blueprint.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via the Europe PMC API.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.