Lab Newsletter — August 7, 2026: The Layer You Can't Amplify

We spend a lot of these digests on what a cell could do — its genome — and what it intends to do — its transcriptome. Today is about the layer where biology actually happens: the proteome, the working machinery, the molecules that are doing the job right now. It’s the readout closest to phenotype and, by a wide margin, the hardest to get. There’s no PCR for proteins — you can’t amplify them — so sensitivity is a brutal wall, and the raw signal, a forest of fragment-ion peaks in a mass spectrum, was nearly unreadable at scale unless the answer already sat in a reference database. That last constraint is the one deep learning just took down.

🧩 Read the protein straight from the spectrum — no database

The unlock is de novo peptide sequencing: predict a peptide’s amino-acid sequence directly from its mass spectrum, with no database to match against — the only way to find peptides that aren’t catalogued (immune peptides, environmental metaproteomes, ancient proteins, venoms, novel or mutated sequences). The field turned a corner with Casanovo (Yilmaz et al.), the first transformer built for the task: it “translates peaks in MS/MS spectra into amino acid sequences,” reading raw peaks and precursor mass with no discretization, and it’s open-source and still actively maintained (v5.0 in 2025). The 2025 flagship pushes further — InstaNovo (Eloff et al., Nature Machine Intelligence, 2025; InstaDeep + DTU) pairs a transformer with InstaNovo+, a diffusion model that iteratively refines each prediction like a second, more careful reading. The result isn’t incremental: the authors report “improved therapeutic sequencing coverage, discover novel peptides and detect unreported organisms in diverse datasets,” widening what a proteomics search can even see. Database-free identification turns the proteome from a lookup problem into a discovery one. Why it matters for the lab: this is the open-model, database-free ethos we like — a benchmarked tool that finds the thing your reference didn’t contain, shared for anyone to run.

🔬 Down to a single cell, up into a foundation model

Two frontiers are advancing at once. Down: single-cell proteomics now routinely quantifies 1,000–3,000 protein groups from one mammalian cell, prying protein-level state out of individual cells the way scRNA-seq did for transcripts — except, again, with no amplification to lean on, which makes every gain a hard-won sensitivity win. Up: the field is building foundation models. A 2025 model for tandem-MS proteomics trained on de novo sequencing “learns generalizable representations of spectra” that transfer to data-scarce downstream tasks, and a 2026 preprint, scpFormer, aims for a unified representation of single-cell proteomics. scpFormer also names a trade-off worth holding onto: antibody-based methods give broad coverage at scale, while mass spectrometry gives “substantially greater proteomic depth per cell.” Why it matters for the lab: that trade-off is our two hands. The lab works with the Human Protein Atlas — antibody-and-imaging proteomics, all breadth and spatial context; MS proteomics is the depth-per-cell complement. Reading proteins both ways — the picture and the spectrum — is how you get coverage and resolution instead of choosing.

🎯 The ladder to a virtual cell — and the check

The reason to care beyond method is where this points. A brand-new review from Tiannan Guo’s group — “AI proteomics: from protein identification to virtual cells” (Nature Methods, 2026) — draws the arc without hedging: AI in mass-spec proteomics now runs “from protein identification to building AI virtual cells,” across identification, quantification, protein–protein interactions and complexes, spatial and perturbation proteomics, and multi-omics integration — “ultimately, enabling AI virtual cells.” That’s the Human Cell Simulator’s missing molecular layer, named by the proteomics field itself. But the same discipline we keep insisting on applies in full: a de novo peptide is a prediction, and a confidently sequenced peptide that isn’t real is a false discovery. Which is why the field is building shared benchmarks like NovoBench and holding onto rigorous error control. Why it matters for the lab: it’s the same rule as hallucination-checked stains and virtual cells that show their work — an identification you can’t audit isn’t a measurement.

Read together, the arc is the one we keep meeting from new angles: a signal that used to need a reference to interpret can now be read directly, at finer and finer resolution, by a model that learned the mapping. The proteome was the omics layer that most resisted this — un-amplifiable, database-bound, spectrum-noisy — and it’s giving way. The genome is the blueprint; the proteome is the machine actually running. If a virtual cell is ever going to be judged against what a real cell does, this is the layer it will be judged on — and we’re finally learning to read it, one spectrum at a time.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.