Posts

Metabolites are the molecules closest to what a cell is actually doing — ‘a direct functional signature of cellular state.’ Untargeted mass spectrometry can see thousands of them in a sample, but here’s the uncomfortable secret: ’the vast majority of metabolites remain unknown.’ This is the dark metabolome, and machine learning is finally lighting it up. CSI:FingerID turns a fragmentation spectrum into a predicted molecular fingerprint and searches it against databases like PubChem; SIRIUS 4 packaged that into a fast tool hitting ‘identification rates of more than 70%.’ When a compound is in no database, CANOPUS uses a deep network to at least name its chemical class — 2,497 of them, even for molecules ‘for which neither spectral nor structural reference data are available.’ MSNovelist goes further and generates structures de novo from the spectrum alone, ‘without having ever seen the structure in the training phase.’ Spec2Vec brings word2vec to spectra for better similarity, and GNPS makes it a shared, open ’living data’ commons. It’s the metabolome — the omics layer a virtual cell can’t do without.

AICell Lab warmly welcomes Florence Stadelmann, an M.Sc. Molecular Biotechnology student from Uppsala University joining us as a research intern to operate and validate our automated live-cell imaging platform.

A single cell speaks several molecular languages at once — which genes its chromatin lets it read, which RNAs it’s transcribing, which proteins are actually doing the work. New assays measure two or three of these in the same cell, and the AI job is fusion: ’tying together data across different modalities’ that, as GLUE puts it, ’typically have distinct feature spaces.’ Seurat’s weighted-nearest-neighbor learns ’the relative utility of each data type in each cell’ across a 211,000-cell CITE-seq atlas, looking ‘beyond the transcriptome toward a unified and multimodal definition of cellular identity.’ totalVI models RNA and protein jointly as ‘a composite of biological and technical factors’; MultiVI extends the idea to chromatin and even to ‘cells for which one or more modalities are missing.’ GLUE fuses unpaired datasets with a regulatory graph, scaling to a ‘human cell atlas construction over millions of cells.’ And the scIB benchmark — 68 methods, 85 batches, 1.2 million cells — keeps everyone honest about the real tension: removing batch effects ‘while retaining biological variation.’ It’s the multi-layer definition of a cell a virtual cell still needs.

AlphaFold and ESMFold gave us shape at genome scale — the ESM Metagenomic Atlas alone folds ‘more than 617 million metagenomic protein sequences.’ But a fold is not a job. Today’s digest is about the other half of the problem: reading a protein’s function from its sequence or structure. DeepGOPlus predicts GO terms ‘from sequence alone,’ fast enough to ‘annotate around 40 protein sequences per second’; ProteInfer runs a CNN that predicts EC numbers and GO terms ‘directly from an unaligned amino acid sequence,’ even in-browser with ’no data uploaded to remote servers.’ For enzymes, CLEAN uses contrastive learning to assign EC numbers ‘with better accuracy, reliability, and sensitivity compared with the state-of-the-art tool BLASTp’ — and to correct mislabeled ones. DeepFRI reads function off structure with a graph network, down to ‘site-specific annotations at the residue-level.’ And CAFA — the community benchmark that drove ’new functional annotations for more than 1000 genes’ — keeps everyone honest, admitting some categories still haven’t improved. It’s the parts list a virtual cell needs: not just what each protein looks like, but what it does.

Three weeks ago we followed the RNA map of a tissue; today it’s the protein map. Multiplexed imaging like CODEX stains one slice for dozens of proteins at once — ‘antibody binding events’ read out ‘at a single-cell and cellular neighborhood’ level — turning a piece of tissue into a ~40-channel image where every pixel carries a molecular fingerprint. The AI job is a three-step relay: find every cell (Mesmer, trained on TissueNet’s ‘more than 1 million manually labeled cells,’ hit human-level segmentation across tissue types); name every cell from its protein profile (STELLAR, ‘a geometric deep learning method for cell-type discovery,’ applied straight to CODEX data); then read the neighborhoods. And the honest wall: a 2024 Nature Methods benchmark of ‘more than 50 diverse biological experiments’ found methods ‘often tailored to specific modalities or require manual interventions’ — generalization, not accuracy on one panel, is the frontier. It’s the imaging×omics intersection at the lab’s core, and the protein half of the spatial map a virtual cell still needs.

A cryo-EM dataset isn’t one picture of a protein — it’s millions of noisy 2D snapshots, each a different molecule flash-frozen mid-wiggle in a slightly different shape. For decades the job was to average them into a single sharp structure, which meant averaging the motion away. The AI turn flips that: treat the heterogeneity as the signal. cryoDRGN used neural networks to ‘directly reconstruct continuous distributions of 3D density maps,’ even ’large-scale continuous motions of a spliceosome complex.’ 3DFlex modeled the motion itself as a physical deformation; DynaMight (open-source, in RELION-5) learns per-particle deformations — while candidly warning that atom-model priors ‘may lead to important artifacts due to model bias.’ And CryoBench names the reckoning underneath it all: without standardized ground truth, how do you know a recovered motion is real and not the algorithm’s imagination? It’s the experimental twin of the predicted-motion story we told on Aug 11 — two roads to the same conformational landscape a virtual cell will have to get right.

What if you could just ask your microscope a question? That’s the promise of vision-language models — systems that connect a picture to words. BiomedCLIP learned the mapping from 15 million image-text pairs mined from 4.4 million scientific articles; LLaVA-Med turned figure captions into ‘a vision-language conversational assistant that can answer open-ended research questions of biomedical images,’ trained ‘in less than 15 hours.’ Then the reality check: on real microscopy, μ-Bench finds ‘current models struggle on all categories, even for basic tasks such as distinguishing microscopy modalities,’ and specialist fine-tuning can make things worse. CARES adds that medical vision-language models ‘consistently exhibit concerns regarding trustworthiness, often displaying factual inaccuracies.’ The gap between a fluent answer and a correct one is the whole story — and it’s exactly why the lab’s answer is a grounded, tool-calling assistant like the BioImage.IO Chatbot, not a model asked to know everything on its own.

A single-cell sequencing run is a still photograph — every cell frozen at the instant it was lysed. So how do you recover motion from a photo? RNA velocity’s answer, from Sten Linnarsson’s group at Karolinska and colleagues, is beautiful: count a gene’s unspliced (new) versus spliced (mature) transcripts, and the imbalance becomes ’the time derivative of the gene expression state’ — a vector that ‘predicts the future state of individual cells on a timescale of hours.’ scVelo dropped the brittle steady-state assumptions; CellRank turned velocity into probabilistic fate maps; and the 2024 deep-learning turn — veloVI, CellRank 2 — added the one thing the method most needed: a way to know when to trust its own arrows. It’s the molecular twin of yesterday’s imaging-based cell tracking — same question, where is this cell going next, from the opposite kind of data — and it’s exactly the cell-state dynamics a virtual cell would have to reproduce.

Segmentation finds the cells in one frame. The harder job is following each one through the next thousand — as it crawls, brightens, and above all divides, one cell becoming two. That temporal-linking problem is cell tracking, and it has its own quiet history. The classical answer was global optimization: the KTH tracker of Magnusson and Jaldén used the Viterbi algorithm to link whole tracks across an entire movie at once, handling mitosis, apoptosis, and cells wandering out of view. The Cell Tracking Challenge made progress objective — 21 algorithms on 13 datasets, with the honest verdict that ’none obtained fully correct solutions.’ Then deep learning moved in: TrackMate 7 wired modern segmentation into an open Fiji pipeline; DeepCell’s Caliban learned tracking-with-lineage from human-in-the-loop-labeled movies; and Trackastra replaced hand-tuned linking with a transformer that learns cell-to-cell associations and even handles division. It’s the temporal spine of every live-cell experiment the lab runs — and a field KTH helped build. But a track is a hypothesis about a cell’s identity over time: errors compound frame by frame, and the benchmark still isn’t solved.

CRISPR made cutting the genome easy; the hard part was knowing what would happen next. Which guide will actually cut? And once the cell repairs the break, what sequence will you be left with — long dismissed as random? Machine learning answered both. Guide-activity models like DeepSpCas9 and DeepCpf1 predict which guides work from sequence alone. Then the surprise: inDelphi showed template-free repair is ‘predictable and capable of precise repair,’ hitting a single dominant genotype at 5–11% of sites; FORECasT confirmed outcomes ‘are not random, but depend on DNA sequence’ across a billion measured events. The idea generalized — BE-Hive predicts base-editing outcomes at R≈0.9, PRIDICT predicts prime-editing efficiency and even flags the pegRNAs worth trying. It’s the design half of the perturbation the virtual cell tries to predict the response to, and the design brain of the design→edit→measure loop a self-driving lab runs. The honest frontier is our native language: benchmarks are still thin, outcomes are cell-line-dependent, and a predicted genotype stays a hypothesis until sequencing agrees.