deep-learning

A protein’s job depends on its address — the same molecule means one thing in the nucleus and another at the membrane. Today’s digest is about teaching machines to find that address, and it runs close to the lab’s own roots in the Human Protein Atlas. Almagro Armenteros et al.’s DeepLoc predicted localization ‘relying only on sequence information,’ using ‘a recurrent neural network … and an attention mechanism.’ Stärk et al. swapped alignments for language-model embeddings, their ’light attention’ beating the state of the art ‘by about 8 percentage points.’ DeepLoc 2.0 went multi-label with protein language models and ‘highly accurate prediction of nine different types of protein sorting signals.’ On the image side, Sullivan et al. turned the HPA Cell Atlas into a video-game mini-game — ‘322,006 gamers’ making ’nearly 33 million classifications.’ Ouyang et al. ran the HPA competition, where ‘2,172 teams’ produced models that beat the prior effort ‘by ~20%.’ And Kobayashi et al.’s cytoself learned a localization atlas ‘fully self-supervised,’ ‘from coarse classes … to the subtle localization signatures of individual protein complexes.’ Every protein has an address — and we can predict it.

AI can now design a promising molecule in seconds — but a molecule you can’t synthesize is just a picture. Today’s digest is about the other half of discovery: teaching machines to plan how to make things. Segler & Waller showed deep nets could ‘resolve reactivity conflicts and prioritize the most suitable transformation rules,’ hitting ‘95%’ top-10 retrosynthesis. Coley et al. went knowledge-free, ranking disconnections by ‘molecular similarity … without the need to encode any chemical knowledge.’ Then Segler et al.’s landmark used ‘Monte Carlo tree search and symbolic artificial intelligence’ to solve ‘almost twice as many molecules, thirty times faster.’ Schwaller et al.’s Molecular Transformer solved the forward problem as machine translation, ‘above 90%’ top-1 with calibrated uncertainty. Genheden et al.’s open-source AiZynthFinder made it a tool anyone can run ‘in less than 10 s.’ And Coley et al. closed the loop with ‘a robotically controlled experimental platform,’ making 15 drug-like compounds. Plan, predict, and press go.

An antibody binds its target through six tiny hypervariable loops — and for decades the only way to improve one was to make thousands and screen them. Today’s digest is about the deep-learning models that learned to read, write, and refine antibodies. Ruffolo et al.’s DeepAb tackled structure with ‘interpretable deep learning,’ then IgFold scaled it — ‘a pre-trained language model trained on 558 million natural antibody sequences followed by graph networks,’ predicting structures ‘of similar or better quality than alternative methods (including AlphaFold) in significantly less time (under 25 s).’ Olsen et al.’s AbLang showed that for ‘antibody specific problems … a model trained solely on antibodies may be more powerful,’ restoring missing residues better than ’the general protein language model ESM-1b.’ Shuai et al.’s IgLM turned it generative, ‘a deep generative language model for creating synthetic antibody libraries’ via ’text-infilling.’ Hie et al. proved general PLMs ‘can efficiently evolve human antibodies … despite providing the model with no information about the target antigen,’ improving affinities ‘up to 160-fold’ with ‘20 or fewer variants.’ And Mason et al. attacked the screening bottleneck, ‘predicting antigen specificity from antibody sequence via deep learning.’ Reading, writing, and refining the immune system’s keys.

Before you can compare two images, track a cell, or stack tissue slices into a 3D atlas, you have to line them up — and for years that meant a slow optimization run for every single pair. This week we read molecules; today we return to imaging’s quiet, load-bearing step. Balakrishnan et al. built VoxelMorph, reframing registration ‘as a function that maps an input image pair to a deformation field,’ computed ‘orders of magnitude faster’ than classical methods. De Vos et al.’s DLIR trained that function ‘unsupervised … by exploiting image similarity,’ since ‘obtaining example registrations is not trivial.’ Kim et al.’s CycleMorph added ‘cycle consistency’ as ‘an implicit regularization to preserve topology during the deformation.’ Chen et al.’s TransMorph brought transformers — whose ‘substantially larger receptive field enables a more precise comprehension of the spatial correspondence’ — with diffeomorphic and ‘well-calibrated registration uncertainty’ variants. Hoffmann et al.’s SynthMorph learned ‘without acquired imaging data,’ training on synthetic shapes to become contrast-agnostic. And Zeira et al.’s PASTE carried it into spatial omics, aligning tissue slices ‘using an optimal transport formulation’ to build ‘a stacked 3D alignment of a tissue.’ Lining images up, learned.

The genome isn’t a string — it’s a folded 3D object, and where a gene sits in that fold decides which enhancers can reach it. This week we read the sequence; today we watch it fold. Fudenberg et al. built Akita, ‘a convolutional neural network … that accurately predicts genome folding from DNA sequence alone,’ learning ‘an orientation-specific grammar for CTCF binding sites.’ Schwessinger et al.’s DeepC used ‘megabase-scale transfer learning’ to ‘predict the impact of both large-scale structural and single base-pair variations.’ Zhou’s Orca went multiscale — ‘from kilobase to whole-chromosome scale’ — recapitulating variant effects from ‘300 bp to 90 Mb.’ Tan et al.’s C.Origami does ‘de novo prediction of cell-type-specific chromatin organization,’ powering ‘high-throughput in silico genetic screening’ in leukemia vs normal T cells. Yang et al.’s Epiphany predicts ‘cell-type-specific Hi-C contact maps from widely available epigenomic tracks,’ with a GAN ’to encourage contact map realism.’ And GraphReg closes the loop, using ‘graph attention networks to exploit the connectivity of distal elements up to 2 Mb away’ to predict gene expression — validated by ‘CRISPRi-FlowFISH and TAP-seq.’ Decoding genome function from sequence, through structure.

The COVID vaccines made one thing obvious: mRNA is a medicine you can design — and the design problem is machine learning. Sample et al. paired ‘polysome profiling of a library of 280,000 randomized 5′ untranslated regions (UTRs) with deep learning,’ then used the model ’to engineer new 5′ UTRs that accurately direct specified levels of ribosome loading’ — extensible to ‘chemically modified RNA … for applications in mRNA therapeutics.’ Karollus et al. added ‘frame pooling, a novel neural network operation,’ to predict ribosome load ‘for 5′UTR of any length,’ and read a beta-thalassemia HBB variant. Wayment-Steele et al. attacked shelf-life: mRNA hydrolysis is beaten by designing structure to lower the ‘average unpaired probability,’ yielding ‘superfolder’ mRNAs with ‘≥two-fold’ half-life. LinearDesign faced ‘around 2.4 × 10^632 candidate mRNA sequences for the SARS-CoV-2 spike protein’ and, reframing it ‘as a lattice parsing problem,’ found an optimum ‘in just 11 minutes,’ raising ‘antibody titre by up to 128 times in mice.’ UTR-LM brought a ‘5′ UTR language model’ whose designs beat a therapeutic baseline by ‘32.5%.’ And Angenent-Mari et al. showed RNA that computes — deep nets predicting toehold-switch function at ‘R2 = 0.43–0.70’ vs ‘0.04–0.15’ for thermodynamic models. Designing the message, base by base.

You never actually read a genome. A sequencer hands you a noisy electrical trace or billions of short, error-prone reads, and an algorithm infers the DNA underneath. Increasingly that algorithm is a neural network. DeepVariant reframed variant calling as image recognition — a CNN reading ‘images of read pileups’ — and ‘outperforms existing state-of-the-art tools’ while generalizing across species. Chiron went a step upstream, the ‘first deep learning model to achieve end-to-end basecalling,’ turning raw nanopore signal straight into sequence. Clairvoyante called variants from single-molecule reads despite a ‘~5-15%’ error rate, surfacing 3,135 variants Illumina missed. PEPPER-Margin-DeepVariant pushed long reads into ‘segmental duplications and low-mappability regions where short-read-based genotyping fails,’ and DeepConsensus used a transformer to cut read errors ‘42%.’ The referee is a blind community benchmark — precisionFDA Truth Challenge V2 on Genome in a Bottle truth sets — where ‘machine learning approaches, combining multiple sequencing technologies performed particularly well.’ Open, callable, benchmarked: the foundation every omics story stands on.

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.

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.