newsletter

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.

For a week we’ve modeled molecules one at a time — designed, evolved, force-fielded, screened. Today the lens widens: biology is a web of relationships, and machine learning on that web predicts the connections we haven’t drawn. Himmelstein et al. fused biomedical knowledge into Hetionet — ‘47,031 nodes of 11 types and 2,250,197 relationships of 24 types’ — and scored ‘209,168 compound-disease pairs’ for repurposing, ’entirely open.’ Zitnik et al.’s Decagon brought graph neural networks: ‘a new graph convolutional neural network for multirelational link prediction,’ predicting ’the exact side effect’ of a drug pair and ‘outperforming baselines by up to 69%.’ When COVID hit, a network-medicine consensus ranked 6,340 drugs and screened the top ones at a ‘62% success rate, in contrast to the 0.8% hit rate of nonguided screenings’ — and ‘76 of the 77’ hits act through mechanisms ’that cannot be identified using docking-based strategies.’ PrimeKG released an open precision-medicine graph of ‘17,080 diseases with 4,050,249 relationships,’ and TxGNN turned it into ‘a graph foundation model for zero-shot drug repurposing,’ finding candidates ’even for diseases with … no existing drugs,’ with ‘multi-hop’ explanations. As Li, Huang & Zitnik put it, ‘graphs are universal descriptors of systems of interacting elements.’ Reasoning over the web of biology.

Antibiotic resistance is outrunning the discovery pipeline — so researchers taught machines to look where humans can’t. Stokes et al. ’trained a deep neural network capable of predicting molecules with antibacterial activity’ and surfaced halicin, ‘structurally divergent from conventional antibiotics.’ Liu et al. screened ~7,500 molecules and found abaucin, with ’narrow-spectrum activity against A. baumannii,’ a priority Gram-negative pathogen. Wong et al. made the black box talk — graph neural networks predicting activity and cytotoxicity for ‘12,076,365 compounds’ and yielding a new structural class ‘selective against methicillin-resistant S. aureus.’ Das et al. designed antimicrobials outright with ‘deep generative models and molecular dynamics simulations.’ And two teams went mining nature’s own arsenal: 181 of 216 gut-microbiome peptides proved active, and AMPSphere catalogued ‘863,498 non-redundant peptides’ from the global microbiome — an open resource. AI widens where we look for medicine.

To simulate a molecule you need the forces on every atom. The gold standard, quantum mechanics, is ‘computationally demanding … making long simulations of large systems unfeasible.’ The workaround is a neural network that learns the potential-energy surface itself. Behler & Parrinello introduced one ‘several orders of magnitude faster than DFT’; ANI-1 delivered ‘DFT accuracy at force field computational cost,’ transferable across organic chemical space; SchNet showed deep nets are ‘ideally suitable for representing quantum-mechanical interactions.’ DeePMD made it scale — quantum-accurate molecular dynamics ‘at a cost that scales linearly with system size’ — and NequIP baked in the symmetries of 3D space, ‘challenging the widely held belief that deep neural networks require massive training sets’ by learning from ‘up to three orders of magnitude fewer training data.’ As Unke et al. put it, the aim is ’to narrow the gap between the accuracy of ab initio methods and the efficiency of classical FFs.’ It’s the physics engine a virtual cell will run on.

Yesterday we read the genome; last Friday we designed proteins from scratch. Today’s verb is evolve — using machine learning to improve the proteins nature already gave us. Directed evolution won a Nobel; adding ML lets it ‘predict how sequence maps to function in a data-driven manner without requiring a detailed model of the underlying physics.’ DeepSequence showed you can score mutations ‘in an unsupervised manner solely on the basis of sequence information,’ beating prior methods across deep mutational scans. UniRep distilled proteins into a representation that’s ‘structurally, evolutionarily and biophysically grounded,’ buying ’two orders of magnitude efficiency improvement.’ Low-N built ‘an accurate virtual fitness landscape’ from ‘as few as 24 functionally assayed mutant sequences’ and screened ten million in silico. Hsu et al. showed a plain ridge regression ‘is competitive with, and on average outperforms more sophisticated methods’ — and preached ’the importance of systematic evaluations and sufficient baselines.’ And at the bench, ML-guided evolution engineered a new-to-nature enzyme to ‘93% and 79% ee.’ Design–build–test–learn, made real.

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.

A cell’s identity isn’t a list of active genes — it’s a circuit. Transcription factors switch each other on and off in tangled feedback loops, and reconstructing that wiring from data is, as one landmark put it, ‘one of the pressing open problems of computational systems biology.’ Gene regulatory network inference is how AI reads the circuit. GENIE3 turned it into a stack of regression problems — feature importance as a regulatory link. DREAM5’s blind test of 30+ methods delivered a humbling verdict — ’no single inference method performs optimally’ — and a fix: ensembles, ‘wisdom of crowds.’ SCENIC brought it to single cells, reconstructing networks and cell states at once; BEELINE benchmarked the field and found accuracy only ‘moderate,’ with pseudotime-free methods ahead. Then the payoff: CellOracle uses inferred networks ’to perform in silico transcription factor perturbations … using only unperturbed wild-type data’ — knock out a gene on a laptop, predict the phenotype, validate at the bench. SCENIC+ adds chromatin for enhancer-driven wiring. It’s the causal layer a virtual cell can’t do without.

For most of biology, proteins were things you discovered. Now they’re things you can write. De novo protein design flips the deep-learning revolution around: instead of predicting the structure of a protein that already exists, it invents proteins that never did. It started with physics — Top7, a ’novel sequence and topology’ whose crystal structure matched the design to 1.2 Å, proving we could ’explore the large regions of the protein universe not yet observed in nature.’ Then deep learning: ‘hallucination’ inverts a structure predictor to dream new folds (27 of 129 designs folded in the lab); ProteinMPNN writes the sequence for a given shape with ‘52.4% sequence recovery compared with 32.9% for Rosetta’; RFdiffusion borrows image-generation diffusion to design binders and enzymes ‘from simple molecular specifications,’ with a cryo-EM-confirmed influenza binder ’nearly identical to the design model’; and Chroma treats design ‘as Bayesian inference under external constraints’ — steered by shape, symmetry, even ’natural-language prompts.’ The predictor that made it possible, AlphaFold, is also the referee that validates it. Prediction and design: two directions of one map.

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.