Published digests
Search
4 results tagged AI & machine learning · Clear the tag
-
A reference genome without a virus: cDNA reconstruction reveals the provenance and function of the MS2 phage sequence
Rebuilding the RNA phage MS2 from cDNA showed that the current NCBI reference genome does not yield infectious phage, whereas sequences from published isolates do, meaning the reference reflects curation across studies rather than one validated genome. When asked to find the minimal corrections restoring infectivity, a human experimentalist prioritised the right changes while the genome language model Evo2 did not. The result is a cautionary benchmark for relying on AI genome models to interpret or repair viral sequences.
-
Divergent microbial preludes to necrotising enterocolitis defined by gut phages and bacterial resistomes.
Analysing 1,825 longitudinal stool samples from 43 preterm infants who developed necrotising enterocolitis and 86 matched controls across three hospitals, the authors combined metagenomics, metatranscriptomics and machine learning to show distinct viral and prophage trajectories before disease onset. Models integrating pre-onset prophage, resistome and bacteriome data, stratified by early versus late onset, indicate the gut virome carries predictive signal for a condition with no reliable clinical predictors. This is one of the largest attempts to turn phage data into clinical risk prediction.
-
Predicting Capsid Protein Binding Sites in Single-Stranded RNA Viruses Using Machine Learning from Local Geometric Features
The authors built a sequence-based pipeline that models RNA 3D structure, extracts local geometric descriptors of the RNA backbone and trains a neural network to predict where capsid proteins bind the viral genome, using bacteriophage Qβ as the test system. On a benchmark of experimentally verified binding and non-binding fragments it reached an AUC of 0.88 in five-fold cross-validation. This offers a cheaper computational alternative to laborious experimental mapping of genome-packaging signals.
-
LAMBDA: a prophage detection benchmark for genomic language models.
LAMBDA is a new benchmark that tests how well genomic (DNA) language models can tell phage sequences from bacterial ones, across probing, fine-tuning, diagnostic and genome-wide prophage-detection tasks. Results show that domain-specific training data matter more than sheer model size, and expose clear limits of current models on this basic annotation problem. It gives the field a concrete yardstick for AI models applied to phage genomics.