Murto's Lab

Picking the surface protein a rapid antigen test for Cyclospora cayetanensis should target. Public assemblies are QC'd and lifted onto one annotation, DeepLoc 2.1 finds the proteins that reach the surface, and OrthoFinder borrows PaxDB abundances from relatives to rank what's left.

Classifies CATH architecture straight from structure. Each domain is rendered as an RGB contact map — red for inverted Cβ–Cβ distance, green for helix–helix pairs, blue for strand–strand pairs — and read by a dilated CoordConv residual CNN. Benchmarked out-of-distribution by vaulting whole superfamilies out of training, the way DockGen and Chainsaw do.

Identified 71 natural alternatives to a $200 million drug entirely in silico. Screened the COCONUT natural-products database by chemical similarity to acarbose, then docked the survivors with GNINA — an AutoDock Vina fork that scores poses with a CNN — and ranked them on CNN-VS. Decoys, redocking replicates, and stability analysis stripped out the false positives.

Assembling and comparing four different ML/DL models on identifying peptide-protein binding sites. Built alongside a custom ML-friendly dataset extracted and parsed directly from the PDB.

Utilized over 47,000 SLURM jobs to generate AlphaFold-Multimer predictions for wild-type and mutant protein complexes. Structural features were extracted to train an XGBoost classifier, achieving a 91% AUC for predicting PPI disruption.

Finds variants of known effect at protein–protein interfaces to use as in silico controls. Takes pathogenic and benign variant tables from ClinVar, keeps the ones PIONEER places at PPI-encoding sequences, and outputs a JSON of positive and negative controls. Built to test AlphaFold-Multimer at SickKids.

Compares wild-type protein complexes predicted with AlphaFold-Multimer against the same complexes with missense mutations introduced. Its scripts measure Cα and Cβ distances across the interface, find consensus interface residues, score Jaccard similarity between wild-type and missense interfaces, average PI-score features, and run a U-test between the two.

Pulls every hyperlink and its anchor text from a page, scores each linked domain's authority with the Moz API, and reports the scores and their average. Built in under a day with GPT-4o, to automate a fact-checking step from editing work at a marketing agency.

Idiom illustrations drawn during the pandemic for articles on Medium, released free to download and use. My first launch on Product Hunt.