Software

Cancer genomes carry many inactivating events at once, raising the possibility of higher-order vulnerabilities: a gene that becomes essential only after two others are lost. We screened the cancer dependency map for such third-gene targets by defining the cell lines that have lost a pair of tumor suppressors and asking which gene each background can no longer spare. The screen recovers textbook biology, the PRMT5 axis in CDKN2A and MTAP-deleted cells and the E2F machinery in TP53 and RB1-deleted cells, and nominates targets in real cancers, including VPS4A in SMAD4-deleted pancreatic cancer and, in its original form, ENO2 in ENO1-deleted glioma. But these copy-number-based collateral targets share a limitation: they require homozygous deletion of the partner, whereas many clinically defining events, such as the 1p/19q co-deletion of oligodendroglioma, are heterozygous and halve a gene's dosage rather than remove it. Pushing the higher-order logic further, we show that a dosage-sensitive, expression-based approach fits the heterozygous genotype: scanning 1p and 19q for paralogs that become essential as their partner's expression falls returns ATP1A1, which cells come to require as the 19q-encoded neuronal pump subunit ATP1A3 is reduced. ATP1A1 is a dangerous target for a small-molecule drug, which cannot avoid the closely related neuronal subunit ATP1A3. An antisense oligonucleotide sidesteps this: it is selective by sequence, so it lowers ATP1A1 while sparing ATP1A3, and because it is too large to cross the blood-brain barrier, delivery into the spinal fluid keeps its action confined to the nervous system. Throughout, the discipline is the same: separate a driver from a passenger, a marker from a mechanism, and the depth of a deletion from its dosage.

Expression-aware qPCR primer prioritization. Scores any gene and primer pair for its probability of validation success using target biology rather than thermodynamics alone. Reaches AUC 0.75 on held-out assays; roughly halves failed validations while preserving good ones.

Deep-learning platform for predicting TDP-43 and RNA-binding protein binding footprints across neurological disease genes. Nominates antisense oligonucleotide target sites from transcriptome-wide eCLIP data.