Our method opens up the possibility of discovering drug-tolerance related genes from single clinical samples, not just limited to settings where genetic engineering is feasible. Incredible work by #SuvranilGhosh, and a really fun and enjoyable collaboration with #ArchishmanRaju.

We demonstrated that memory genes discovered from many independent samples of a melanoma cell line (Memory-Seq), are recovered using Power-Seek, but with just one scRNA-seq dataset. Excitingly, we also demonstrate its applicability in human breast cancer tissue. Many ECM and EMT genes showed up.

We showed that single samples will also exhibit power laws in the presence of memory genes. We developed a simple and easy to use algorithm, 'Power-Seek' (all puns intended). This detects memory genes using variations in power-law upon removing genes one at a time from a single scRNA-seq dataset.

Demonstrating the theory in single samples posed intriguing difficulties, since cell-cell correlations get mixed up with gene-level variations, precluding use of methods like PCA. This took us into a deep dive in Random Matrix Theory, from a beautiful paper on protein evolution: tinyurl.com/3y42v7zt

While seemingly impossible, this can indeed be done due to a beautiful underlying theoretical result that we show: memory genes, which give rise to lineage correlations, generate detectable Power-Laws in the eigenspectrum of the cell-covariance matrix of a scRNA-seq dataset.

However, these methods are conceptually based on variants of the classic Luria-Delbruck framework, requiring either many samples or lineage information from barcodes to identify memory genes. We asked whether *all* these requirements can be done away with using a completely different approach?

We now know that non-genetic states inherited across cell divisions are associated with cancer drug tolerance. Beautiful methods: Memory-Seq, Rewind, PATH, GEMLI have recently been developed to discover these states. We reviewed one such method (Rewind) recently: rdcu.be/d11Jm