Thread by @MoritzGerstung

Can you see mutations in cancer cells? Kind of.

We trained a neural network on 17k tumour slides with known genomics transcriptomics to assess how histopathology, molecular tumour characteristics and survival correspond. 1/8 nature.com/articles/s4301…

This analysis discovered histopathological patterns of 167 different mutations ranging from whole genome duplications to point mutations in cancer driver genes - about 1/4 mutations tested. 2/8

Further, around 40% of the transcriptome is correlated with histopathology reflecting tumour grade and composition. This is probably best illustrated at the example of infiltrating lymphocytes TILs, which can be identified and localised through their expression signature. 3/8

Lastly, we found that one can also automatically learn prognostic patterns (such as necrosis and TILs), which provide decent predictions. 4/8

For many of these associations the underlying histopathological patterns didn't appear to be captured by conventional histopathological grades and subtypes, suggesting that there is indeed a complementary value in digital pathology analyses. 5/8

Thus molecularly informed digital pathology may provide clues when genomics is unavailable, help shed light on genetic variants of unknown significance and serve as an additional layer of information for developing integrated classification schemes. 6/8

This was a hugely inspiring collaborative effort by @yufu0413, @alex_w_jung, @luiza_moore, @LucyYat47076319 and many others. Reassuringly, and published back to back, @jnkath and colleagues arrived at similar conclusions, using a different network. 7/8 nature.com/articles/s4301…

All these findings are nicely summarised in the accompanying News & Views article by Nicolas Coudray & Aristotelis Tsirigos. Thanks! 8/8 nature.com/articles/s4301…

And here is an open access version of the article: rdcu.be/b5R5y