1/n: 1 month ago the team at @RecursionPharma partnered with a Biosafety Level 3 Lab to assess if they could 1) phonemically characterize in vitro COVID19 and 2) run high throughput screens to find therapeutic candidatures as potential cures
https://www.biorxiv.org/content/10.1101/2020.04.21.054387v1.full.pdf+html

2/n: Industry standard in vitro assays for viruses are usually 1-dimensional "viral load" readouts from monkey kidney epithelial (VERO) cells. Rather, @RecursionPharma took *unbiased* images of healthy cells, COVID19 infected cells, and irradiated cells.

3/n: Doing so both in VERO and human renal cortical epithelial (HRCE) cells. Here's what these look like under a microscope. Impossibly complex, and rich features with complex and non-linear relationships. A viral load score is worth a single number. An image? A thousand words...

4/n: The study then trained deep neural nets to both distinguish between "health" and COVID19 infected cells AND C19 infected cells and all independent toxicity (the irradiated cells). The nets were able to distinguish statistically significant features for VERO and HRCE

5/n: Finally, these models were projected against 2 axes.
X-axis: a quantitative phenotypic score of C19
Y-axis: a quantitative phenotypic score of C19 independent toxicity

6/n: With a quantified phenotypic projection in hand, they then ran a blinded screen of 1,670 approved & reference compounds that through literature or active clinical trials are considered candidates for C19. Given the global urgency from C19, the results were stunning..

7/n: each of the blindly screened drugs produced a graph like the below, showing effect on diseases cells as dosage increased. A high score would take an initial dot from inside the purple circles to inside the blue circles..

8/n: A drug that furthered the C19 phenotype would push the dots to the right. A drug that furthered C19 independent toxicity would push the dots up. Highest performers pushed the dots down and to the left.

9/n: Still blinded, the distribution of performance looked essentially like a bell curve. The vast majority -- including those under active clinical trial -- did not look promising, and only a small minority showed any signal. Drug discovery is HARD.

10/n: Importantly for scientists globally, the study found that the in vitro gold standard does not reliably translate to human cell lines. Rather the ability of a drug to translate from VERO to human cells is likely to be mechanism and target specific.

11/n: You may have noticed CQ and HCQ *underperforming* on human cells while *overperforming* on monkey cells. Perhaps more damning, CQ and HCQ increased C19 and toxicity scores in the HRCE phenotypic screens as indicated by the blue and orange dots moving up and to the right.

12/n: In that context, this isn't so surprising -- though politically tragic: https://twitter.com/ricfulop/status/1253697521001680897

13/n: Remdesivir? Well it turns out it was the highest performing candidate from the 1,670 drugs. Remdesivir and its parent nucleoside both pushed diseased human cells down and to the left. Curing C19 phenotypes with minimal orthogonal effects to the cells.

14/n: But what about yesterday's leaked preliminary trial results from China? Again, drug discovery is HARD and humans are COMPLEX. The only way to validate true human efficacy of a drug is via Randomized Control Trials. No matter how good your AI gets ;) https://www.statnews.com/2020/04/23/data-on-gileads-remdesivir-released-by-accident-show-no-benefit-for-coronavirus-patients/

15/15: In the meanwhile we're left with a heightened new knowledge of the challenges outstanding for finding a C19 cure, a re-ordering of global leads, & a budding internal pipeline at @RecursionPharma for entirely novel mechanisms and classes to fight back against COVID19.