Fine, here we go.

First let's start with some baseline expectations. Pancreatic cancer is scary, 5y survival is ~10%.

So if you see a trial with an apparent survival asymptote at ~50% then certainly *something* is happening, right?

https://twitter.com/TheVinodLab/status/1892245578446184873

Well, if you look at the # at risk, we're down to *1* patient by 4y, so the confidence intervals by the end of this curve should be widening to include 0%. So let's say the 5y OS in a bigger cohort would end up 30-40%. Still has to be some kind of therapeutic activity, right?

Let's say you want to publish in a top-tier journal and need to have a high accuracy predictor of something of great medical importance, such as survival of cancer patients in response to immunotherapy. The easiest route is just to cheat: (X,y).predict(X) model.fit
Since you're just memorizing the test data (or maybe a superset of it), then the predictor works better as you add more features.

Finishing up my Defense Against the Dark Arts syllabus and realized I don't have anything that's really specific to large deep learning models applied to -omics or other biomedical data.

What are some egregious missteps there which show up in high profile papers? Tentative list of topics:

1) Just make it up: image manipulation and tables which don't add up
2) Try everything and report the small p-values: uncorrected multiple hypothesis testing

Preprint live: biorxiv.org/content/10.110…
tl;dr We did not disprove the role of T-cells in clearing SARS-CoV-2 (sorry to disappoint) but we did manage to make a vaccine which induces strong but useless T-cell responses against SARS-CoV-2. Details & guess at interpretation below
1/n

https://twitter.com/iskander/status/1496236919130308616

Many SARS-CoV-2 vaccines present the immune system with the spike protein or the parts of the spike most vulnerable to neutralization (S1 portion or even just the RBD). This antigenic content can be made directly (protein subunit), encoded as mRNA, DNA, DNA in a virus, &c
2/n

It's been a preprint for a while but a much improved version of our paper is out in @GenomeMedicine

"Landscape and selection of vaccine epitopes in SARS-CoV-2"

genomemedicine.biomedcentral.com/articles/10.11…

With @ChristofCSmith1 @KOimmuno @BenjaminGVincen + many others

Tweetorial:

1/ In Feb 2020 SARS-CoV-2 still felt far away, a twitter feed of China + isolated cases in other countries (remember #ncov2019?)

I was driving to NC to start my new job at @UNC_Lineberger and was thinking about making a peptide vaccine for SARS-CoV-2...

2/

New paper on @biorxivpreprint:

"Landscape and Selection of Vaccine Epitopes in SARS-CoV-2"

(biorxiv.org/content/10.110…)

Collaboration with @ChristofCSmith1 @bgvincentlab @JuliaKodysh @timodonnell @glycam @Olivercgrant @PEPperCHIP @erikgarrison & others.

Thread:

(1/) There are a zoo of vaccine approaches to SARS-CoV-2:

https://twitter.com/AndyBiotech/status/1268700820188606465

Front-runners: inactivated virus (SinoVac), mRNA (BioNTech, Moderna), electroporated DNA (Inovio), adenovirus (CanSino, Oxford), rVSV (IAVI), rS (Novavax) &c

Missing from the list: peptides

2/