Light travels at about 1 foot per nanosecond.

With @franklowery our new study on attributes of T cells that contribute to successful cell therapy in cancer patients in @ScienceMagazine today w/ colleagues @slgoff_SB, @NCI_CCR_SB @theNCI a TL;DR thread on key findings with caveats :) 1/10

has a long history of using tumor infiltrating T cells (TILs) to treat cancers since well.. before I was even born. We analyzed our most successful melanoma ACT trial for cell surface phenotypes in TIL infusion products of patients (aPD1/immunotherapy naive) /2

Surprisingly we found a CD39- TIL subset (CD39-CD69-, DN) associated with ACT-response. TBH we were expecting the opposite (CD39+). We only included CD39 bcz multiple groups (e.g. Simoni et al, 2018) had reported CD39+ as enriching for anti-tumor/neoantigen reactive T cells. /3

CD39- DN TILs RNA/epigenetics resemble stem-like memory progenitors, and in vitro were able to self-renew, and give rise to other CD39+ subsets. OTOH the most dominant subset of patient infusion products were CD39+ CD69+ (DP) and these guys were terminally differentiated.. /4

So, to clarify we specifically analyzed tumor-specific mutation-reactive Tcells. Turns out, ACT-responders had pool of neoantigen-reactive TILs in the CD39- phenotype, while non-responders did not (despite other irrelevant CD39- Tcells) -> not all CD39- T cells are bystanders /5

1/

Get a cup of coffee.

In this thread, I'll help you understand the basics of Binomial Thinking.

The future is always uncertain. There are many different ways it can unfold -- some more likely than others. Binomial thinking helps us embrace this view.

2/

The S&P 500 index is at ~3768 today.

Suppose we want to predict where it will be 10 years from now.

Historically, we know that this index has returned ~10% per year.

If we simply extrapolate this, we get an estimate of ~9773 for the index 10 years from now:

3/

What we just did is called a "point estimate" -- a prediction about the future that's a single number (9773).

But of course, we know the future is uncertain. It's impossible to predict it so precisely.

So, there's a sense of *false precision* in point estimates like this.

4/

To emphasize the uncertainty inherent in such predictions, a better approach is to predict a *range* of values rather than a single number.

For example, we may say the index will return somewhere in the *range* of 8% to 12% (instead of a fixed 10%) per year.

5/

10 years from now, this implies an index value in the *range* [8135, 11703]:

Today, we are bringing other exciting results involving black holes and AI. We released a new paper:

"Black Hole Weather Forecasting with Deep Learning: A Pilot Study"

Work by Roberta Duarte (@import_robs), Rodrigo Nemmen (@nemmen) and João Paulo Navarro (from @NVIDIABrasil).

The authors used deep learning to simulate the dynamics of gas accreting onto a black hole, i.e., black hole weather forecasting.

They trained the model (U-Net) with frames from numerical solutions of the hydrodynamical equations.

Numerical simulations are time-consuming. A simple simulation can take as long as 7 days to finish. If we go with more complex simulations, this time may increase.

We want to investigate if deep learning can be a new method to simulate accurately in less time!

In the paper, they discussed two examples:

1- The model simulating only one system after learning only from this system
2- The model simulating an unseen system after training with several systems with different initial conditions

In the first example, they trained the model with a single system and analyzed how the model simulates by iterative predictions.

The result is that the model can simulate up to 8e4 gravitational time accurately with a speed-up of 30000x faster!