My favorite part is this -> For the first time, we describe the ability of a model to regress a proxy for TCR binding affinity with a deep learning model. We demonstrate in doing this from TCR-TetSeq, we can determine the binding contacts of a TCR from high-throughput NGS data!

When using this block within a supervised sequence classification task, we see (unsurprisingly) leveraging antigen-specific labels improves the learning of these models. Furthermore, the convolutional layers of the network allow us to extract the learned "motifs."

We first utilize this block in a Variational Autoencoder (VAE) and demonstrate improved antigen-specific clustering over current state-of-the-art methods.

The core of all our deep learning methods is a deep learning "featurization" block which learns a joint representation of TCR-Seq inputs (CDR3 sequence, V/D/J gene usage). In our latest version, we even incorporate HLA background as a possible input (more on this later).

#DeepTCR is a comprehensive deep learning framework for doing both unsupervised & supervised analyses at the sequence and repertoire level. Github 👇 https://t.co/aEHJlBszn3 Docs 👇 https://t.co/FZJeFWa2MA Tutorials 👇 https://t.co/U1zNnhngVa

In 2017, I attended a talk by @Google at @AACR on #DeepLearning. I realized then the potential for deep learning for analyzing TCR-Seq data & thus, the idea for #DeepTCR was born. 4 years later, our manuscript is now available at @NatureComms https://t.co/Gx6ujCt9ux

Finally, we provide "explainable AI" by incorporating an integrated gradients approach to reveal the relevant morphological features that are characteristic of APL. Surprisingly, we found our model did not identify Auer rods as being specific/sensitive for APL.

Our #ASH2020 abstract is now live! We present a multiple-instance deep learning model capable of rapidly identifying t(15;17) #APL from peripheral smear, potentially allowing more timely and appropriate therapy to this aggressive form of leukemia. https://t.co/rkeP6RHtyK