Resume

Education

Stanford 2014-2020 Ph.D.	Ph.D. in Computer Science, applying deep learning to study regulatory genomics Ph.D. Advisor: Anshul Kundaje.
MIT 2009 - 2013 B.S.	B.S. Computer Science with Molecular Biology, Minor in Mathematics Undergraduate GPA: 5.0/5.0

Experience

Kundaje Lab Sep 2014 - Sep 2020

Developed DeepLIFT for explaining the individual predictions made by a deep learning model Developed TF-MoDISco for understanding the patterns learned by a deep learning model. Developed a flexible framework for deciding which examples a predictive model should abstain on when optimizing metrics other than the standard classification accuracy. Developed GkmExplain for explaining the individual predictions made by nonlinear gapped-kmer Support Vector Machines. Developed Separable Fully-Connected Layers and Reverse-Complement Parameter Sharing to exploit unique properties of genomic data for deep learning. Developed an engine for creating simulated genomic datasets; used in multiple projects to generate a simulated dataset to empirically validate a method. Developed an internal framework to facilitate the training of deep learning models in the Kundaje lab.

Palantir Technologies June 2013 - Sep 2014

Forward Deployed Engineer for the Healthcare Team. Primary developer for pilot project which successfully became the Healthcare Team's first enterprise deployment. Continued as primary developer until I left the company. Created a way to integrate Palantir's two major platforms at the time into a seamless experience; success was highlighted company-wide. Went beyond my assigned duties to build multiple handy software utilities for use team-wide and company-wide. Consistently rated high on bi-annual reviews. Was told that I had at times done the work of 2-3 coders and had set a new standard for productivity among my teammates.

Selected Talks

ICML 2020 July 2020	Maximum Likelihood with Bias-Corrected Calibration is Hard-To-Beat at Label Shift Domain Adaptation Video here.
ISMB 2019 July 2019	GkmExplain: Fast and Accurate Interpretation of Nonlinear Gapped k-mer SVMs Video here.
Biological Data Science Nov 2018	Suggested best practices for interpreting deep learning models of regulatory DNA Slides here.
NVIDIA GTC March 2018	Not Just a Black Box: Interpretable Deep Learning for Genomics and Beyond Video here.
NIPS MLCB Dec 2017	TF-MoDISco: Deep learning non-redundant, predictive sequence motifs of transcription factors Video here.
ICML 2017 Aug 2017	Learning Important Features Through Propagating Activation Differences Video and slides here and here, paper here.
Deep Learning In Healthcare Summit, Boston May 2017	Not Just a Black Box: Interpretable Deep Learning for Genomics Interview here.
CEHG Symposium March 2016	Not Just a Black Box: Interpretable Deep Learning for Genomics Video here.

Selected Publications

ICML Workshop on Computational Biology July 2020	Look at the Loss: Towards Robust Detection of False Positive Feature Interactions Learned by Neural Networks on Genomic Data. Mara Finkelstein*, Avanti Shrikumar*† Anshul Kundaje† *co-first authors † co-corresponding authors Novel strategy to detect when feature interactions learned by a neural network may be false positives, by looking at the impact that the learned interaction effect has on the model's prediction loss on held-out data.
ICML 2020 July 2020	Maximum Likelihood With Bias-Corrected Calibration is Hard-To-Beat at Label Shift Adaptation. Amr Alexandari*, Anshul Kundaje†, Avanti Shrikumar*† *co-first authors † co-corresponding authors Algorithm that gives state-of-the-art results at a problem called domain adaptation to label shift, which arises when adapting a trained classifier to perform well in a scenario where the class proportions are different compared to when the classifier was first trained (e.g. adapting a disease predictor to account for a surge in cases due to a pandemic)
Proceedings of the ISMB July 2019	Gkmexplain: Fast and Accurate Interpretation of Nonlinear Gapped k-mer SVMs. Avanti Shrikumar*†, Eva Prakash*, Anshul Kundaje† *co-first authors † co-corresponding authors Computationally efficient algorithm for explaining individual predictions made by nonlinear gapped-kmer SVMs trained on genomic sequences.
Nature Biotechnology May 2019	Kipoi: accelerating the community exchange and reuse of predictive models for genomics. Žiga Avsec*†, Roman Kreuzhuber*, Johnny Israeli, Nancy Xu, Jun Cheng, Avanti Shrikumar, Abhimanyu Banerjee, Daniel S. Kim, Lara Urban, Anshul Kundaje†, Oliver Stegle†, Julien Gagneur† *co-first authors † co-corresponding authors Model zoo for genomics. I was involved in designing the API and converted the DeepBind models to Keras using the code here
arXiv July 2018	Computationally Efficient Measures of Internal Neuron Importance. Avanti Shrikumar*†, Jocelin Su*, Anshul Kundaje†. *co-first authors † co-corresponding authors Showed an equivalence between Total Conductance, a recently-proposed method for computing internal neuron importance, and Path Intergrated Gradients, thereby providing a computationally efficient way to compute the former. The reformulation of Total Conductance was referred to as Neuron Integrated Gradients. Benchmarked Neuron Integrated Gradients against DeepLIFT. Colab notebook reproducing results here
Journal of the Royal Society Interface April 2018	Opportunities And Obstacles For Deep Learning In Biology And Medicine Collaboratively written review on deep learning for biology and medicine. I wrote the section on interpretation. Note that the final published version was stripped down due to word limits. I have linked to the original submission from my end.
arXiv Feb 2018	A Flexible and Adaptive Framework for Abstention Under Class Imbalance. Avanti Shrikumar*†,Amr Alexandari*, Anshul Kundaje†. *co-first authors † co-corresponding authors Proposes a framework for identifying which examples to abstain on in order to optimize for a specific metric of interest. Leverages the insight that because the calibrated probabilities can be used as a proxy for the true label, optimization is possible even when the ground-truth labels are not known. Derived computationally efficient algorithms for optimizing auROC, sensitivity at a target specificity and weighted cohen's kappa. Showed that by leveraging strategies for domain adaptation to label shift, the abstention algorithms can apply even in situations where the test-set distribution has a different class imbalance compared to the training-set distribution.
ICML Workshop on Computational Biology (Spotlight Talk, Best Poster) June 2017	Separable Fully-Connected Layers Improve Deep Learning Models For Genomics. Amr Alexandari*, Avanti Shrikumar*, Anshul Kundaje. *co-first authors Adapts deep learning models for genomics by leveraging known patterns in transcription factor binding data.
ICML April 2017	Learning Important Features Through Propagating Activation Differences. Avanti Shrikumar, Peyton Greenside, Anshul Kundaje. Details a computationally efficient algorithm to explain individual predictions of a deep learning model by assigning contribution scores to individual parts of the input. Code here.
BioRxiv January 2017	Reverse-Complement Parameter Sharing Improves Deep Learning Models For Genomics. Avanti Shrikumar*, Peyton Greenside*, Anshul Kundaje. *co-first authors Adapts deep learning models for genomics by leveraging the reverse-complement property of DNA sequence.
Circulation Research Dec 2014	Transcriptional Reversion of Cardiac Myocyte Fate During Mammalian Cardiac Regeneration. O'Meara CC, Wamstad JA, Gladstone RA, Fomovsky GM, Butty VL, Shrikumar A, Gannon JB, Boyer LA, Lee RT Collaboration between Boyer lab at MIT and Lee lab at Harvard. I analysed RNA-seq data to study transcriptional reversion.
Cell Sep 2012	Dynamic and coordinated epigenetic regulation of developmental transitions in the cardiac lineage. Wamstad JA, Alexander JM, Truty RM, Shrikumar A, Li F, Eilertson KE, Ding H, Wylie JN, Pico AR, Capra JA, Erwin G, Kattman SJ, Keller GM, Srivastava D, Levine SS, Pollard KS, Holloway AK, Boyer LA†, Bruneau BG†. † co-corresponding authors Collaboration between Boyer lab at MIT and Gladstone Institutes. I performed the bulk of bioinformatics analysis at the Boyer lab. 355 citations as of Jan 2018.

Recognition

HHMI International Student Research Fellowship 2016	Awarded to 20 international students. Announcement here.
Stanford Bio-X Fellowship 2016	Bio-X fellowships are awarded to about 25 students annually for interdisciplinary research. Announcement here.
Microsoft Women's Fellowship 2016	Awarded to one woman per participating University pursuing or interested in pursuing a PhD. Announcement here.
Outstanding Research Award Spring 2013	Awarded to 3 projects completed as part of MIT's SuperUROP program. My project was done in the Kellis lab. Announcement here.
Sophomore Academic Excellence Award Fall 2011	AIChE Sophomore Academic Excellence Award for the student with the highest GPA among chemical engineers after sophomore year at MIT. Announcement here.
IGCSE Examinations June 2006 & 2007	The IGCSE was administered in roughly 300 schools in India. I had the highest score in India in Extended Mathematics Without Coursework (June 2006; press release), Physics (June 2007) and Geography (June 2007).

Selected Coursework

Stanford	Probabilistic Graphical Models	CS 228	Winter 2015	A+
Stanford	Machine Learning	CS 229	Fall 2015	A
MIT	Statistics for Applications	18.443	Spring 2013	A
MIT	Advanced Computational Biology	6.878	Fall 2012	A+
MIT	Design and Analysis of Algorithms	6.046	Spring 2012	A
MIT	Software Construction	6.005	Spring 2012	A+
MIT	Evolutionary Biology	7.33	Spring 2012	A+