Kiran Krishnamachari

I am a Scientist at the Genome Institute of Singapore, part of the Agency for Science, Technology and Research (A*STAR), where I work on AI for cancer genomics. I completed my PhD in Computer Science with a focus on Deep Machine Learning from the National University of Singapore in 2024. During my PhD, I had the pleasure to work with Foo Chuan Sheng, Anders Skanderup and See-Kiong Ng. My PhD thesis was titled Towards Robust Deep Learning with Real World Applications.

I am interested in developing fundamentally better deep learning algorithms and applications of these algorithms to real-world problems such as healthcare and genomics.

My Ph.D. research on using AI to detect cancer mutations has been published in Nature Communications and featured in some popular media outlets, such as The Straits Times, Genome Web, Asian Scientist, Channel News Asia, and more.

I have a B.Eng in Industrial and Systems Engineering from National University of Singapore.

☕ Support my work by buying me a coffee!

Email  /  GitHub  /  LinkedIn

Research

I'm interested in deep learning, AI for healthcare and genomics, predictive coding.

Comprehensive benchmarking of methods for mutation calling in circulating tumor DNA

We benchmark somatic mutation callers in plasma samples for liquid biopsy.

Identification of somatic variants in cancer genomes from tissue and liquid biopsy samples

A contributed book chapter on somatic variant calling in tissue and liquid biopsy samples.

Identifying somatic mutations from tumor-only sequencing using deep learning

We develop an accurate weakly supervised deep learning variant caller for tumor-only sequencing without a matched normal sample.

FFPEnet: Somatic variant calling in FFPE tumor samples using deep transfer learning

We develop a deep learning variant caller for tumor samples that reduces FFPE-associated artifacts.

RNAlign: Alignment of tumor and cell line transcriptomes using conditional VAEs

We propose a conditional variational autoencoder (VAE) to align transcriptomic data between cell lines and tumors.

Towards Robust Deep Learning with Real World Applications

My PhD Thesis

Uniformly Distributed Feature Representations for Fair and Robust Learning

We show that encouraging deep feature representations to be uniformly distributed makes models fairer and more robust, empirically and theoretically.

Mitigating Real-World Distribution Shifts in the Fourier Domain

We developed an simple yet effective unsupervised domain adapation method that operates by matching Fourier spectrum statistics between source and target domains. Works for multiple modalities e.g. images, audio, time-series.

Fourier Sensitivity and Regularization of Computer Vision Models

We analyse the Fourier sensitivity of deep learning models in a principled fashion and also study their effect on robustness. We also propose regularizers that can modify the frequency sensitivity of any model.

Accurate somatic variant detection using weakly supervised deep learning

We develop VarNet, the first accurate deep learning method for detecting cancer mutations.

ARMOURED: Adversarially Robust MOdels using Unlabeled data by REgularizing Diversity

We propose an adversarially robust method for training model ensembles using semi-supervision.

Design and source code from Jon Barron's website