Vivek Kanpa

Novartis

What cardiovascular conditions are associated with different features of ECG readouts?

Recently, the Patient Contrastive Learning of Representations (PCLR) featurization—a vector of ~320 numbers—was found to more "expressive, performant, and practical" than raw electrocardiogram (ECG) data for cardiovascular machine learning tasks (Diamant et al., 2022). Traditionally, cardiologists read an ECG to diagnose adverse cardiac events, monitor drug toxicity, and screen patients who might be at risk for certain conditions. Using XAI and generative learning techniques on large datasets from MIMIC-IV (MIT) and the UK Biobank (NHS), I seek to understand internal feature relationships in PCLR and associate feature perturbations with clinical cardiac conditions.

AI_PREMie

Goal: To develop a best-in-class AI-assisted clinical decision making tool for maternal preeclampsia.

Preeclampsia is a pregnancy disorder characterized by hypertension that affects 1 in 10 women, claiming the lives of 50,000 mothers and 500,000 babies annually. Because its warning signs are difficult to detect and may only appear close to delivery, preeclampsia often goes undetected until delivery, when complications can become serious. Following the discovery of novel blood biomarkers, the AI_PREMie team began developing an AI risk stratification tool to assist clinicians. I am running various machine learning optimization and feature engineering techniques to improve the performance of ensemble models at diagnostic classification. My optimization techniques, in addition to novel biomarker data, enables our model to outperform market competitors by large margins.

Revolution Medicines

Can protein structure inform strategic priorities for discovery biology and pharmacogenomics?

Performed structural alignment and hierarchical clustering (Dali, AlphaFold) of 600+ cancer driver proteins identified by OncoKB. Constructed a polar dendrogram where proteins were clustered by structural similarity, and encoded patient mutation frequency (AACR GENIE) with node radius. By clustering cancer driver proteins by structural homology, I informed the decision-making for next generation drug priorities and enabled efficient cancer targeting strategies. I documented this project's workflow over here.

Goals: develop generalizable tools to increase pharmacogenomics throughput and expidite ADME screening.

• Trained graph convolution network (DeepChem) using molecular featurization of RevMed hit library to improve permeability (Caco-2) predictions by 28% compared to conventional molecular weight cutoff.
• Designed a light-weight software system to automate mass spectrometry (MS) cross-linking analysis, calculating a normalized percent cross-linking value given raw relative abundance outputs. I enabled high throughput screening for drug-target binding, which shaved dozens of hours off each 384-well MS experiment and led to the discovery of a c-Myc-drug heterodimer complex a few weeks later.
• Performed in vitro dose-response assays of KRAS-mutant CRC cell lines and quantified c-Myc mRNA levels (PAGE, qRT-PCR) and protein levels (BCA assay, western blotting/ELISA) to understand downstream effect of KRAS inhibition on Myc activity.

Apfeld Laboratory

Goal: To be updated soon!

Description coming soon!

Takeda Pharmaceuticals

Full stack web developer for Global Biologics team in Cambridge, MA. Two of my major projects included:

• Designed end-to-end web apps with Django for protein quantification assay (HTP Lunatic) and HPLC chromatography (Agilent 1260 Infinity) from assay completion to database upload, increasing analysis throughput, data availability, and data standardization exponentially.
• Enabled high throughput assay workflows by building a translator for the Lynx liquid handler to interpret a scientists' experimental plate layout (96/384-well) into fulfillment instructions with picomolar precision.