Rishabh Sharma

• Research Associate Data Scientist in the department of computational biomedicine. • Developed a PyTorch-based U-Net pipeline for semantic segmentation of H&E-stained prostate cancer biopsy images, classifying tissue into background, stroma, healthy epithelium, and cancerous epithelium (Gleason grades 3-5) • Conceptualized strategies to mitigate class imbalance and improve segmentation accuracy on hard-to-classify pixels • Designed a histology-informed tiling approach to extract semantically meaningful regions from whole-slide images • Led hands-on instruction in Python and ML fundamentals for pathology through Cedars-Sinai’s AI Campus program

• Machine Learning Intern on the Apps & Algorithms team at Enable Medicine • Developed a ViT-based Masked Auto-Encoder (MAE) model in PyTorch, achieving a reconstruction RMSE of 0.005, to prototype a company-scale, self-supervised, foundation model for immunofluorescence, histopathology, whole-slide images • Executed distributed training over a grid of MAE architectures and hyper-parameters, optimizing tradeoffs between visible and masked patch reconstruction loss • Implemented end-to-end pipelines for linear probing, partial fine-tuning, and full fine-tuning of pre-trained MAE models, enabling a wide range of downstream image classification tasks

• Developed TensorFlow-based Convolutional-Recurrent Variational Auto-Encoder (VAE) models using molecular string representations for generative AI-driven chemical latent space exploration, in collaboration with Genentech • Analyzed VAE latent space for positional generative properties (uniqueness, novelty, diversity, density, relevance boundary, mode-collapse, and molecular property/similarity) as a function of latent decoding radius • Developed Variational Hetero-Encoder models using molecular string augmentation methods to condition latent properties • Developed De-noising VAEs using molecular string mutation methods to enhance generative molecular diversity • Investigated the VAE as a fuzz-tester of molecular representational robustness through outlier latent decoding (paper published in JCIM: Fuzz testing molecular representation using deep variational anomaly generation)

AI driven small molecule and protein design.

Applied multiple machine learning methods for materials classification, discovery, and analysis (Binary and multi-class classification models for Perovskite phase prediction)

• Developed a Restricted Boltzmann Machine (RBM) model using TensorFlow, generating up to 25% chemically valid and 90% drug-like molecules represented as SMILES strings, in collaboration with the Generative Therapeutics Design team • Developed a Deep Belief Network (DBN) model, generating up to 30% chemically valid and 70% drug-like molecules • Deployed RBM and DBN models as Pipeline Pilot components, enabling intra-company usage by computational chemists

Relevant Coursework - Basics of Machine Learning for Materials, Numerical Methods in Chemical Engineering