Ewa Nowara

Senior Machine Learning Scientist at Prescient Design (Genentech)

Email  /  Resume  /  GitHub  /  Google Scholar  /  LinkedIn

I study how to adapt, steer, and post-train generative models for reliable scientific discovery, especially under scarce or imperfect feedback and competing objectives. My work draws on reinforcement learning, agentic search, autoregressive transformers, and diffusion and flow models. At Prescient Design, I lead research from method development through deployment for molecular and protein design.

Previously, I was a Research Scientist at Meta Reality Labs working on latent diffusion models and 3D generative models for AR/VR. I received my Ph.D. from Rice University in 2021 with Prof. Ashok Veeraraghavan, followed by a postdoc at Johns Hopkins with Prof. Rama Chellappa.

Full list on Google Scholar.

Autoregressive transformer for prompt-guided, multi-property molecule generation in low-data settings, further improved through RL post-training.

We investigate when 3D-rotational equivariance can be learned from data during training rather than hard-coded into model architectures, and how it emerges over the course of training.

We examine whether explicit equivariance is necessary for 3D molecular generative models, or whether comparable performance can be achieved with non-equivariant architectures and appropriate training.

A latent generative model for controllable design of molecular libraries, combining neural empirical Bayes with learned 3D latent representations to enable efficient exploration of chemical space.

We evaluate electron-density-based representations for 3D molecular learning as a richer alternative to atom-centric representations.

Selected work from my Ph.D. and postdoc on computer vision and camera-based physiological sensing.

A review of state-of-the-art physiological and medical imaging modalities that leverage recent advances in computational photography.

TransLocator: a dual-branch Vision Transformer for geolocalizing images anywhere in the world, using both RGB and semantic segmentation features for robustness to appearance variation.

We use the inverse of an attention mask to estimate corruption in video regions without the signal of interest, then use it to denoise temporal observations of physiological signals.

We improve the generalizability of deep learning models for non-contact cardiac monitoring by augmenting the training set videos with “magnified” videos. These data augmentations are geared towards revealing useful features for recovering the physiological signals.

We animate blood flow patterns to augment the appearance of synthetic avatars and photo-realistic faces based on a data-driven physiological model.

We show that deep learning models can learn how noise at different video compression levels affects the physiological signals and are able to reliably recover vital signs from highly compressed videos, even in presence of large motion.

We evaluate how much gender and skin tone affect vital signs estimation from video. We find that performance drops significantly on videos of people with very dark skin tones, especially for machine learning algorithms.

We reduce most outside light variations using narrow-band near-infrared (NIR) video recordings to obtain reliable heart-rate estimates. We present a novel optimization algorithm, AutoSparsePPG, that leverages the quasi-periodicity of physiological signals and outperforms prior methods.

We use 3D face tracking to estimate the position of facial landmarks with pixel-level accuracy to improve motion robustness of camera-based vital sign estimation.

Small intensity variations in the skin related to physiological signals can be recovered even from very compressed videos with supervised deep learning.

Narrow-bandwidth near-infrared active illumination for camera-based vital signs, with a novel signal tracking and denoising algorithm based on Robust PCA and sparse frequency estimation.

A machine learning system to prevent face-spoofing attacks by detecting and analyzing the heartbeat signal from face videos.

M.R. Maser, E.M. Nowara, P.H.O. Pinheiro, S. Saremi. Three-Dimensional Molecule Generation in Latent Voxelized Space. US Patent App. 19/390,374, 2026.

P.H.O. Pinheiro, S. Saremi, M.R. Maser, E.M. Nowara. Three-Dimensional Molecule Generation by Denoising Voxel Grids. US Patent App. 19/390,401, 2026.

T. Marks, H. Mansour, E.M. Nowara, Y. Nakamura, A. Veeraraghavan. System and Method for Remote Measurements of Vital Signs of a Person in a Volatile Environment. US Patent App. 17/199,696, 2021.

T. Marks, H. Mansour, E.M. Nowara, Y. Nakamura, A. Veeraraghavan. System and Method for Remote Measurements of Vital Signs. US Patent App. 16/167,668, 2019.

• Invited speaker, Houston Methodist – Rice Digital Health Institute Industry Day (2026)
• Best Paper Award, AI for Science Workshop at NeurIPS, for Training Dynamics of Learning 3D-Rotational Equivariance (2025)
• Featured in the RSIP Vision Magazine as a Woman of the Month in Computer Vision (2022)
• Invited speaker, Microsoft Research AI Breakthroughs (2020)
• Best graduate poster and demo award, ECE Corporate Affiliates Day at Rice, for “SparsePPG: Towards Driver Monitoring Using Camera-Based Vital Signs Estimation in Near-Infrared” (2019)
• Presidential Award, St. Mary’s University (top 14 graduating undergraduate seniors) (2015)

Website credits to Jon Barron.