Ava Amini

I am a Principal Researcher at Microsoft Research in Cambridge, MA.
My research focuses on developing new AI methods to understand and design biology, with the ultimate aim of realizing precision biomedicines to improve human health. To this end, I co-lead Project Ex Vivo, a collaborative effort between Microsoft and the Broad Institute, that is focused on defining, engineering, and targeting cell states in cancer.

I completed my PhD in Biophysics at Harvard University, where I worked with Sangeeta Bhatia at the Koch Institute for Integrative Cancer Research and was supported by the NSF Graduate Research Fellowship. I received my Bachelor of Science in Computer Science and Molecular Biology from MIT, where I was recognized as a Henry Ford II Scholar and with the AMITA Academic Award.

Research  /  Teaching  /  Talks

Email  /  CV  /  Google Scholar  /  Twitter

My research focuses on developing new AI methods to understand and design biology, with the ultimate aim of realizing precision biomedicines to improve human health. My work bridges the computational and experimental worlds to create new diagnostic and therapeutic biotechnologies and to achieve new insights into cancer.

We introduce a hierarchical cross-entropy loss for single-cell cell type annotation, finding that this simple modification significantly improves out-of-distribution performance without added computational cost.

We evaluate the performance of single-cell foundation models in ``zero-shot`` settings where they are used without any further training, finding that they are outperformed by simpler methods.

ProtNote is a multimodal deep learning model that leverages free-form text to enable both supervised and zero-shot protein function prediction.

We introduce a new deep learning framework, MICON, that models chemical compounds as treatments that induce counterfactual transformations of cell phenotypes for improved representation learning in microscopy-based morphological profiling.

We develop a method for protecting against over-clustering in analysis of single-cell RNA sequencing data, by controlling for the impact of reusing the same data twice when performing differential expression analysis.

We present CleaveNet, an end-to-end AI pipeline for the design of peptide-based protease substrates, enabling generation of peptides guided by a target cleavage profile for the design of efficient and selective substrates.

We review recent advances in deep generative models for protein design, with a particular focus on sequence-structure co-generation methods.

We systematically investigate the consequences of training dataset composition on the behavior of deep learning models of single-cell transcriptomics, focusing on human hematopoiesis as a tractable model system and including cells from adult and developing tissues, disease states, and perturbation atlases.

We assess deep learning-based uncertainty quantification methods on protein sequence-function prediction tasks.

We investigate the role of pre-training dataset size and diversity on the performance of single-cell foundation models on both zero-shot and fine-tuned tasks, finding that current methods plateau in performance with pre-training datasets that are only a fraction of the full size.

We take a deep dive into scBERT, one recently developed transformer model for single-cell RNA-sequencing data, to develop a deeper understanding of the potential benefits and limitations of single-cell foundation models.

We develop EvoDiff, a general-purpose discrete diffusion over protein sequences, that combines evolutionary-scale data with the distinct conditioning capabilities of diffusion models for controllable protein generation in sequence space.

We establish a method of pooling perturbations, like chemical compounds, followed by computational deconvolution to reduce required sample size, labor, and cost in high-throughput phenotypic screens.

Utilizing patient-derived xenograft (PDX) models and clinical trial specimens of acute lymphoblastic leukemia (ALL), we examined how genetic and transcriptional features co-evolve to drive progression during prolonged tyrosine kinase inhibitor response, uncovering a landscape of cooperative mutational and transcriptional escape mechanisms that differ from those causing resistance to first generation inhibitors.

To understand how the features learned in pretraining protein language models (PLMs) relate to and are useful for downstream tasks, we perform a systematic analysis of transfer learning using PLMs, conducting 370 experiments across a comprehensive suite of factors including different downstream tasks, architectures, model sizes, model depths, and pretraining time.

We developed a non-parametric infinite mixture model that leverages Bayesian sparse priors to identify marker genes while simultaneously performing clustering on single-cell expression data.

We present FoldingDiff, a diffusion-based generative model that generates protein backbone structures via a procedure inspired by the natural folding process.

We develop intravenous priming agents that are given prior to a blood draw to increase the abundance of cell-free DNA in circulation, improving the sensitivity of liquid biopsy cancer diagnostic assays.

We develop a self-supervised masked language model for biosynthetic gene clusters in bacteria, and leverage it for natural product classification.

A commentary on a recent image-inspired, diffusion model to generate new protein structures.

We discover "sanctuaries" within lymph nodes that contain low proteolytic activity and act as a safe haven for vaccines, and demonstrate that this heterogeneity can be exploited to enhance vaccine-induced antibody response.

We engineer an integrated set of methods for measuring specific protease activities across the organismal, tissue, and cellular scales, and unify these methods into a methodological hierarchy that powers new biological insights into cancer.

We build Protease Activity Analysis (PAA), a Python software package with a collection of data analytic and machine learning tools for analyzing protease activity data.

We develop a sensor-based, ML-driven system to diagnose pneumonia and classify its etiology, using machine learning to classify directly from molecular barcodes.

We establish a sensor-based, ML-driven diagnostic for noninvasive, real-time monitoring of disease in a preclinical model of lymphangioleiomyomatosis (LAM), a rare lung disease.

We design an easily applicable, non-invasive formulation to deliver diagnostic nanosensors through the skin, enabling a sustained release diagnostic monitoring system for detecting thrombosis.

We design a sense-and-respond system that integrates a synthetic gene circuit and nanotechnology detection tools for tumor-specific expression of heterologous biomarkers.

A fast, scalable approach for uncertainty quantification in neural networks enables uncertainty-aware molecular property prediction, accelerated property optimization, and guided virtual screening.

We engineer a new class of enzyme activity probes that can be applied to fresh-frozen tissue sections to spatially localize protease activty, enabling new insights into the biology of protease dysregulation.

We develop a novel algorithm for fast, scalable uncertainty quantification in highly complex, non-linear neural networks trained for regression tasks.

We optimize the immunogenicity of peptide-based antitumor vaccines in mice by tuning their pharmacokinetics via fusion of the peptide epitopes to protein carriers.

Review detailing how integrating techniques from multiple disciplines has developed engineered diagnostics that are selectively activated in disease states, highlighting their potential to realize the goals of precision medicine.

We couple protease-responsive nanoparticle sensors with machine learning to engineer a sensitive and specific urinary test for lung cancer detection.

Magnetic nanoparticles that display genetically encoded targeting peptides to promote tumor accumulation and enhance MRI contrast.

By leveraging the unique properties of catalytic nanomaterials, we develop a simple color-change urine test for detection of cancer in mice.

Convolutional neural networks for automated segmentation and uncertainty estimation of microscopy images of malaria infection.

Engineered microrobots that use magnetism to push drug-delivery nanoparticles out of blood vessels and into diseased tissue.

Generalizable algorithm for mitigating hidden biases within training data, by leveraging learned latent distributions to adaptively re-weight the importance of certain data points while training.

Estimating uncertainty in neural networks for end-to-end control by exploiting feature map correlations during training.

Programming biological state machines that enable cells to remember and respond to a series of events.

In addition to research, I am passionate about education and leadership, and strive to help and empower others to excel in their own pursuits.

I am an organizer and lecturer for Introduction to Deep Learning (6.S191), MIT’s official introductory course on deep learning foundations and applications. Together with Alexander Amini, I have organized and developed all aspects of the course, including developing the curriculum, teaching the lectures, creating software labs, and collaborating with industry sponsors. All materials can be found online on the course website.

I am a co-founder and director for Momentum AI, an outreach program that teaches AI and machine learning to under-resourced and under-served high school students from the greater Boston area. Our two-week capston program is a free, project-based deep dive into AI and is held on MIT's campus.