Ce projet vise à démontrer comment l'intelligence artificielle interprétable peut conduire à la génération de nouvelles connaissances biologiques en biologie computationnelle. Notre domaine d'application sera l'analyse des données protéomiques unicellulaires (CyTOF) des patients cancéreux. CyTOF est une technologie de pointe qui permet la lecture simultanée de ~40 protéines dans des millions de cellules individuelles. Appliqué à la recherche sur le cancer, CyTOF permet aux chercheurs de comprendre la composition moléculaire des patients cancéreux ainsi que de caractériser l'hétérogénéité tumorale, qui donne de l’information sur l’évolution clinique. S'appuyant sur une collaboration existante entre IBM Research et l'Université de Zurich, nous développerons d'abord de nouveaux modèles d'apprentissage en profondeur pour répondre aux besoins cliniques communs, tels que la stratification des patients en fonction du risque ou l'identification des communautés cellulaires importantes pour l'apparition du cancer, la progression de la maladie ou la réponse au traitement. Ensuite, nous développerons des techniques interprétables pour comprendre comment le modèle d'apprentissage profond a atteint chaque prédiction. Ce faisant, nous pourrons découvrir des règles moléculaires qui pointent vers des mécanismes moléculaires sous-jacents associés au cancer. En résumé, ce projet montrera comment les approches d'interprétabilité peuvent approfondir notre compréhension des mécanismes à l'origine des processus biologiques clés.
Since its reintroduction in seminal works by Bengio and others, deep learning has become one of the most active fields in machine learning, with astounding performance in a broad area of applications such as computer vision, speech recognition and natural language processing. However, with the improving performance of artificial intelligence (AI), the community has come to realize that for a model to be trusted, it must be not only accurate, but also explainable. Simply put, if users do not trust a model or a prediction, they will not use it, especially in a sensitive domain such as health care.
This proposal focuses on demonstrating how Interpretable Artificial Intelligence can lead to the generation of new biological insights in computational biology, a flourishing field where the recent application of deep neural networks (DNNs) to long-standing problems such as the prediction of functional DNA sequences, the inference of protein-protein interactions or the detection of cancer cells in histopathology images has brought a break-through in performance and prediction power. However, despite initial achievements, many of these models are built as black-boxes and have failed to provide new biological insights. This project rests on the recent demonstration that interpretable approaches can shed light on the underlying biological principles driving model decisions.
Our application area will be the analysis of single-cell mass cytometry (CyTOF) data from tumor samples. CyTOF technology is achieving tremendous high throughput in the number of profiled cells, with some of the most recent datasets consisting of millions of single cells. In this project, we will develop state-of-the-art deep learning networks to model CyTOF data, and use recent methodologies developed at IBM Research - Zürich on automatic architecture-search in order to optimize the architectures and hyperparameters and identify the best performing model for each learning task with a minimal computational effort. We will develop novel interpretable approaches exploiting different conceptual approaches. For instance, we will investigate how to produce disentangled representations that clearly and disjointly map latent and data generative factors. We will also explore architecture-constraint models, which can enhance the interpretability of certain components of the network. Finally, we will investigate parallels with Lagrangian mechanics in physics and develop approaches to identify invariances that can highlight underlying mechanisms. We will validate the generated explanations first through a literature search, to rule out unlikely explanations, and then through experimental perturbation experiments, which will be performed by Prof. Bernd Bodenmiller’s lab, a pioneer of the development and application of mass cytometry to tumor profiling and a long-established collaborator of the group of computational biology at IBM Research - Zürich.
In summary, this research programme will provide an in-depth understanding of how new biological insight can be extracted cumbersome deep learning models, and how interpretability approaches can deepen our understanding of the mechanisms driving key biological processes.
Last updated:18.06.2022
