IA pour la découverte de molécules et matériaux

Mots-clefs : Artificial intelligence, machine learning, drug discovery, materials discovery, deep learning, reinforcement learning, discrete optimization, search, generative models, causality, active learning, out-of-distribution optimization, exploration algorithms, physique des matériaux, élaboration de nouveaux matériaux assister par l’apprentissage machine, cycle de vie des matériaux

Organisations pertinentes :

• Mila, IRIC, Médicament-Québec, Génome Québec, NRC, Valence Discovery, IBM, Samsung, Recursion Pharmaceuticals, Novartis, Genentech/Roche, GPAI, GARDP, DnDI, Bill and Melinda Gates Foundation, Chan-Zuckerberg Initiative.
• Valence Discovery, IBM, Samsung, Recursion Pharmaceuticals, Genentech/Roche, and Novartis have partnered with Mila and have a very keen interest in supporting research in AI and drug discovery, both financially and scientifically. If the theme we are proposing is selected by IVADO, we will pursue our discussions with these companies to further explore their involvement and level of support. Mila will also look into the possibility to allocate to the proposed project additional funding that our institute already secured from MEI (Ministère de l’Économie et de l’Innovation). This funding is aimed at fostering technology transfer in specific strategic application areas including drug and material discovery; as such, it would be a relevant complement to the IVADO funding.
• Our team is composed of a critical mass of world-class academic researchers with expertise in drug discovery and materials discovery, and a track-record of securing competitive funding for interdisciplinary projects in these areas, e.g. from Genome Quebec, Bill and Melinda Gates Foundation, Consortium Québécois du médicament (CQDM) among others. We will be seeking funding from these organizations and from others (e.g. NRC).

Personnes pertinentes suggérées durant la consultation : 

Les noms suivants ont été proposés par la communauté et les personnes mentionnées ci-dessous ont accepté d’afficher publiquement leur nom. Notez cependant que tous les noms des professeur.e.s (qu’ils soient affichés publiquement ou non sur notre site web) seront transmis au comité conseil pour l’étape d’identification et de sélection des thèmes stratégiques. Notez également que les personnes identifiées durant l’étape de consultation n’ont pas la garantie de recevoir une partie du financement. Cette étape sert avant tout à présenter un panorama du domaine, incluant les personnes pertinentes et non à monter des équipes pour les programmes-cadres.

Machine Learning:

• Yoshua Bengio (UdeM)
• Doina Precup (McGill)
• Jian Tang (HEC)
• Pierre-Luc Bacon (UdeM)
• Sarath Chandar (Polytechnique)
• Guy Wolf (UdeM)
• Mathieu Blanchette (McGill)
• David Rolnick (McGill)
• Irina Rish (UdeM)

Drug discovery:

• Mike Tyers (UdeM)
• Anne Marinier (UdeM)
• Yves Brun (UdeM)

Materials discovery :

• Yelena Simine (McGill)
• Jim Wuest (UdeM)
• Mickael Dollé (UdeM)
• Audrey Laventure (UdeM)
• Michel Côté (UdeM)
• Normand Mousseau (UdeM)

The task of searching a high-dimensional space such as that of molecules, looking for a few needles in a haystack, calls upon methodological advances in AI which not only have potentially transformative applications of major societal and economic value, they also connect with some of the frontiers of AI research begging to be pushed in order to approach human-level intelligence. In this project we propose to focus in particular on the following question: where and how should a learning agent explore in order to acquire knowledge which will be most useful later? We will focus on the specific space of molecules or multi-atom configurations and we propose to work with domain experts in drug discovery and materials discovery to evaluate the methodological advances made in this project. We propose to also work with industrial and non-profit partners who could provide practical experience and feedback and take up the most promising candidates for further study and potentially for clinical trials or more expensive assays.

The general methodological framework in which we propose to situate the above question can be summarized as follows. The learning agent interacts through a series of rounds with the real world (which we call an oracle). The oracle can score how useful a candidate molecule or material might be (not necessarily with full certainty, e.g., cellular assays do not give the full picture that mice assays provide, which do not yet reveal everything that clinical trials do). At each round, the learning agent can propose a batch of candidates, and receives a scalar score for each of them, thus augmenting its training set. The learning agent itself is broken down into two main parts: a world model and an experiment or query generator. The world model can be trained on the available data in a supervised way to predict the outcome (score) of experiments. The experiment generator produces the batch of candidates (queries for the oracle) at each round and can be trained by interrogating the world model, constructing a generative model which amortizes the cost of searching virtually for candidates, so it can then be used to actually generate those sent to the oracle.

The above raises many research questions, which different researchers in the project may focus on, including the following:

• How to represent molecules and materials as input and output (to generate new ones), e.g., using graph neural networks?
• How to predict properties of a candidate and more generally train a world model trying to capture the relevant aspect of the experimental assays?
• How to discover, represent and exploit epistemic uncertainty (lack of knowledge due to insufficient or incomplete data) in the world model?
• How to efficiently optimize in the discrete space of molecules or materials?
• How to train a generative policy which can sample from as many as possible of the high modes of some reward function (which combines the score estimated by the world model and epistemic uncertainty about it)?
• How to set up experimental assays from which large quantities of data can be obtained and that may have significant economic or societal value if the search succeeds?
• How to assess the value of a candidate molecule or material in a way that can guide the AI-driven search in the directions most likely to be of significant economic or societal value?
• What properties of a molecule or material matter to the above and what constraints on these properties (like synthesizability) must be taken into account by the AI candidate generator?
• How to exploit domain knowledge (about the molecules or materials and their effect) in the design of the world model used by the AI search to score candidates? For example this may include models of molecular dynamics or of low-energy joint atomic conformations.
• How to use machine learning to accelerate physical simulations that may be exploited by the world model (or are used as a proxy to an experimental assay)?
• How to design the world model and the generative model so as to take advantage of the causal view on the whole process, as a series of interventions aiming at uncovering causal structure?
• Given that the learning agent will have a single «  »life » » (made of say a few dozen rounds of interrogations of the oracle), how should it pick its hyper-parameters when it is in the middle of its life and has not yet experienced the end of it? This calls upon questions about the exploration-exploitation and the possibility of using meta-learning to simulate under different scenarios how the rest of the «  »life » » could unfold.
• How to explore efficiently in the high-dimensional space of molecules, taking into account the fact that the stream of experimental data feeding the learning agent is non-stationary (because the agent explores and pushes the boundaries of its knowledge as it progresses)?
• If the world model is causally structured (broken down into cause-effect mechanisms), how to discover some of its structure and how to reason over it in order to guide the exploration and the optimization?
• Considering that targeted pathogens can develop resistance to antimicrobials, how can we characterize the evolutionary distribution of likely mutations in order to favour candidate treatments while minimizing the probability of resistant mutations to these treatments?

Ajout 13/07 :
Voici quelques exemples d’axes de recherche qui pourraient être envisagés dans le cadre de ce thème stratégique.

Batteries :
défi : rendre les matériaux utilisés dans le stockage d’énergie conformes avec les principes de développement durable.

Alliages métalliques aciers, alliages d’aluminium, etc.
défi : développer les propriétés mécaniques recherchées en évitant les éléments les plus polluants ou les plus rares

Électronique :
défi : retrouver les propriétés électriques recherchées en utilisant des éléments facilitant le recyclage ou l’intégration.

Matériaux piézoélectrique :
défi : élaboration de matériaux piézoélectriques qui pourraient revaloriser la chaleur dégagée par les serveurs roulant les applications d’intelligence artificielle afin de limiter leurs impacts en termes de consommation d’énergie et de production de gaz à effet de serre.

Matériaux neuromorphiques :
défi : développer des matériaux qui sont plus près du fonctionnement du cerveau humain. Essentiellement, les algorithmes basés sur les réseaux de neurones utilisent l’architecture binaire des ordinateurs d’aujourd’hui pour simuler des processus analogiques. Les matériaux neuromorphiques pourraient remplir cette fonction directement.

(pas de documentation complémentaire pour le moment)