진학프로석·박사 채용 접수 플랫폼

Motivation: Machine learning (and deep neural network methods in particular) have revolutionised image processing and text/speech recognition domains. They have also made a tremendous impact in different branches of life sciences. These range from the prediction of protein structures by Google's DeepMind [1] and academic teams [2] to drug design, molecular interactions, and many more. Thus, AI methods are readily applicable, with various extensions, to open problems in structural bioinformatics and computational physics. The fundamental methodological challenges, however, remain. Many of them relate to the conformational variability of proteins and other macromolecules, as well as the integration of sparse and low-resolution experimental data. This conformational variability is common among biological entities and is not incidental. Indeed, it has been evolutionarily selected for specific purposes. Building on our previous works, we aim to develop innovative computational techniques and approaches adapted to flexible macromolecules. The project aims to improve computational tools for small-angle X-ray (SAXS) and neutron (SANS) scattering applications, extending them for flexible molecules.

Pepsi-SAXS and Pepsi-SANS are the state-of-the-art approaches for analyzing biological profiles at small angles, developed by our team [3]. Currently, the user provides the method with the initial conformation of a protein or molecular system. However, if the initial model is far from the system in solution, or if the user lacks access to the structural model, the method is practically inapplicable. This project aims to address this limitation by proposing structural models of proteins conditioned by SAXS and SANS profiles and reconstructed using a retrained OpenFold-based architecture.

Technical description: We will learn the protein deformation field to bias a precomputed pair representation of a structure predictor, such as AlphaFold2. This biased pair representation will be decoded into an atomic model using a frozen-weight structure module, and then converted into a scattering profile by applying the Pepsi model. We will then train the network parameters by minimizing a certain loss. We will utilize standard variational inference with a variational autoencoder. We will use SAXS data, widely available on the public SAXSDB repository (https://www.sasbdb.org/). We will also use synthetic scattering profiles generated from PDB structural models (https://www.rcsb.org/).

References:

[1] Jumper, John, et al. "Highly accurate protein structure prediction with AlphaFold." Nature 596.7873 (2021): 583-589.

[2] Baek, Minkyung, et al. "Accurate prediction of protein structures and interactions using a three-track neural network." Science 373.6557 (2021): 871-876.

[3] Sergei Grudinin et al. Pepsi-SAXS: an adaptive method for rapid and accurate computation of small-angle X-ray scattering profiles. https://journals.iucr.org/paper?S2059798317005745

Skills/Qualifications

We are looking for creative, passionate, and dedicated individuals with a background in applied mathematics, computer science, or computational physics. Candidates should possess exceptional skills in computer science and mathematics, along with an interest in computational chemistry or biology. Excellent oral and written communication skills, as well as strong interpersonal abilities, are essential, with English as the primary working language (knowledge of French is a plus).

Specific Requirements

A solid understanding of machine learning, PyTorch, and structural bioinformatics will be considered advantageous.