Machine Learning Intern — Stochastic Gradient Descent

Nova In Silico company; Lyon; Hybrid; EUR 8,900 - 13,000; Part time

Job summary

Nova In Silico is a health tech company in Lyon that develops an in silico clinical trial platform jinkō to simulate drug efficacy and optimize development. The internship provides significant responsibilities and a steep learning curve within a highly motivated team.

You will work on stochastic gradient estimation for QSP surrogate models, review literature, prototype algorithms, and help integrate solutions into Nova's simulation platform.

Qualifications

• Proficiency in machine learning using Python and PyTorch.
• Experience with NLME models and statistical modelling is a plus.
• Professional English communication is required.

Responsibilities

• Review the literature on relevant ML algorithms.
• Prototype the stochastic gradient algorithm under Nova's constraints.
• Evaluate benchmark cases against the SAEM alternative.
• Integrate the solutions into Nova's simulation platform.

Skills

Python

PyTorch

NLME modelling

Statistical modelling

Tools

NLME

ML in Python

Job description

Nova In Silico is a health tech company that develops an in silico clinical trial platform jinkō to simulate drug efficacy and optimize clinical development using virtual patients and disease modeling. As an innovative company, we offer a dynamic work environment distinct from larger, established organizations. Interns will gain significant responsibilities and benefit from a steep learning curve, supported by a highly motivated team.

KEYWORDS

Expectation Maximization, Gradient Descent, Non-Linear Mixed-Effects Model, Surrogate Model, PyTorch

BACKGROUND — QUANTITATIVE SYSTEMS PHARMACOLOGY AND ITS CHALLENGES

Quantitative Systems Pharmacology (QSP) is a critical discipline in modern drug development. It involves creating complex, mechanistic mathematical models that describe the dynamic interactions between a drug and a biological system. These models integrate pathophysiology and pharmacology to predict a drug's effect, safety, and efficacy across diverse patient populations. At Nova In Silico, our R&D efforts are focused on building and applying these high-fidelity QSP models.

A significant challenge arises when fitting these models to real-world clinical data. To account for variability between individuals, QSP models are often formulated as Non-Linear Mixed-Effects (NLME) models. Parameter estimation for NLME models, which is typically performed via Maximum Likelihood Estimation (MLE), is a difficult and computationally intensive task. Traditional estimation algorithms can take hours or even days to converge, creating a substantial bottleneck in the R&D pipeline.

PYTORCH-BASED SURROGATE MODELS

To address this computational bottleneck, Nova In Silico has successfully developed surrogate models for some of our key QSP models. These surrogates, built using the PyTorch deep learning framework, are lightweight, fast-to-execute approximations of the full, complex QSP models. They are designed to capture the essential input-output behavior of the original model while dramatically reducing computation time.

This speed-up has enabled us to more efficiently perform parameter estimation. Currently, we leverage our surrogate models within Expectation-Maximization (EM) type algorithms. EM is a powerful and standard method for finding maximum likelihood estimates in models with latent variables (such as the random effects in NLME models). This approach has proven effective for our existing model structures.

FLEXIBLE ESTIMATION VIA STOCHASTIC GRADIENT DESCENT

While effective, EM-type algorithms are often tailored to specific model structures and statistical assumptions. As our R&D pipeline evolves, we aim to explore more diverse and complex surrogate model architectures and apply them to various types of clinical data. The mathematical framework of EM can be restrictive in these more general cases.

Stochastic Gradient Descent (SGD) offers a compelling and flexible alternative, as these algorithms:

• Can be applied to a much broader family of models and data structures.
• Are often more computationally efficient, as they can process large datasets in small batches.
• Integrate natively with the PyTorch ecosystem, as gradient computation is the framework's core function.

OBJECTIVE

The intern will implement the stochastic approximation gradient algorithm, drawing from the principles in the reference articles, and apply it to our existing surrogate models. This will equip Nova In Silico with a novel, flexible, and powerful estimation tool, expanding our capabilities to fit next-generation QSP models to complex clinical data.

YOU ARE

• team player, a good listener, and an effective communicator
• Curious and proactive, ready to face real-life engineering challenges
• Autonomous and self-motivated with strong analytical and problem-solving skills
• Eager to learn mathematical modeling and simulations of biological systems
• Willing to explore latest advances in science and technology
• Responsive and capable of tackling time-sensitive issues with agility

YOU WILL

• Review the scientific literature on relevant machine learning algorithms
• Prototype the stochastic gradient algorithm under Nova's specific constraints
• Evaluate benchmark cases against the alternative SAEM algorithm
• Integrate solutions into Nova's simulation platform

METHODOLOGY AND TECHNICAL SKILLS

We are looking for people who know some of the following or are eager to learn and work with them:

• Machine learning in Python, PyTorch
• Statistical modeling, NLME models

A professional English level (written and oral) is required for this role.

PRACTICAL INFORMATION

• Contact: [e-mail]
• Salary: Competitive
• Start date: Flexible