Mots-Clés

protein language models explainability protein structure prediction interpretability transformers protein evolution

Description

Why do protein structure predictors confidently produce nonsense, and can we see it coming?

You will lead the first systematic explainability study of protein language model failures, using hallucinated predictions with physically grounded, residue-level quality labels. You will also contribute to conceive novel flow-matching-based protein structure predictors.

About the Project

Protein structure prediction has been transformed by deep learning, yet two fundamental problems remain unsolved.
• Deep-learning structure predictors hallucinate on about 5–10% of their targets – they confidently produce physically implausible outputs whose failure modes are poorly understood and almost entirely opaque [1].
• pLMs rely heavily on the evolutionary signal they absorbed during training: when this signal is weak or missing, as for evolutionarily diverged and de novo proteins, they fall back into hallucinated patterns [2]. This project attacks both problems.

Track 1. On the interpretability side, we are compiling FLoPS, a curated dataset of pLM hallucinations scored with abstraqt [1], an ad-hoc scoring function which provides a physically grounded quality signal at the residue and contact level. Using this dataset, we want to develop explainability methods that identify which input features, attention patterns, and latent representations drive hallucination events, thus contributing to the first systematic XAI study of pLM failure modes grounded in physical plausibility rather than held-out structural benchmarks.

Track 2. On the generative modelling side, we are developing a novel flow matching architecture for protein folding that operates with little or no evolutionary information (homology, coevolution). The model will learn a mixture of physically motivated velocity fields (for secondary structure formation, hydrophobic collapse, and side-chain adjustment) with learned gat-
ing functions that capture the correct hierarchical order of folding events. Training data will consist of kinetically annotated all-atom trajectories generated by enhanced sampling methods (Self-Consistent Path Sampling, Weighted Ensemble) across the CATH fold space. The postdoctoral researcher appointed to this position will lead the XAI track as their primary responsibility,and will have the opportunity to significantly contribute to the flowm atching track as an exploratory, high-risk/high-gain secondary activity, depending on interest and background.

[1] E. Sarti and F. Cazals. “Fold or flop: quality assessment of AlphaFold predictions on whole proteomes”. In: Bioinformatics Advances 6.1 (2026), vbag190. doi: 10.1093/bioadv/vbag190.
[2] E.Seçkin, D.Colinet, E. G. J. Danchin, and E. Sarti.“Evaluating transformer-based models for structural characterization of orphan proteins”. In: Bioinformatics Advances (2026), vbag215. doi: 10.1093/bioadv/vbag215.

What We Offer

• Contract: 12 months, with the possibility of renewal for a further 12 months.
• Start date: 4 January 2027, with 1-2 months flexibility if necessary.
• Salary: from 2200€ per month, depending on experience.
• Compute: access to Grid’5000, the French national testbed for large-scale computing, plus local resources.
• Location: Inria Sophia Antipolis, near Nice and Antibes on the Mediterranean coast.

Requirements

Essential

• PhD in machine learning, computational biology, bioinformatics, biophysics, or a closely related field, completed or imminent at the time of appointment.
• Strong working knowledge of deep learning, including transformer architectures and attention mechanisms, demonstrated by publications or substantial open-source contributions.
• SolidPython programmingskillsand fluency withatleast one major deep learning framework (PyTorch strongly preferred).
• Demonstrated ability to work independently, manage a research timeline, and write clearly in English.

Nice to have (you don’t need all of these)

• Prior experience with explainability or interpretability methods for large neural networks (attribution methods, probing, mechanistic interpretability).
• Familiarity with protein language models (ESM family, ProtTrans, or similar).
• Experience with generative models: diffusion, flow matching, score matching, or variational autoencoders applied to structured or geometric data.
• Background in statistical mechanics, molecular dynamics, or enhanced sampling, even at a foundational level.
• Experience with HPC cluster environments and GPU-accelerated training.