Post-Doctoral Research Visit F/M Why Do Protein Structure Predictors Confidently Produce Nonsense, and Can We Anticipate These Failures?

Context
Deep learning has transformed protein structure prediction. However, current models can still produce physically implausible structures with high confidence. These failures, known as
hallucinations
, remain poorly understood and raise fundamental questions about the reliability, interpretability, and limitations of protein language models (pLMs).
This project aims to systematically characterize these failures and develop methods to understand and anticipate them. It builds on FLoPS, a dataset of hallucinated protein language model predictions, with physically grounded quality annotations generated using the abstraqt method. These annotations provide residue- and contact-level information, going beyond conventional metrics based on comparisons with reference structures.
The project comprises two complementary research directions:
Model interpretability and explainability:
identify sequence features, attention mechanisms, and internal representations associated with hallucinations, with the goal of understanding why models generate physically implausible structures.
Generative protein structure modelling:
depending on the candidate’s interests and expertise, contribute to the development of a novel flow-matching architecture for protein structure prediction, aiming to reduce reliance on evolutionary information and incorporate physically motivated mechanisms of structural formation.
Main objectives
Develop a systematic explainability methodology for protein language model failures, grounded in residue- and contact-level physical plausibility annotations.
Identify input features, internal mechanisms, and latent representations associated with hallucinations.
Investigate whether model failures can be detected or predicted from internal signals and model outputs.
Depending on the candidate’s background and scientific interests, contribute to the design of flow-matching-based generative protein structure predictors that rely less heavily on evolutionary information.
The position is based at Inria Sophia Antipolis on the French Riviera, within an interdisciplinary research environment at the intersection of machine learning, bioinformatics, and biophysics. The successful candidate will have access to local computing resources and Grid’5000, the French national testbed for experimental research in distributed computing and high-performance computing.
Position highlights:
an emerging research topic with strong scientific potential, access to physically annotated data, opportunities to develop novel explainability methods for large neural networks, and the possibility of contributing to generative protein structure modelling.
Assignment
The successful candidate will primarily lead the research track devoted to explaining hallucinations in protein language models. They will design and implement analytical methods that link model-internal mechanisms to the physical quality of predicted structures, using the FLoPS dataset and the annotations generated by abstraqt.
They will define experimental protocols, conduct quantitative and qualitative analyses, assess the robustness of the proposed methods, and contribute to the scientific interpretation of the results. Particular attention will be paid to relationships between input sequence properties, model activations and attention mechanisms, latent representations, and the different failure modes observed.
The successful candidate will oversee the scientific and technical development of their research, from experimental design to dissemination. They will contribute to scientific publications and the dissemination of methods and results, including the development of reproducible code.
Depending on their expertise and interests, they may also contribute, on an exploratory basis, to the project’s second research track: the development of flow-matching-based protein structure predictors, in collaboration with researchers working on generative modelling and molecular simulation.
Collaboration and responsibilities
Collaborate with the research team on the development, analysis, and exploitation of the FLoPS dataset.
Develop and benchmark explainability methods tailored to protein language models and physical plausibility annotations.
Design rigorous and reproducible evaluation protocols to identify mechanisms associated with hallucinations.
Contribute to interdisciplinary scientific discussions at the intersection of machine learning, bioinformatics, and biophysics.
Depending on expertise and project needs, contribute to methodological developments in flow matching and protein structure prediction.
Disseminate research outcomes through scientific publications, presentations, and, where possible, the sharing of reproducible code and resources.
References
[1] E. Sarti and F. Cazals. “Fold or flop: quality assessment of AlphaFold predictions on whole proteomes”.
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”.
Bioinformatics Advances
, 2026, vbag215.
Main activities
Data and model failure analysis:
use the FLoPS dataset and abstraqt quality annotations to characterize hallucinations at the residue and contact levels.
Explainability method development:
adapt and develop attribution methods, attention mechanism analyses, representation probing, and latent-space analysis techniques for protein language models.
Failure mechanism identification:
investigate relationships between sequence features, internal representations, and violations of physical plausibility constraints.
Hallucination detection and anticipation:
assess whether internal model signals and explainability methods can identify unreliable predictions, particularly when evolutionary information is limited.
Experimental evaluation:
design comparative experiments, define appropriate metrics, conduct statistical analyses, and assess the robustness and reproducibility of results.
Scientific computing and software development:
implement methods in Python, primarily using PyTorch, and leverage available computing resources for large-scale analyses.
Scientific dissemination:
prepare research papers, present findings at scientific meetings and conferences, and contribute to the dissemination of reproducible code and experimental protocols.
Exploratory contribution to generative modelling:
depending on the candidate’s background, contribute to the design and evaluation of flow-matching-based protein structure predictors incorporating physically motivated mechanisms of structural formation.
Skills
Fluent in English