LMRL will be at

ICLR 2025

in Singapore

Learning Meaningful Representations of Life (LMRL) Workshop at ICLR 2025

Since the last time that the LMRL workshop was held at NeurIPS 2022, interest in representation learning for biology has surged, with new ideas challenging traditional approaches and sparking discussions on how best to capture the complexity of biological systems through machine learning. The availability of large-scale public DNA and RNA sequencing, protein sequences and 3D structures, mass spectrometry, and cell painting datasets (JUMP-CP, RxRx3, Human Cell Atlas) has fueled the development of numerous large-scale “foundation models” for biological data (Rozenblatt-Rosen et al. 2021; Fay et al. 2023; Chandrasekaran et al. 2023). These models aim to extract “meaningful” representations from noisy, raw and unstructured high-dimensional data to address a variety of biological questions.

The AIxBio community has two important questions to answer: (i) what data, models and algorithms do we need to ensure that we extract meaningful representations (sufficient for their intended applications); and (ii) what are the appropriate methods for evaluating the quality of these embeddings, both in terms of the richness of information they capture, and their ability to generalize and improve performance on downstream tasks?

We believe that the early stage of this field presents a remarkable opportunity to foster discussion, collaboration, and insight sharing through our workshop on “Learning Meaningful Representations of Life”. Our agenda will encourage discussion both about new methods for representation learning in biology as well as biologically relevant & substantive evaluations to probe the generalization capabilities of the learned representations. Building upon the themes of previous years, the workshop will focus on multiple layers of biological information: genomes, molecules, cells, phenotype and beyond.

It is essential for such “meaningful representations” to not only generalize across modalities but also to capture biological information across different scales, from subcellular to multi-cellular and organism-wide processes. Harmonizing representations from molecules, proteins, cells, and tissues enables in-silico simulation of biological processes, interactions, and causal mechanisms, ultimately building towards a foundation model of AI-powered virtual cell (Bunne et al. 2024) , i.e. universal simulators of cellular function and behavior.

For the LMRL workshop at ICLR 2025, our objectives are (i) to convene those engaged in learning representations within and across different modalities of biological data, (ii) to discuss cutting-edge methods for assessing and measuring the significance of learned biological representations, (iii) to create a platform for developing open-source standardization of datasets and evaluation metrics for benchmarking new methods, and (iv) to envisage potential real-world problems that could be solved with improved strategies for learning meaningful representations of life.

Call for papers

The LMRL Workshop returns to ICLR 2025 to foster discussion and collaboration in the growing field of representation learning for biological data. With the increasing availability of large-scale biological datasets—spanning genomics, proteomics, cell imaging, and more—the development of innovative machine learning methods to extract and evaluate meaningful representations has never been more critical. This year, we aim to bring together researchers at the forefront of AI and biology to address two key questions:

What data, models, and algorithms are needed to extract meaningful biological representations that generalize well to downstream tasks?
How can we evaluate the quality and utility of these learned representations?

We invite submissions on a wide range of topics, including but not limited to:

Foundation models for biological data
Multimodal representation learning
Multiscale representation learning to connect molecular and biological data
Generalizability and interpretability in biological datasets
Causal representation learning in biology
Active learning for experimental design
Generative models for molecular design
Modeling biological perturbations and their effects
Long-range dependency modeling in sequences and spatial omics
New datasets, benchmarks, and evaluation metrics

We welcome submissions from diverse areas of biology, including omics, protein design, cell profiling, tissue and systems biology, and beyond. Submissions don’t need to align perfectly with AI-focused conferences, as our goal is to attract interdisciplinary contributions and foster new collaborations.

Submission Tracks

1. Full Paper Track

Length: 4–8 pages (excluding references)
Scope: Comprehensive studies with clear descriptions of biological problems, AI methods, implementation details, and benchmarking metrics.

2. Tiny Paper Track

Length: 1–2 pages (extended abstract format)
Scope: Focused on work-in-progress, novel methodologies, or emerging ideas without requiring extensive results.

Review Process

Double-blind review for all tracks.
Full papers require at least one reviewer nomination; tiny paper reviewer nominations are optional.
Submissions may be moved between tracks based on reviewer recommendations.

Submission Details

All submissions require a “Meaningfulness Statement” (100 words max), to explain what the authors consider a “meaningful representation of life” and how the work contributes to this direction.

A Latex template can be downloaded here.

All submissions are non-archival but will be hosted online unless explicitly requested otherwise.