Laurits Fredsgaard Larsen
I develop, train, and evaluate machine-learning models, with recent work on molecular representation learning and bioassay-supervised pretraining.
I am a postdoctoral researcher at DTU Compute through the end of 2026. My PhD thesis is submitted, and I am awaiting my defense.
I am interested in ML scientist and research engineer roles in Copenhagen.
Model contribution
I optimized and ran Atom-JEPA-Bio’s second-stage training on LUMI’s AMD GPUs, using mixed precision and custom ROCm/Triton kernels. Combining self-supervised learning with experimental bioassay supervision gave state-of-the-art ADMET prediction performance.
International Conference on Machine Learning (ICML).
Changing a molecule can change its properties. We use agreement on unlabelled molecules to learn without augmentations, with individual models often outperforming conventional ensembles.
AISTATS Workshop: Towards Trustworthy Predictions — Theory and Applications of Calibration for Modern AI.
Using AutoGraph’s SENT encodings ↗, transformers learn to generate molecular graphs as strings. On QM9, we found that comparing likelihoods across equivalent strings provides a better quality signal than generation likelihood alone.
AIDaR Workshop at NeurIPS. Camera-ready version forthcoming.
Across 22 ADMET datasets, we compare eleven calibrated prediction pipelines for accuracy and uncertainty estimates. Neural ensembles perform more consistently across tasks, while LightGBM leads on many individual assays.
Transactions on Machine Learning Research (TMLR).
Are we training deep ensembles the right way? Stopping based on joint performance let members train longer and improved ensemble accuracy and predictive likelihood in our experiments.
AXIOM: Foundations of Efficient Deep Learning — NeurIPS Workshop. Camera-ready version forthcoming.
I developed ICICLE for resource-efficient hyperparameter tuning, building on LC-PFN ↗ and FT-PFN ↗ with a new embedding for more hyperparameters. It integrates with Ray Tune for parallel trials and found better configurations within fixed training budgets on QuickTune.
Research software
I designed a training setup that keeps ensemble members independent across GPUs while making their joint performance easy to track.
Vaccines.
I combined vaccine experiments with quantitative analysis of IgG responses, comparing AP205 and HPV particle backbones and covalent versus affinity-based antigen attachment. This work formed the basis of my MSc thesis.
Nature Communications.
As part of the vaccine development work, I cloned, expressed, and purified candidate Spike and receptor-binding domain (RBD) antigens using a baculovirus–insect cell system.
At LEO Pharma, I extended an existing omics application with mappings among transcripts, genes and proteins, implemented views requested by scientists, and set up RNA-seq analysis pipelines.
For DaSSCo’s digitisation of natural history specimens ↗, I created a Python package for printable Data Matrix labels ↗ for pinned insects and extended existing camera-control and imaging software.
I scaled multi-node training of NequIP and Allegro interatomic potentials on LUMI. I also built boltz-lumi ↗, which adds ROCm kernels for Boltz-2 protein-ligand structure prediction on AMD GPUs.
BSc in Machine Learning & Data Science, and MSc and BSc in Molecular Biomedicine, University of Copenhagen.
My undergraduate projects with Wouter Boomsma covered protein inverse folding with a 3D U-Net and mutation-effect prediction with DeepSequence-inspired variational autoencoders.
My submitted PhD thesis is titled “Uncertainty and Evaluation in Molecular Machine Learning”.
Teaching assistant at DTU: