Niklas Rindtorff, MD

I am a PhD student working with Alexander Tong and Michael Bronstein at Aithyra.

Previously, I developed image-based drug testing platforms for patient-derived cancer models — both organoids and primary cells — with Michael Boutros and Jesse Boehm, respectively.

I hold an MD, summa cum laude, from Heidelberg University and a Masters in Biomedical Informatics from Harvard Medical School.

I am working towards a future where computational and cellular models make the discovery of new medicines dramatically faster, cheaper, and more accessible.

Strong Stochastic Flow Maps

Diffusion Models, such as Alphafold 3, are powerful step-wise generative models. Flow Maps are a new, powerful class of models that aim to improve them by speeding up the sampling process by taking fewer steps. However, this has usually come at the cost of reduced stochasticity during sampling. Here we introduce strong stochastic flow maps, an approach that allows us to maintain the pathwise stochasticity of the diffusion SDE. For example, Strong Stochastic Flow Maps allow us to sample diverse, high quality conformers of the protein Chignolin with just 8 steps, while comparable diffusion models require up to 100 steps for similar fidelity.

Cycle-0: An Inversion-based Cyclic Peptide Generator

Cyclic peptides (ca. 150 heavy atoms) sit between small molecules (ca. 15 heavy atoms) and proteins (ca. 1500+ heavy atoms) in medicinal chemical space. Given their size and composability they are tractable molecules for both physical simulations, and generative modeling approaches. Recently, model-inversion based generative design methods, such as Bindcraft, have enabled the design of high-affinity protein-based binders. Can a similar approach be used to design cyclic peptides, and does the smaller size of these binders allow us to perform higher fidelity simulations of their binding properties in-silico, thus improving the design proccess? In this project, I developed Cycle0, an inversion based design approach for cyclic peptides which outperformed current diffusion-based generative modelling methods in in-silico designability metrics.

A Biologically Plausible Benchmark for Contextual Bandit Algorithms in Precision Oncology Using in vitro Data

Assigning cancer patients to different treatment options based on the genetic makeup of their tumor can be understood as a policy learning problem. What methods allow us to find strong policies while minimizing regret? In this small project, which I presented at a NEURIPS 2019 workshop, we created a benchmark dataset for treatment policy learning from in vitro cancer cell line drug sensitivity data.

The drug-induced phenotypic landscape of colorectal cancer organoids

Over the last century, we've slowly accumulated a collection of ca. 1000 publicly available cancer cell lines. Recently, new approaches such as 3D patient-dervied organoids have lowered the cost of establishing new in-vitro cancer models considerably, putting us on track to double the number of cancer models within the decade. But what determines the structural organisation of these multicelluar models? In this project, we identified two distinct molecular factors, IGFR1 and Wnt-signaling, which can be manipulated to control cancer organoid structure and modify treatment response. These factors can guide future therapeutic discovery and tissue engineering.