Rafael Enrique Velásquez Rojas
Rafael Enrique Velásquez Rojas

Universidad de los Andes · Bogotá

Rafael Velásquez

Physicist and biomedical engineer building generative models for single-cell biology.

I am finishing an M.Sc. in Biomedical Engineering under Pablo Arbeláez, after undergraduate degrees in Physics and in Biomedical Engineering. My work runs from the stochastic physics of gene regulation to flow matching and equivariant graph models for molecules, materials and medical images.

Rafael Enrique Velásquez Rojas
Group
Biomedical Computer Vision, Uniandes
Graduating
April 2027
Seeking
PhD positions, autumn 2027

Four cells, one promoter. Bursty transcription with first-order decay, simulated.

Research

How much of what a cell does is written in its fluctuations

My undergraduate physics thesis argued that the variance between genetically identical cells carries functional regulatory information rather than noise to be averaged away. Modern single-cell machine learning keeps rediscovering the same quantity from the data side, through within-cell-type correlation structure. The two literatures barely cite each other.

Building the generative machinery that can actually read that information is the work I want to do for the next five years.

Generative models and flow matching

Equivariant graph neural networks, flow matching over structured scientific data, and simulation-based inference for mechanistic models.

Single-cell multi-omics

scRNA-seq, scATAC-seq and multiome data, enhancer to gene regulatory links, and the methods that make them learnable.

Biophysics and information theory

Stochastic dynamical systems, stochastic gene expression, and information-cost trade-offs in regulatory network motifs.

Biomedical vision

Topology-aware segmentation of tubular structures, with losses and metrics that score connectivity instead of volume overlap.

Machine learning for matter

Crystal structure prediction, interatomic potentials, and benchmarks that measure what a DFT practitioner actually cares about.

Publications and preprints

Selected work

2026

OPAL-Bench: a multi-metric benchmark for crystal structure prediction

NeurIPS 2026, Datasets and Benchmarks Track

Crystal structure prediction scored across several metrics at once rather than collapsed into a single number, released together with a flow matching model built on an equivariant transformer.

AcceptedBenchmark and method
2026

Beyond Volume Overlap: Surface Matching for Topology-Aware Coronary Artery Segmentation

STACOM Workshop, MICCAI 2026

A bipartite surface metric and a surface distance loss with explicit removal and creation terms, plus instance weighting by inverse squared radius so that thin distal branches count. Evaluated on nnU-Net, SwinUNETR and NexToU over ImageCAS and ASOCA.

PublishedFirst author
2026

Predicting When Cheap Simulators Break Neural Inference of Bursty Gene Expression

LatinX in AI Workshop, NeurIPS 2026, Sydney

When an approximate simulator is good enough to train a neural posterior estimator, and when the model quietly inherits the simulator's error instead of the biology.

AcceptedPoster
2026

Information-cost trade-offs as a possible contribution to the feed-forward loop census

Physical Review E

Why feed-forward loops are over-represented in transcriptional networks, read through the information a motif transmits against the cost it pays to transmit it, under stochastic gene expression. Grew out of my undergraduate physics thesis with Juan Manuel Pedraza.

Under reviewFirst author
2025

ESCAPE: a multilabel benchmark for antimicrobial peptide classification

NeurIPS 2025, poster

Multilabel classification of antimicrobial peptides, with a dataset and evaluation protocol that treat overlapping activity classes as the default case rather than an edge case.

PublishedSecond author

In progress

Current projects

Ongoing

Medical Tubathlon

A segmentation decathlon for treelike anatomy

Ten aggregated datasets spanning vessels, airways and other tubular structures across body regions, aimed at a generalist 3D segmentation model and at an evaluation that rewards connectivity rather than voxel agreement.

Ongoing

GLaMP

Drug discovery as next geometric token prediction

Binding affinity prediction cast as autoregressive generation over geometric tokens, with roughly a 20 percent MSE improvement over the BindingDB baseline. Presented as a poster, manuscript in preparation.

Background

Education, positions and service

Current

Research Projects Professional

Biomedical Computer Vision group, Universidad de los Andes

Machine learning for biomedical and biological data under Pablo Arbeláez, across generative modelling, graph neural networks and medical image segmentation.

3 years

BIOS, AI for Science subgroup

Universidad de los Andes

An interdisciplinary team of physicists, chemists, biologists, microbiologists and biomedical and electrical engineers working on machine learning for the natural sciences.

Research stay

Mitacs Globalink research intern

Dalhousie University, Canada

Dynamics of ageing with Professor Andrew Rutenberg, modelling how damage accumulates across networks of interacting health deficits.

2027 (exp.)

M.Sc. in Biomedical Engineering

Universidad de los Andes, artificial intelligence focus

Coursework and degree requirements complete in December 2026, graduation in April 2027. Advisor: Pablo Arbeláez.

B.Sc. ×2

Physics, and Biomedical Engineering

Universidad de los Andes

Physics thesis on information-cost trade-offs in genetic feed-forward loops, supervised by Juan Manuel Pedraza.

Teaching

Computer vision instructor

Universidad de los Andes

I prepare and teach the labs for the computer vision course and supervise the students taking it.

2021

Co-founder, Network Andes AI

Student chapter

Founded with close friends to widen access to AI knowledge on campus, bringing back alumni who had started deep tech companies to talk to students who had no other route to them.

Side work

Biosensors and early-stage ventures

Microplastics detection, speech translation, neonatal screening

Co-founded a systems and synthetic biology team designing a biosensor for microplastics in water, and made two attempts at building products out of research: real-time speech to speech translation, and a web application for screening neonatal jaundice.

Contact

Get in touch

Email is the fastest way to reach me. I am glad to talk about stochastic gene expression, flow matching, or whether a benchmark measures the thing it claims to measure.

Email