Research
My research combines probabilistic modeling, machine learning and algorithm design to make large biological data tractable. I build probabilistic and graph-based models of biological sequence collections, design algorithms that search and compress enormous sequence spaces efficiently, train machine-learning and deep-learning models where a closed-form answer is out of reach, and write simulators with full ground truth to test all of the above.
- Markov and graph models
- Data compression and string algorithms
- Information theory
- Probability and statistical modeling
- Machine learning
- Computational immunology and immune repertoire analysis
Research interests 5
01Probabilistic modeling of biological sequencesGenerative and statistical models of sequence collections, with quantities that can be computed exactly rather than estimated.
I am interested in models that assign probabilities to biological sequences and to whole collections of them, and in what those probabilities reveal: how diverse a population is, how likely a sequence is to arise, and which sequences are anomalous. My preference is for models on which the interesting quantities, such as generation probabilities, entropy and Hill diversity, follow from the model structure by dynamic programming instead of sampling or extrapolation. This is the thread that runs from LZGraphs to FlashbackGraph, where a graph built from billions of observed sequences yields exact repertoire-level statistics.
02Representing immune repertoiresAlignment-free graph models of T- and B-cell receptor repertoires, built from lossless decompositions of strings.
Compression is a lens on structure. A decomposition that is reversible loses nothing, and if it is also deterministic and context-free, sequences that share structure share paths in a graph. I study such decompositions and the graphs they induce, from Lempel-Ziv parsings to Flashback, a bilateral run-peeling decomposition with an exact run-pairing theorem, and I use information-theoretic quantities on these graphs as features, distances and diversity measures. The same ideas apply to any large collection of discrete sequences, not only to immune receptors.
03Machine learning and deep learning for sequence analysisMulti-task neural models trained on simulated ground truth, with uncertainty that can be trusted.
Where no closed form exists I train models, and I care as much about their calibration as about their accuracy. AlignAIR is the main example: a multi-task deep network that assigns germline genes and boundaries to antibody sequences and returns likelihoods that track the true probability of being right. Before that, I worked on natural-language models over source code and on machine-learning analysis of DNA-methylation data, and I bring the same toolkit of representation learning, sequence models and careful evaluation to whichever domain the problem lives in.
04Algorithm design for very large sequence spacesData structures and algorithms that make billions of sequences searchable, comparable and compressible.
Much of my work is engineering for scale: a data structure and server for membership queries against tens of thousands of repertoires (PHOENIX, filed as a US provisional patent), linear-time decompositions, and Rust and C kernels behind the Python packages. I aim for algorithms that are as efficient as the problem allows and that someone outside the field can run without reading the paper first.
05Simulation and benchmarkingSimulators with full ground truth, and community benchmarks, so that methods are compared on equal terms.
Methods are only as good as their evaluation. Because ground truth is usually unavailable in experimental biological data, I build simulators that produce it: GenAIRR generates annotated immune-receptor sequences with realistic recombination, mutation and sequencing error across 23 species. It underpins the training of AlignAIR, a systematic comparison of immunoglobulin aligners, and community benchmarks such as the Adaptive Immune Profiling Challenge, which I co-organized.