Alan Murphy

Alan Murphy

he/him

Postdoctoral Research Scientist at CSHL | Deep learning for genomics

Alan Murphy

Computational Biologist | Postdoctoral Researcher, Cold Spring Harbor Laboratory, New York

   

Hi, I'm Alan Murphy, a computational biologist conducting my postdoctoral research in Dr. Peter Koo's lab at Cold Spring Harbor Laboratory. I build sequence-to-function (seq2func) models of gene regulation — models that map DNA sequence to measured regulatory activity — and I work on making them reliable where they are currently weakest: on genetic variants, on designed regulatory elements, and in the cell types where disease acts.

Most variants associated with common disease fall outside protein-coding sequence, where their effect on gene regulation is usually unknown, and the same sequence behaves differently from one cell type to the next. Models that read both the sequence and the cell offer a route to interpreting those variants and to designing regulatory elements active in a single cell type — but only if their predictions hold outside the data they were trained on. Establishing where they do, extending them where they do not, and using them to design elements for the cell types where disease acts is the programme I will build my laboratory around.

Beyond research, I co-created and serve as editor for the Genomics × AI blog, a growing community hub highlighting methods and applications at the intersection of genomics and machine learning. This initiative reflects my dedication to scholarly leadership and knowledge sharing in computational biology.

 

Research Programme

Two obstacles stand between a disease locus and a mechanism: which variant is causal, and which cell type it acts in. Seq2func models could resolve both — a variant's effect is the difference between scoring its two alleles, and a promising regulatory element is one the model scores highly in the target cell type — yet they are least reliable in exactly these settings. Two causes account for that, and my laboratory will be organised around measuring and correcting each: the sequence, because a rare variant or a designed element is unlike anything in the training data, and the cell, because an element is only active where the regulators that drive it are present.

Aim 1

Establish where regulatory models can be trusted

Accuracy is reported as a single number over held-out regions of the genome, which says little about a variant in the cell type where a disease acts. I will measure how accuracy falls across classes of sequence and as cellular context moves away from a model's training data. The result is a public resource reporting the accuracy to expect for a given model, sequence and cell type.

Aim 2

Correct models with perturbation data

Genome-wide profiling shows which sequence features accompany regulatory activity; perturbation experiments show what changing one of them actually does. My laboratory will refine models on the full body of published reporter and CRISPR interference screens, and establish whether a correction learned in one cell type transfers to the cell types where disease acts.

Aim 3

Design regulatory elements selective for a chosen cell type

Gene therapies can restrict a transgene to one cell type through the regulatory element placed in front of it, but candidates must be narrowed down long before anything is synthesised. I will build models that score candidate elements across hundreds of cell types, so an element can be designed for one and screened against all the others, then validated experimentally.

Throughout, the models, datasets and evaluation protocols behind this work are released openly. The longer goal is a model that can be trusted on a variant nobody has assayed, in a cellular state nobody has screened.

 

CV Google Scholar Genomics × AI Blog

 

News

  • 2026-07-27

    Our review, Toward generalizable and interpretable AI in regulatory genomics, is out in Nature Genetics! Sequence-to-function models do well on held-out genomic regions but generalise inconsistently across genetic variation and cellular contexts. We ask why, and argue a way forward with targeted perturbation experiments feeding back into iterative model refinement.

  • 2026-06-25

    Published a blog post benchmarking sequence-to-function models against CRISPRi enhancer-knockdown screens on the Genomics x AI blog, showing that leading models (including AlphaGenome and Borzoi) systematically underestimate distal enhancer effects, with the gap widening as enhancers get farther from their target genes.

  • 2026-05-07

    Presented Improving genomic deep learning models with perturbation data, our work on continual learning for genomics, at Biology of Genomes 2026, Cold Spring Harbor Laboratory.

  • 2026-04-08

    Co-authored a research highlight in Cell Research on AlphaGenome and predicting non-coding variant effects, with Masayuki Nagai and Peter Koo.

  • 2026-02-20

    Published a blog post on fine-tuning AlphaGenome for MPRA and STARR-seq data on the Genomics x AI blog, highlighting the approach of treating sequence-to-function models as modular regulatory encoders leading to state-of-the-art results across perturbation assays.

  • 2025-10-10

    Presented our work on causal refinement for genomic deep learning models through continual learning at MLCB 2025. Check out the talk on the MLCB YouTube channel.

  • 2025-03-03

    Joined Peter Koo’s group at Cold Spring Harbor Laboratory, New York to develop methods to improve genomic sequence-to-function and genomic language models.

  • 2024-12-11

    Our ChromExpress paper investigating the relationship of histone marks with expression using deep learning is out in Nucleic Acids Research (NAR)! See a quick overview on X/Twitter.

  • 2024-11-16

    Our Enformer Celltyping paper, a genomic DNN to accurately predict epigenetic signals in previously unseen cell types, is out in Nature Communications! See more on Twitter or on BlueSky.

  • 2023-12-04

    Our re-analysis paper of the first single-cell RNA-seq Alzheimer’s disease dataset is out in eLife! Check out an overview here.

  • 2023-09-25

    Presented my PhD work predicting the cell type-specific effects of genetic variants on the epigenome at the Kipoi Summit for computational regulatory genomics.

  • 2023-07-29

    Presented a session on single-cell genomics for Alzheimer’s disease as part of ADDI’s Summer Learning Series.

  • 2022-12-22

    Our paper benchmarking differential expression methods for single-cell RNA-seq is out in Nature Communications! Check out our overview here.

  • 2021-10-02

    MungeSumstats, our software for rapid standardisation and quality control of GWAS or QTL summary statistics, is now out in Bioinformatics.

  • 2021-07-19

    Thrilled to officially start my PhD with Dr. Nathan Skene’s group in the Department of Brain Sciences, Imperial College London as part of the UK DRI.

Selected Publications

This is a selection of recent work. For a complete and always up-to-date list, see my Google Scholar profile.