WDWill Dumm

Will Dumm

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Will Dumm has a total of 21 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Entrenchment of germline amino-acid differences in antibody affinity maturation

    by Noam Harel, Kevin Sung, Will Dumm, Mackenzie M. Johnson, David Rich, Julia Fukuyama, Hugh K. Haddox, Frederick A. Matsen

    Entrenchment — epistasis that locks in amino acid differences between homologous proteins, so each disfavors substitutions toward the other’s state — has been demonstrated along individual protein lineages over deep evolutionary time. Antibodies offer a unique system for studying entrenchment: multiple homologous germline V gene paralogs provide diverse starting points, and the rapid somatic evolution of affinity maturation generates dense phylogenies from which selection on germline-encoded residues can be inferred. Using DASM, a deep-learning model that separates selection from mutation in antibody repertoire data, we test for entrenchment across immunoglobulin heavy chain variable (IGHV) genes. We detect entrenchment at two levels of germline divergence, driven by different sources of epistasis. Within V gene families (up to ∼20% amino acid divergence), entrenched sites cluster at the borders of the complementarity-determining regions (CDRs, the antigen-binding loops) and show high germline diversity. These sites contact antigen, light chain, and the heavy chain CDR3 loop, all of which are encoded independently of the IGHV germline. This pattern is consistent with epistasis from genetically uncoupled partners. Between V gene families, at deeper levels of divergence (25–40%), entrenchment additionally includes positions in the framework scaffold distant from binding interfaces, suggesting a larger contribution from intra-heavy-chain structural constraints. Observed mutation frequencies in human repertoires corroborate these predictions where data are sufficient. Together, these results demonstrate that the rapid somatic evolution of antibodies can serve as a lens for revealing epistatic constraints acting on germline-encoded residues, including constraints imposed by genetically uncoupled partners assembled during B cell development.

  • Open Access

    Separating selection from mutation in antibody language models

    by Frederick A Matsen, Will Dumm, Kevin Sung, Mackenzie M Johnson, David H Rich, Tyler N Starr, Yun S Song, Julia Fukuyama, Hugh K Haddox

    Antibodies are encoded by nucleotide sequences that are generated by V(D)J recombination and evolve according to mutation and selection processes. Existing antibody language models, however, focus exclusively on antibodies as strings of amino acids and are fitted using standard language modeling objectives such as masked or autoregressive prediction. In this paper, we first show that fitting models using this objective implicitly incorporates nucleotide-level mutation processes as part of the protein language model, which degrades performance when predicting effects of mutations on functional properties of antibodies. To address this limitation, we devise a new framework: a deep amino acid selection model (DASM) that learns the selection effects of amino acid mutations while explicitly factoring out the nucleotide-level mutation process. By fitting selection as a separate term from the mutation process, the DASM exclusively quantifies functional effects: effects that change some aspect of the function of the antibody. This factorization leads to substantially improved performance on standard functional benchmarks. Moreover, our model is an order of magnitude smaller and multiple orders of magnitude faster to evaluate than existing approaches, as well as being readily interpretable.


  • Open Access

    Separating selection from mutation in antibody language models

    by Frederick A Matsen, Will Dumm, Kevin Sung, Mackenzie M Johnson, David Rich, Tyler Starr, Yun S Song, Julia Fukuyama, Hugh K Haddox

    Antibodies are encoded by nucleotide sequences that are generated by V(D)J recombination and evolve according to mutation and selection processes. Existing antibody language models, however, focus exclusively on antibodies as strings of amino acids and are fitted using standard language modeling objectives such as masked or autoregressive prediction. In this paper, we first show that fitting models using this objective implicitly incorporates nucleotide-level mutation processes as part of the protein language model, which degrades performance when predicting effects of mutations on functional properties of antibodies. To address this limitation, we devise a new framework: a Deep Amino acid Selection Model (DASM) that learns the selection effects of amino-acid mutations while explicitly factoring out the nucleotide-level mutation process. By fitting selection as a separate term from the mutation process, the DASM exclusively quantifies functional effects: effects that change some aspect of the function of the antibody. This factorization leads to substantially improved performance on standard functional benchmarks. Moreover, our model is an order of magnitude smaller and multiple orders of magnitude faster to evaluate than existing approaches, as well as being readily interpretable.

  • Open Access

    larch: mapping the parsimony-optimal landscape of trees for directed exploration

    by Mary Barker, Ognian Milanov, Will Dumm, Dave Rich, Yatish Turakhia, Frederick A Matsen

    Phylogenetic inference algorithms for large data sets typically return a single tree. However, there are often many optimal trees, especially when sequence data is closely related. We develop a compact representation of large collections of maximally parsimonious histories —trees with mutations mapped onto tree edges. Our C++ implementation, larch , leverages this representation for a highly parallel search algorithm. The storage component uses our history DAG structure to compactly represent large families of optimal trees. The search algorithm integrates this storage with matOptimize for rapid tree optimization; the DAG structure allows us to accept thousands of conflicting tree rearrangements in parallel. The integration enables a new type of tree search: one that systematically maps out the collection of good trees, enabling moves that are directed away from the current set of optimal trees to cross valleys and increase the diversity of the set of optimal trees. It is able to identify more parsimonious trees than are found by other methods. We find diverse optimality landscapes for viral datasets, including many distinct plateaux. We also find that our implementation produces similar results whether using a variety of single starting trees or an ensemble of starting trees, indicating effective global optimization.


  • Open Access

    Separating selection from mutation in antibody language models

    by Frederick A. Matsen, Will Dumm, Kevin Sung, Mackenzie M. Johnson, David Rich, Tyler Starr, Yun S. Song, Julia Fukuyama, Hugh K. Haddox

    Antibodies are encoded by nucleotide sequences that are generated by V(D)J recombination and evolve according to mutation and selection processes. Existing antibody language models, however, focus exclusively on antibodies as strings of amino acids and are fitted using standard language modeling objectives such as masked or autoregressive prediction. In this paper, we first show that fitting models using this objective implicitly incorporates nucleotide-level mutation processes as part of the protein language model, which degrades performance when predicting effects of mutations on functional properties of antibodies. To address this limitation, we devise a new framework: a Deep Amino acid Selection Model (DASM) that learns the selection effects of amino-acid mutations while explicitly factoring out the nucleotide-level mutation process. By fitting selection as a separate term from the mutation process, the DASM exclusively quantifies functional effects: effects that change some aspect of the function of the antibody. This factorization leads to substantially improved performance on standard functional benchmarks. Moreover, our model is an order of magnitude smaller and multiple orders of magnitude faster to evaluate than existing approaches, as well as being readily interpretable.

  • Open Access

    A Sitewise Model of Natural Selection on Individual Antibodies via a Transformer–Encoder

    by Frederick A Matsen, Kevin Sung, Mackenzie M Johnson, Will Dumm, David Rich, Tyler N Starr, Yun S Song, Philip Bradley, Julia Fukuyama, Hugh K Haddox

    During affinity maturation, antibodies are selected for their ability to fold and to bind a target antigen between rounds of somatic hypermutation. Previous studies have identified patterns of selection in antibodies using B cell repertoire sequencing data. However, these studies are constrained by needing to group many sequences or sites to make aggregate predictions. In this paper, we develop a transformer–encoder selection model of maximum resolution: given a single antibody sequence, it predicts the strength of selection on each amino acid site. Specifically, the model predicts for each site whether evolution will be slower than expected relative to a model of the neutral mutation process (purifying selection) or faster than expected (diversifying selection). We show that the model does an excellent job of modeling the process of natural selection on held out data, and does not need to be enormous or trained on vast amounts of data to perform well. The patterns of purifying vs diversifying natural selection do not neatly partition into the complementarity-determining vs framework regions: for example, there are many sites in framework that experience strong diversifying selection. There is a weak correlation between selection factors and solvent accessibility. When considering evolutionary shifts down a tree of antibody evolution, affinity maturation generally shifts sites towards purifying natural selection, however this effect depends on the region, with the biggest shifts toward purifying selection happening in the third complementarity-determining region. We observe distinct evolution between gene families but a limited relationship between germline diversity and selection strength.


  • Open Access

    Thrifty wide-context models of B cell receptor somatic hypermutation

    by Kevin Sung, Mackenzie M Johnson, Will Dumm, Noah Simon, Hugh Haddox, Julia Fukuyama, Frederick A Matsen

    Somatic hypermutation (SHM) is the diversity-generating process in antibody affinity maturation. Probabilistic models of SHM are needed for analyzing rare mutations, for understanding the selective forces guiding affinity maturation, and for understanding the underlying biochemical process. High throughput data offers the potential to develop and fit models of SHM on relevant data sets. In this paper we model SHM using modern frameworks. We are motivated by recent work suggesting the importance of a wider context for SHM, however, assigning an independent rate to each k-mer leads to an exponential proliferation of parameters. Thus, using convolutions on 3-mer embeddings, we develop “thrifty” models of SHM of various sizes; these can have fewer free parameters than a 5-mer model and yet have a significantly wider context. These offer a slight performance improvement over a 5-mer model, and other modern model elaborations worsen performance. We also find that a per-site effect is not necessary to explain SHM patterns given nucleotide context. Also, the two current methods for fitting an SHM model — on out-of-frame sequence data and on synonymous mutations — produce significantly different results, and augmenting out-of-frame data with synonymous mutations does not aid out-of-sample performance.

  • Open Access

    Leveraging DAGs to improve context-sensitive and abundance-aware tree estimation

    by Will Dumm, Duncan Ralph, William DeWitt, Ashni Vora, Tatsuya Araki, Gabriel D. Victora, Frederick A. Matsen Iv

    The phylogenetic inference package GCtree uses abundance of sampled sequences to improve the performance of parsimony-based inference, using a branching process model. Our previous work showed that GCtree performs competitively on B-cell receptor data, compared with other similar tools. In this article, we describe recent enhancements to GCtree, including an efficient tree storage data structure that discovers additional diversity of parsimonious trees with negligible additional computational cost. We also describe a suite of new objective functions that can be used to rank these trees, including a Poisson context likelihood function that models sequence evolution in a context-sensitive way. We validate these additions to GCtree with simulated B-cell receptor data, and benchmark performance against other phylogenetic inference tools. This article is part of the theme issue ‘"A mathematical theory of evolution": phylogenetic models dating back 100 years’.


  • Open Access

    Thrifty wide-context models of B cell receptor somatic hypermutation

    by Kevin Sung, Mackenzie M. Johnson, Will Dumm, Noah Simon, Hugh Haddox, Julia Fukuyama, Frederick A Matsen

    Somatic hypermutation (SHM) is the diversity-generating process in antibody affinity maturation. Probabilistic models of SHM are needed for analyzing rare mutations, for understanding the selective forces guiding affinity maturation, and for understanding the underlying biochemical process. High throughput data offers the potential to develop and fit models of SHM on relevant data sets. In this paper we model SHM using modern frameworks. We are motivated by recent work suggesting the importance of a wider context for SHM, however, assigning an independent rate to each k-mer leads to an exponential proliferation of parameters. Thus, using convolutions on 3-mer embeddings, we develop “thrifty” models of SHM of various sizes; these can have fewer free parameters than a 5-mer model and yet have a significantly wider context. These offer a slight performance improvement over a 5-mer model, and other modern model elaborations worsen performance. We also find that a per-site effect is not necessary to explain SHM patterns given nucleotide context. Also, the two current methods for fitting an SHM model — on out-of-frame sequence data and on synonymous mutations — produce significantly different results, and augmenting out-of-frame data with synonymous mutations does not aid out-of-sample performance.


  • Open Access

    Representing and extending ensembles of parsimonious evolutionary histories with a directed acyclic graph

    by Will Dumm, Mary Barker, William Howard-Snyder, William S. DeWitt III, Frederick A. Matsen Iv

    In many situations, it would be useful to know not just the best phylogenetic tree for a given data set, but the collection of high-quality trees. This goal is typically addressed using Bayesian techniques, however, current Bayesian methods do not scale to large data sets. Furthermore, for large data sets with relatively low signal one cannot even store every good tree individually, especially when the trees are required to be bifurcating. In this paper, we develop a novel object called the “history subpartition directed acyclic graph” (or “history sDAG” for short) that compactly represents an ensemble of trees with labels (e.g. ancestral sequences) mapped onto the internal nodes. The history sDAG can be built efficiently and can also be efficiently trimmed to only represent maximally parsimonious trees. We show that the history sDAG allows us to find many additional equally parsimonious trees, extending combinatorially beyond the ensemble used to construct it. We argue that this object could be useful as the “skeleton” of a more complete uncertainty quantification.