S. Joshua Swamidass, MD PhD

Research

Machine learning for medicine, chemistry and biology

Our group applies deep learning and computational modeling to problems at the intersection of chemistry, biology and medicine, from medicinal chemistry and drug safety to cancer genomics, image analysis and clinical studies. A hallmark of our work is close collaboration with experimental and clinical scientists to produce useful tools that enable more efficient discovery.

Drug metabolism and toxicity

Many drugs fail, or are withdrawn, because the liver converts them into reactive metabolites that bind to proteins and DNA. Our group builds models that predict where on a molecule metabolism occurs, which reaction takes place, and whether the resulting metabolite is likely to be reactive.

This work began with XenoSite, a neural-network model of cytochrome P450 sites of metabolism. We showed that these models can be trained from ambiguous, region-level data where the exact site of metabolism is unknown, and the work grew to cover epoxidation, quinone formation, N-dealkylation, UGT conjugation, glutathione and macromolecule reactivity, and all of Phase I metabolism in a single model. More recent work chains these predictions together to infer full metabolic pathways and the structures of the metabolites they produce. Close collaborations with experimental pharmacologists test these predictions on drugs such as terbinafine, meloxicam and the NSAIDs. The XenoSite models are used by pharmaceutical companies to guide the design of safer medicines.

All 30 papersXenoSite web tools

Diagram: cytochromes P450 convert carbamazepine to carbamazepine-10,11-epoxide, which binds protein to form a hapten complex that triggers an immune response and adverse drug reactions.
How a reactive metabolite causes harm: carbamazepine is epoxidized, binds protein, and can trigger hypersensitivity.Hughes, Miller & Swamidass, ACS Cent Sci 2015, Fig. 1 · ACS AuthorChoice
Four molecules with predicted reactive sites highlighted, each shown reacting with cyanide, DNA, glutathione or protein.
The four kinds of reactivity the XenoSite model predicts: with cyanide, DNA, glutathione and protein.Hughes et al., ACS Cent Sci 2016, Fig. 1 · ACS AuthorChoice

Key papers

Deep learning of molecular structure

Predicting the behavior of small molecules and proteins requires representations that capture their underlying physics. We develop deep learning methods that learn these representations directly, rather than relying on hand-crafted descriptors.

Examples include a local-variable model of aromatic and conjugated systems, coordinate-free networks that reproduce quantum-chemical calculations, site-level bioactivity predictions built on learned quantum chemistry, and models that identify the structural determinants of protein properties by comparing structural ensembles.

All 5 papers

A polycyclic hydrocarbon and two porphyrins colored by aromatic system size, with error maps comparing Weave and wave network models.
Wave networks learn the size of aromatic and conjugated systems with less error than convolutional models.Matlock, Dang & Swamidass, ACS Cent Sci 2018, Fig. 3 · ACS AuthorChoice

Key papers

Chemical informatics

My early work, in Pierre Baldi’s group at UC Irvine, applied machine learning to predict the biological properties of small molecules. This included one of the first applications of kernel methods and support vector machines to chemical data, a widely cited study that helped spark a resurgence of chemical informatics within the bioinformatics community, and Influence Relevance Voting, an accurate and interpretable neural-network approach to virtual screening.

This work also produced algorithms for fast, exact search of chemical fingerprints, including the first sub-linear search algorithm, the first algorithmic progress in decades on this problem. It also produced ChemDB, one of the early public databases of commercially available chemicals, which together with ZINC and ChemBank renewed interest in chemical data and helped lead to PubChem. Later work examined how chemical data can be shared without revealing structures, and how to measure early retrieval performance (CROC).

All 18 papers

Key papers

Screening and drug repurposing

Working with the screening team at the Broad Institute, we used economic theory to design more efficient high-throughput screens. Explicitly encoding costs and utility to direct confirmatory testing led to large efficiency gains, at times tripling the rate at which new active scaffolds were found.

Later work, funded first by Pfizer and then by GlaxoSmithKline, built software that used historical screening data to direct internal drug repurposing projects. It also addressed practical problems in screening data: missing measurements, hidden workflows in PubChem, and combining phenotypic with target-based assays. This work also produced the Scaffold Network Generator, an open-source tool for mining molecular collections.

All 13 papers

Key papers

Genomics and bioinformatics

Sequencing studies identify thousands of mutations in each tumor, but only a few drive the disease. We developed methods to distinguish drivers from passengers. These include a parsimony-guided, unsupervised approach to detecting driver mutations and a statistical test that identifies tumor suppressors and oncogenes from pan-cancer data.

With collaborators at the McDonnell Genome Institute, we also built deep learning tools that automate the refinement of somatic variant calls and improve neoantigen prediction by accounting for nearby variants. Other collaborations apply machine learning to gene expression, transcription factor binding and the microbiome. Related work outside genomics includes cryptosets, a way to securely measure the overlap between private datasets.

All 13 papers

Key papers

AI in pathology

Our work has expanded into image analysis, where deep learning can guide critical clinical decisions. Pathologists often must decide, from a frozen section and under time pressure, whether a donor kidney or liver can be transplanted. We build deep learning systems that quantify these features consistently, including glomerulosclerosis in kidney biopsies, steatosis in donor livers, and Paneth cell density in inflammatory bowel disease, and that predict transplant outcomes.

Our group was among the first academic groups to deploy AI into a pathology workflow: the donor kidney biopsy model, developed with Joseph Gaut, is licensed to Trusted Kidney and is used in clinical practice by Mid-America Transplant.

All 7 papersTrusted Kidney

A frozen-section kidney biopsy shown three ways: H&E stained, annotated by a pathologist, and as predicted by the deep learning model, with sclerosed and non-sclerosed glomeruli in different colors.
Donor kidney frozen section: the stained slide, the pathologist’s annotation, and the model’s prediction of glomeruli.Marsh et al., JAMA Netw Open 2021, Fig. 1A · CC BY 4.0

Key papers

A donor liver biopsy shown as an H&E slide, a pathologist annotation and a model prediction of steatosis, with magnified insets.
Donor liver biopsy: model predictions of steatosis match pathologist annotations.Sun et al., EBioMedicine 2020, Fig. 4 · CC BY-NC-ND 4.0
Paneth cells in an intestinal crypt, annotated crypts, and a diagram of the two-stage deep learning pipeline that counts Paneth cells per crypt.
Counting Paneth cells in inflammatory bowel disease biopsies with a two-stage deep learning model.Kang et al., EBioMedicine 2024, Fig. 1 · CC BY-NC-ND 4.0

AI in healthcare

Electronic health records hold patterns that no single clinician can see, but models trained on them must be trustworthy before they inform care. We apply machine learning to clinical data to find drug interactions, estimate risk and guide treatment, with an emphasis on models that clinicians can question and understand.

Examples include mining a retrospective EHR cohort to identify drugs that increase the risk of liver injury when combined with NSAIDs, “conversational” machine learning that surfaced a protective association between ondansetron and hospital-acquired venous thromboembolism, deep learning estimates of kidney function, and domain-constrained neural networks that guide patient-specific anesthetic dosing. We also study how race and socioeconomic status shape medical data, and how to recognize and mitigate their effects on models.

All 9 papers

Diagram of the drug interaction network: a candidate drug such as diclofenac and a patient’s other drugs feed independent-risk and interaction terms that together predict drug-induced liver injury.
A drug interaction network learned from health records separates a drug’s own liver-injury risk from the risk it adds in combination with NSAIDs.Datta et al., PLoS Comput Biol 2021, Fig. 1A · CC BY 4.0

Key papers

Research support

Our work has been supported by the National Institutes of Health (NLM, NIGMS, NIDDK, NCI and NIMH), by industry partners and by transplant organizations.

  • Systematic discovery of bioactivation-associated structural alerts (NIGMS R01 GM140635; co-PI with Grover Miller)
  • Data and tools for modeling metabolism and reactivity (NLM R01 LM012222; co-PI with Grover Miller)
  • Computationally modeling the impact of ontogeny on drug metabolic fate (NLM R01 LM012482; co-PI with Grover Miller)
  • Paneth cell phenotype as a predictive biomarker for ulcerative colitis (NIDDK R01 DK124274; co-investigator, PI Ta-Chiang Liu)
  • Informatics tools for identification, prioritization, and clinical application of neoantigens (NCI U01 CA248235; co-investigator, PI Malachi Griffith)
  • A deep learning model to improve pathologist interpretation of donor kidney biopsies (NIDDK STTR R41/R42, with PlatformSTL; co-investigator with Joseph Gaut)
  • Early life adversity, biological embedding, and risk for developmental precursors of mental disorders (NIMH R01 MH113883; co-investigator, PI Joan Luby)
  • Deep convolutional neural networks to enhance pathologic assessment of donor organs (Midwest Transplant Services)
  • Industry research support from Pfizer and GlaxoSmithKline for drug repurposing from screening data