Research

Our goal is to help scientists navigate fragmented knowledge, uncover mechanisms underlying complex system behavior, and accelerate the iterative cycle between observations, models, predictions, experiments, and new discoveries. A central theme of our work is the integration of data-driven AI with transparent, interpretable representations of biological mechanisms, enabling models whose predictions can be traced back to underlying knowledge and evidence. By combining explainable AI, large language models, knowledge graphs, graph-based reasoning, and mechanistic modeling, we create computational systems that not only capture what is known about biology, but can reason over that knowledge to explain how biological behaviors emerge. We simulate and analyze these models to predict responses to genetic, molecular, environmental, or therapeutic perturbations, identify important mechanisms and intervention strategies, and generate testable hypotheses for further investigation. 

BOHEME  - Bridging Observations to Hypotheses through Executable Model Engineering An AI-based ecosystem for engineering computable biological knowledge. It transforms scientific observations into structured, verified, and executable models that are simulated to explain biological behavior and generate new hypotheses. BOHEME replaces our previous framework DySE with a new generation of LLM-supported tools. Explore BOHEME through its web interface, where you can access and use the integrated suite of tools directly in your browser.

BioRECIPE - Knowledge Representation Schema

Biological system Representation for Evaluation, Curation, Interoperability, Preserving, and Execution

A standardized, interoperable schema for representing biological knowledge and executable models in alignment with FAIR principles. BioRECIPE captures biochemical interactions and their biological context in a structured, machine-readable format accessible to computational tools, modelers, and biological domain experts. Read more about BioRECIPE in our ACS Synthetic Biology paper and the detailed ReadTheDocs documentation. Use the BioRECIPE web interface within the BOHEME ecosystem to translate existing representations (e.g., JSON, BioPAX, and SIF) into the BioRECIPE schema, or to create your own interaction lists and models directly in BioRECIPE.

ACCORDION - Model Recommendation

ACCelerating and Optimizing model RecommenDatIONs

A model recommendation tool that evaluates and selects relevant knowledge. ACCORDION is a framework for context-aware knowledge selection and automated recommendation of executable biological models. Starting from a baseline model and candidate knowledge retrieved from literature and databases, ACCORDION uses graph-based clustering to identify relevant model extensions, then evaluates candidate models through simulation and formal analysis to recommend models that best reproduce known or desired system behavior. This enables systematic exploration of alternative mechanisms while reducing large collections of candidate interactions to those most relevant to the modeled biological context. Read more about ACCORDION in our Frontiers in Systems Biology paper, explore the detailed ReadTheDocs documentation, access the source code on GitHub, or use the Jupyter notebook to explore ACCORDION workflows and benchmark examples.

BELL - Interaction Verification and Confidence

Biomodel Evidence and LLM-based Logic

An evidence-based verification and decision-support tool for biological interactions and mechanistic models. BELL automatically cross-references interactions against multiple biological knowledge databases, evaluates the strength and consistency of the supporting evidence, and uses LLM-based reasoning to explain the results and recommend curator actions. Its transparent pipeline combines entity grounding, evidence retrieval from seven complementary databases, interpretable confidence scoring, identification of evidence gaps and contradictions, and recommendations to accept, enrich, review, or reject individual interactions. By bringing database evidence, quantitative scores, explanations, and expert annotation together, BELL supports scalable and reproducible verification of biological knowledge before it is incorporated into mechanistic models. BELL was presented at IWBDA 2026, where it received the Best Paper Award, and its functionality is now integrated directly into the KALIMBA web interface within BOHEME.

CELESTA - Relation Extraction with Context

Context Extraction through LEarning with Semi-supervised multi-Task Architecture 

An AI framework for extracting the biological context needed to interpret interactions described in biomedical text. CELESTA uses open-set semi-supervised multi-task learning to identify contextual information associated with biological interactions, including cell type, cell line, disease, organ, and intracellular location. By combining relation and context classification and leveraging both labeled and unlabeled data, CELESTA helps transform literature-derived interactions into richer, context-aware biological knowledge. Read more about CELESTA in our Journal of Biomedical Informatics paper, “An open-set semi-supervised multi-task learning framework for context classification in biomedical texts,” and explore the code, dataset, and implementation documentation on GitHub. Use the CELESTA web interface within the BOHEME ecosystem to extract and assign biological context to interactions from biomedical text.

CLARINET - Knowledge Classification with Context

CLARIfying NETworks

A model recommendation tool that evaluates literature-derived event collaboration graphs to identify useful extensions to existing biological network models. CLARINET organizes interactions extracted from the literature into event collaboration graphs and evaluates them using metrics that capture their occurrence and co-occurrence across publications, as well as their connectivity to an existing baseline model. It groups related interactions and ranks candidate extensions, helping modelers identify connected sets of new biological knowledge that can be incorporated into a model. CLARINET is parametrizable, allowing users to explore different criteria for selecting model extensions, and is designed to efficiently process large sets of literature-extracted interactions. Read more about CLARINET in our Bioinformatics Advances paper, explore the ReadTheDocs documentation, access the GitHub repository, or try CLARINET interactively through our Jupyter notebook.

DiSH - Discrete, Stochastic, Hybrid Simulation

Discrete Stochastic Heterogeneous simulator

A flexible simulator for hybrid biological models with multiple deterministic and stochastic simulation schemes and customizable timing behaviors. DiSH supports hybrid models in which individual elements can have different numbers of discrete activity levels and use either logical or arithmetic update functions, allowing models to combine qualitative regulatory relationships with more detailed quantitative information when available. DiSH supports deterministic and stochastic simulation schemes, with additional flexibility provided by regulator weights, element memory, delayed responses, and spontaneous and unbalancing behaviors. These capabilities enable researchers to explore system dynamics at varying levels of resolution and parametrization without requiring the extensive kinetic parameters typically needed for continuous models. Earlier versions of DiSH were presented at the 2017 and 2021 IEEE Winter Simulation Conferences (WSC'17 paper and WSC'21 paper). We also developed a hardware implementation of DiSH in Verilog and deployed it on an FPGA; read more about this work in our EMBC paper. Stay tuned for DiSH 2.0, coming soon to the BOHEME web interface.

FIDDLE - Graph-based Model Extension

Finding Interactions using Diagram Driven modeL Extension

A model assembly and extension tool that uses dynamic behavior to guide the incorporation of biological knowledge into executable models. FIDDLE evaluates candidate interactions by adding them to an existing baseline model, simulating the resulting models, and determining which additions improve agreement with the expected system behavior. Its Breadth First Addition (BFA) and Depth First Addition (DFA) approaches provide complementary strategies for selecting and incorporating new knowledge, enabling systematic exploration of how network structure, available prior knowledge, and model dynamics affect automated model construction and extension. Read more about FIDDLE in our IEEE Transactions on Computational Biology and Bioinformatics paper, explore the ReadTheDocs documentation, or access the source code on GitHub.

FLUTE - Interaction Confidence and Filtering

The FiLter for Understanding True Events 

A filtering tool that utilizes existing databases to select biochemical interactions with high confidence. FLUTE evaluates biomolecular interactions extracted from biomedical literature by comparing them with evidence from public biological databases and assigning confidence scores, helping reduce inaccurate or irrelevant information before model construction. Read more about FLUTE in our Database paper and explore the detailed ReadTheDocs documentation, which you can open separately using the link or browse directly in the embedded view below. The documentation also includes a Jupyter Notebook tutorial with examples for filtering interactions, analyzing paper sets, querying proteins, identifying recent interactions, and detecting duplicates. Access the complete FLUTE implementation and additional examples through the FLUTE GitHub repository. 

KALIMBA - Interaction Extraction and Curation

Knowledge-Assisted Literature Investigation, Mining, and Biological Annotation

A human-in-the-loop platform for extracting, curating, and exploring biological interactions from scientific literature. KALIMBA integrates NLP- and LLM-based interaction extraction with expert annotation and evidence-grounded conversational querying in a unified workflow. Users can search the literature, extract interactions using NLP-only, LLM-only, or hybrid approaches, review and correct extracted knowledge, and interactively query the supporting source literature through a retrieval-augmented generation (RAG) chat interface. By keeping domain experts actively involved throughout the process, KALIMBA supports efficient and transparent extraction and curation of biological knowledge from the rapidly growing scientific literature. Read more about KALIMBA in our BioNLP paper, and use it through the web interface within the BOHEME ecosystem.

VIOLIN - Knowledge Reconciliation

Versatile Interaction Organizing to Leverage Information in Networks

A modular framework for reconciling new biological knowledge with existing knowledge graphs and mechanistic models. VIOLIN systematically compares literature-derived biological interactions with structured baseline knowledge to determine whether new evidence corroborates, contradicts, or extends what is already known, while accounting for biological context and other interaction attributes. Its configurable graph-based approach supports heterogeneous knowledge sources and alternative reconciliation strategies, enabling transparent and scalable knowledge integration, model verification, refinement, and extension. Read more about VIOLIN in our bioRxiv preprint and use the VIOLIN 2.0 web interface within the BOHEME ecosystem to reconcile your own interaction lists and models.

DySE - Dynamic Systems Explanation Framework (retired)

The amount of data and information produced by experimental laboratories studying complex biological systems is increasing at an incredible rate and is being disseminated at a growing number of publishing venues, scattering the knowledge and limiting the effectiveness and feasibility of manual analysis of all available information. This highlights the need for automated methods to retrieve and connect related pieces of the extensive yet fragmented collective knowledge, for the purpose of understanding, interpreting, explaining, and predicting the behavior of studied systems. To address this challenge, we developed the DySE framework. With DySE, we study the interplay between the immune system and diseases, including T cells, macrophages, glioblastoma multiforme stem cells, pancreatic cancer, melanoma, colon cancer, and ovarian cancer. DySE is featured in DARPA’s World Modelers program, where it assists decision-makers in identifying key points of intervention in complex world systems.

 Funding