SI Pilot Project: Lawrence Berkeley National Laboratory
- Principal Investigator: Paramvir Dehal (Lawrence Berkeley National Laboratory)
- Co-Investigators: Christopher Mungall (Lawrence Berkeley National Laboratory), Ryan Wright (Pacific Northwest National Laboratory), Boris Sadkhin (Argonne National Laboratory)
- Scope/Objectives: Develop BERIL, the BER Intelligence Layer: a super intelligence (SI) co-scientist that helps researchers plan, execute, review, and reuse genome-to-phenotype analyses through the DOE Systems Biology Knowledgebase (KBase).
- Potential Impact and Interface with the American Science Cloud (AmSC) and Transformational AI Models Consortium (ModCon): Connect to ModCon for SI workflows and AmSC through DOE-BRIDGE.
- BER Program Manager: Ramana Madupu
DOE-BRIDGE seeks to provide consistently labeled, harmonized datasets designed for super intelligence (SI)–driven reasoning and inference across multiple heterogeneous data sources. Building on these efforts, this project is developing BERIL, the BER Intelligence Layer, an SI co-scientist that works with researchers to turn data accessed through the DOE Systems Biology Knowledgebase (KBase) Data Lakehouse and BRIDGE’s federated connectivity into hypotheses for experimental testing, reproducible analyses, and new knowledge about how genomes give rise to phenotypes. BERIL is an independent BER SI pilot and KBase’s co-scientist partner; it is built on the KBase Data Lakehouse and delivered to BER researchers through the KBase platform, so the analyses BERIL performs support attributed reuse within applicable access permissions rather than remaining private results.
Co-Scientist
BERIL operates as a research partner or co-scientist. Given a scientific question, BERIL proposes a plan, executes it against the data lakehouse and BER analysis tools, synthesizes the findings, subjects them to review, and stores the results so they can be reused. Researchers set the question and approve the plan; the co-scientist does the work and shows how it was done. This loop is integrated with KBase’s identity and data access controls, and results return to KBase so that knowledge accumulates where BER researchers already work.
Each analysis is recorded in BERIL’s research observatory, and the records are stored in KBase: the plan, analysis steps and tool invocations, data consulted, evidence supporting each conclusion, and available cost information. This record supports inspection and reproducibility of SI-generated results and gives the project a growing dataset of how scientific reasoning succeeds or fails, which the team uses to guide and evaluate improvements.
Specialized Agents
BERIL delegates to specialized agents and integrates their results. The first, developed with the KBase SI/machine learning team, is a gene-function annotation agent that weighs multiple complementary lines of evidence (sequence homology, structure, genomic context, fitness and phenotype data, and the literature) to propose and justify functions for uncharacterized genes. The agent is being evaluated against expert curation and gold-standard datasets. The team is developing agents for structure-based reasoning [with Argonne National Laboratory through the American Science Cloud (AmSC) partnership] and exploring experiment-design support so predictions can be turned into ranked, resource-aware candidates for experimental validation. Agents expose their capabilities through open protocols, allowing tools developed by BER User Facilities and programs to be integrated as they become available.
Evaluation and Trust
Because an SI co-scientist is only useful if its output can be trusted, evaluation is built in. The team is developing expert-reviewed benchmarks of gene-function and genome-to-phenotype assertions, scoring agent outputs against them and expert review, and evaluating how the confidence BERIL reports corresponds to observed accuracy, the basis for calibrated trust: an explicit statement of how much confidence a result deserves, given the evidence and the agent’s track record on similar tasks. A shared evaluation framework, which is under development, will allow co-scientist systems to be compared on the same tasks with independent assessment.
This work is grounded in BERIL’s users. In May 2026, KBase and BERIL ran a one-day “co-scientist paper challenge” in which BER scientists used the co-scientist on their own questions against KBase Lakehouse data; the resulting interview study, Toward Calibrated Trust: What Working Scientists Need from an AI Co-Scientist (Cohoon et al. 2026; 10.5281/zenodo.20595967), identified staying in command of the collaboration and calibrating trust as central needs, which now guide BERIL’s design.
Progress and Impact
BERIL is in use in ongoing collaborations that apply the co-scientist to genome annotation, comparative genomics, and genome-to-phenotype questions in BER-relevant microbes and plants. The engine is open source (github.com/beril-doe) and domain-independent by design. BER biology is its first domain, and as an AmSC infrastructure partner, BERIL is testing agentic workflows on AmSC compute alongside Genesis Mission projects. Pilot demonstrations target genes, pathways, and microbial strains that support BER’s biomanufacturing and bioproduct objectives.
By pairing a co-scientist with the KBase Data Lakehouse, the resources federated through BRIDGE, and KBase’s other agents and tools, BERIL aims to shorten the path from data to testable hypothesis, make SI-driven analysis inspectable and attributable, and give BER researchers a partner whose improvements are guided by reviewed results.