Databricks 为 Genie Code 推出医疗与生命科学技能包 Vital Skills

Code with AI on a unified, governed platform

Genie Code, our coding agent optimized for Databricks, allows you to chat instead of code to perform your data and AI tasks. Critically, Genie Code’s integration with Unity Catalog ensures that the agent understands your enterprise’s data and follows the same security and governance rules as the rest of the Databricks platform. Thus, Genie Code significantly outperforms a leading coding agent by more than twice on real-world data tasks such as exploratory data analysis, machine/deep learning, creating dashboards, agents, apps, declarative automation bundles, Lakeflow jobs and scheduling them.

However, Genie Code and other general AI coding assistants lack domain specificity and assume generic methods and clinical thresholds that do not meet healthcare and life sciences (HLS) standards or best practices.

Databricks Vital Skills bridges this gap by providing packaged playbooks built in the open Agent Skills standard. They equip AI coding agents with industry knowledge so agents know which questions to ask, which steps are non-negotiable, and which tested tools to use.

Vital Skills is designed for Genie Code to make the best use of the unified Databricks ecosystem so generated code is not disconnected from your governed data, and you're not left to stand up serving, GPUs, governance, and scheduling yourself. You get a system, not just a sandbox. With the Databricks ecosystem, you:

• Are grounded in all your data assets. The quality of the answer rises with the reach of the data foundation underneath it.
• Are governed end to end. Data, skills, and models with conversational access are all governed, tracked and secured.
• Get a comprehensive platform. Model Serving, GPUs, and Apps are part of the platform, with no separate software or service to buy and wire together.
• Have open choice. Use any cloud, any model, any framework.
• Design for production, not just play. The same work in prototype can be promoted into scheduled Lakeflow pipelines and jobs with real throughput.

We tested Genie Code with and without Vital Skills on BiomniBench, a public benchmark of biomedical data-analysis tasks graded against expert-written rubrics. Genie Code alone rarely failed on the obvious parts.

It parsed documents and generated charts with ease. However, it sometimes fell short on domain-specific methodology and knowledge and was prone to overassumptions. With Vital Skills, Genie Code scored higher on the relevant tasks and had no negative impact on tasks that were out of scope of Vital Skills.

Skills built by healthcare and life sciences experts, for healthcare and life sciences

Each skill started with a field engineer who had solved the problem several times with customers. We held on to three principles:

1. Ask before assuming. When a decision is the analyst’s to make, e.g. a clinical threshold, the skill pauses, provides options, and lets the user choose.
2. Trust-tested tools, not improvised code. For sensitive work, Genie Code runs tested tools and code rather than writing new logic each time.
3. Prove it. Every skill was run side by side against Genie Code without it, on the same tasks.

A catalog that spans the healthcare and life sciences value chain

Vital Skills puts the know-how of a biostatistician, privacy officer, bioinformatician, or actuary directly in the hands of every analyst, helping teams move faster before engaging additional expertise. The Databricks Healthcare and Life Sciences field team has built skills across the industry, covering key steps in data processing, analysis, model life cycle management, evaluation, report and slide generation.

These skills can be used directly, or further customized to your needs and to leverage your deep expertise in your discipline. We expect this starter set of skills to grow over time and welcome contributions to the open source repository.

Seamlessly get from data to results to a finished product, all on Databricks

Go from data to a finished report or agent or dashboard seamlessly on Databricks.

For example, our semantic-builder skill defines and computes commonly used metrics (e.g. length of stay) into a Metric View from claims or electronic health records. Once Metric Views are defined in Unity Catalog, you can easily re-use them across multiple tools like dashboards, Genie Agents, external BI or spreadsheets.

In the following sections, we will dive into a couple of key skills.

Cohort builder: start with plain English descriptions

An analyst describes a population in plain language — "patients with type 2 diabetes and uncontrolled blood sugar" — and the skill turns that into a cohort they can reproduce, explain, and defend. Cohort building is an integral starting point for nearly every analytics effort: RWE studies, clinical-trial feasibility, payer care-management and population-health outcomes research.

Highlights:

• It works with structured and/or unstructured data. It finds patients in coded records (diagnoses, labs, medications) and optionally, in free-text notes and scanned PDFs.
• It stops to ask at key decision points instead of guessing: which lab value counts as "uncontrolled," how to combine structured and unstructured data, and whether to deliver a table, a notebook, or an answer right in the conversation.
• It checks feasibility first, showing how many patients qualify before building anything.
• It explains its logic. Each patient comes with the reason they were included, without exposing raw clinical notes in the shareable result.
• It consistently gives the same answer. In our tests it produced the same cohort on every run; Genie Code without the skill picked a different definition each time. In paired evaluations, free-text cohort F1 rose from 0.62 to 1.00.

After building cohorts, consider comparative effectiveness studies and share your results using the presentation-builder.

Open weight models: get your favorite healthcare model on Databricks

Some of the most promising AI applications for life sciences are open source biological foundation models that encode single-cell transcriptomics (Geneformer, scGPT, Scimilarity, TEDDY) or protein structure (AlphaFold/OpenFold, Boltz2, etc). They are also difficult to get running, requiring specialized infrastructure and abilities. Because this skill summons the appropriate code, libraries and GPU serving, a scientist can go from “I need this model" to serving a governed model that one can actually use.

Along the way it:

• Asks rather than guess (e.g. how should the model run: a real-time endpoint, batch job, app, or pipeline) so it can deploy the model accordingly with reasoning.
• Makes the model reproducible and auditable by pinning the model and its artifacts (e.g. code, weights, reference databases, and licenses)
• Serves a governed model that is registered in Unity Catalog, deployed, and tested with realistic scientific inputs before handoff.
• Monitors model traffic in inference and system tables for usage/cost tracking and optionally applies sensitive data guardrails.

On a TEDDY single-cell model deployment task, Genie Code without the skill wrote correct code but missed key deployment-readiness steps — pinned dependencies, a post-registration smoke test, and more — completing 19 of 27 checks. With the skill, all 27 completed.

Evaluation: Vital Skills improves Genie Code’s healthcare and life sciences capabilities

Per-skill evaluation

We included a skill to evaluate skills because there is no single universal “accuracy” for healthcare and life sciences workflow. For each skill, it builds a customizable evaluation harness, complete with appropriate tasks, data, scorers, and expected results and behavior, tailored to what each skill is designed to achieve. We ran and scored the evaluation tasks with and without Vital Skills.

BiomniBench evaluation

We also tested Genie Code with and without Vital Skills three times on BiomniBench which curated tasks from top biomedical peer-reviewed papers.

• Better where the skills apply. On the 5 tasks within the skills' scope, Genie Code's average score across 3 trials rose from 71.9 to 76.9 out of 100.
• No cost where they don't. On 18 tasks outside the skills' scope, scores held steady — 69.1 without the skills, 71.2 with them — so the skills stay out of the way when they aren't needed.
• The right skill, every time. With a one-line pointer to the skills in Genie Code's instructions, it opened the right one in every in-scope run.

The right skill makes the right behaviors reliable. On one cohort task, the baseline silently dropped half the eligible patients; with the skill, every one was recovered.

Try Genie Code with Vital Skills

Previous research has shown that Genie Code is more cost-efficient than other coding agents because it uniquely understands your data on Databricks, skipping the typical random walk search.

Extend Genie Code with additional skills like Vital Skills for healthcare and life sciences tasks. Together, one can chain a myriad of skills into complex tasks. For example, after deploying open-weight models, create and schedule a batch inferencing job. If it is an embedding model, prompt Genie Code to create a vector search index.

Browse and install Vital Skills in Genie Code, Genie One or Genie Code CLI. Publish these skills to Unity Gateway to govern and share them with your coworkers. Try the ones that fit your work or customize them to suit your needs. Genie Code even helps you create new skills. Vital Skills is open source and welcomes your contributions.

• https://databricks-industry-solutions.github.io/vital-skills/ , a short "install in Genie Code" text block
• Attend a workshop to use Vital Skills and other BioPharma R&D solutions like Genesis Workbench from Databricks
• "build your own skill" link to CONTRIBUTING.