Agentic AI Systems for Drug Discovery and Business Intelligence

loka · Tech · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Drug discovery and business intelligence

Loka and Domo presented agentic AI systems that move beyond chatbots into assistants for drug discovery, call centers, and financial analysis.

Constraints:

Architecture

The Loka/Domo discussion shows two agentic production patterns. Loka's ADA drug-discovery assistant uses a backend agent to query pathway databases, STRING DB, GEO, UniProt, protein-folding models such as ESMFold, docking models such as DiffDock, and visualizers from natural-language prompts. Domo's platform exposes data sources, agent tasks, tools, human-in-the-loop steps, and Bedrock-backed safeguards for BI and call-center workflows. Both emphasize model agnosticism, routing, security boundaries, and production readiness over one chatbot endpoint.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  drug-discovery-prompt[Drug discovery prompt] --> ada-backend-agent[ADA backend agent]
  ada-backend-agent --> scientific-databases[Scientific databases]
  ada-backend-agent --> predictive-model-tools[Predictive model tools]
  ada-backend-agent --> visualization-outputs[Visualization outputs]
  domo-agent-platform[Domo agent platform] --> business-data-sources[Business data sources]
  domo-agent-platform --> human-review[Human review]
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
Drug discovery prompt Natural-language research request for pathways, proteins, genes, structures, or molecules. ADA backend agent
ADA backend agent Agent that selects databases, predictive models, and output generation steps. Drug discovery prompt, Scientific databases, Predictive model tools, Visualization outputs
Scientific databases KEGG, STRING DB, GEO, UniProt, and other internal/external sources. ADA backend agent
Predictive model tools Protein folding and docking models such as ESMFold and DiffDock. ADA backend agent
Visualization outputs CSV downloads and 3D visualizers for structures and docking outputs. ADA backend agent
Domo agent platform BI platform for data-source-connected agentic workflows. Business data sources, Human review
Business data sources Customer databases, data lakes, and SaaS data exposed to Domo agents. Domo agent platform
Human review Human-in-the-loop path for sensitive business decisions and customer-facing workflows. Domo agent platform

Data flow

In ADA, a researcher asks a drug-discovery question; the backend agent queries pathway, protein-interaction, expression, sequence, folding, docking, and visualization resources; outputs are summarized, downloaded, or rendered. In Domo, business data sources and agent tasks feed agentic workflows, with human review available before sensitive actions or customer-facing changes.

3 supporting quotes
“At its core, ADA is powered by a back-end agent capable of retrieving and interacting with internal and external databases”article, stated
“At its core, ADA is powered by a back-end agent capable of retrieving and interacting with internal and external databases, executing predictive models and generating research outputs, all triggered by natural language prompts.”article, stated
“Um luckily you know the the nature of how Domo lives on top of bedrock it comes with a lot of safeguards security and compliance concerns that that customers won't get outside of that tight collaboration.”transcript, stated

Why this architecture

The presenters argue that production agent systems need routing, model agnosticism, security, governance, and data readiness. ADA decomposes drug discovery into database/model calls because researchers otherwise switch between fragmented tools, while Domo emphasizes agents over governed business data with humans in the loop for higher-risk actions.

3 supporting quotes
“research outputs, all triggered by natural language prompts. No more scripting, no more switching between tools. Just ask and Atta handles the rest. Built for life sciences teams, ADA turns fragmented manual research into seamless scalable progress, saving hundreds of hours on high effort pipelines.”transcript, stated
“um or even one model where you're um you have like one use case and you don't want every single response to go through your most expensive you know model”transcript, stated
“hopefully having a human need to evaluate it every time although there's a lot of value in human in the loop systems. Um, so each of those pieces are something we try to think about some upfront, right, before you get down a path where you've done something you're excited”transcript, implied

Models used: ESMFold · DiffDock · LLMs · Bedrock-hosted models

Scale

Alternatives considered (roads not taken)

Single expensive model for every request

Rejected because: Production systems need model routing for nonsensical, simple, and complex requests.

“um or even one model where you're um you have like one use case and you don't want every single response to go through your most expensive you know model right you want it to be do routing where in a production environment”transcript, stated

Assume generative models can use any data as-is

Rejected because: The panel warns that unstructured tolerance does not remove data-readiness requirements.

“course interpreting you know unstructured data but that doesn't mean that you can just sort of throw total garbage at it and hope for good outcomes. Um and so I think one of the data readiness is something that we look at early on. Um some of this is um around of course like is your data accessible?”transcript, stated

Patterns

multi-agent-orchestration (cross-industry)

The presenters describe moving beyond chatbot/RAG into multi-agent systems over complex data.

“these days a lot about agents multi-agentic systems um really thinking through um more complicated data sources”transcript, stated

human-in-the-loop-gate (oversight)

The systems maintain human oversight while allowing agents to take actions.

“Both companies emphasize the importance of systematic approaches to AI implementation, moving beyond simple chatbots to multi-agent systems that can take autonomous actions while maintaining human oversight through human-in-the-loop architectures.”summary, stated

Tradeoffs

Design for model agnosticism and routing.

Chose Shared infrastructure that can route nonsensical, simple, and complex requests to different model paths over Sending every request to the most expensive model.

Rationale: Production agentic systems needed cost control, scale, and the ability to combine specialized models for different subtasks.

“you don't want every single response to go through your most expensive you know model right you want it to be do routing where in a production environment someone asks something totally nonsensical”transcript, stated

How to control production inference cost.

Chose model routing and model agnosticism over one frontier model for every path.

Rationale: Different requests warrant different model costs and capabilities in production.

“model agnosticism”transcript, stated

Gotchas (what bit them)

Generative AI still depends on mature, secure, and accessible data.

Root cause: Unstructured-data tolerance does not make poor, inaccessible, or insecure data usable in production.

Resolution: Loka evaluates data readiness, security boundaries, governance, and baseline performance early in the discovery period.

“generative AI models have some advantages over traditional models in that they um can do a better job of course interpreting you know unstructured data but that doesn't mean that you can just sort of throw total garbage at it”transcript, stated

Evaluation

The transcript gives production examples but does not provide a consistent metric framework.

Outcomes

The summary characterizes them as production implementations across industries.

“customers um to date. Uh the majority of those have actually moved to production. So have stories of these things working”transcript, stated