Domain-Specific LLMs for Drug Discovery Biomarker Identification

benchsci · Healthcare · 2023 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (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: Preclinical drug discovery and biomarker identification

BenchSci needed to help scientists reason over complex disease biology, evidence, nomenclature, and experiment design so they could identify biomarkers and hypotheses faster without losing scientific veracity.

Constraints:

• Preclinical drug discovery is slow, expensive, and has a high failure rate.
• Disease biology includes trillions of relationships that scientists cannot manually connect at scale.
• Biomedical evidence comes from many publishers, open sources, paywalled sources, and pharma-internal data.
• Scientific naming is ambiguous, with many names for the same entity and many entity relationships.
• Enterprise pharma deployments require scientific accuracy, explainability, and readiness for scrutiny.

Architecture

BenchSci built ASCEND on a decade of structured biomedical evidence. Publisher, open, paywalled, and pharma-internal data are converted into an evidence map and ontology knowledge base. Domain-specific vision and text LLMs extract and structure evidence at scale, while a RAG layer combines that structured data with generative summarization and conversation. The scientist-facing assistant answers workflow questions around hypotheses, experiment design, translation to humans, and biomarker identification with evidence and reasoning.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  biomedical-data-foundation[Biomedical data foundation] --> domain-specific-llm-extraction[Domain-specific LLM extraction]
  domain-specific-llm-extraction --> evidence-map[Evidence map]
  domain-specific-llm-extraction --> ontology-knowledge-base[Ontology knowledge base]
  evidence-map --> ascend-assistant-interface[ASCEND assistant interface]
  ontology-knowledge-base --> ascend-assistant-interface
  ascend-assistant-interface --> scientist-workflows[Scientist workflows]

Components

Data flow

Research data is collected from publisher, open-access, paywalled, and internal pharma sources. Domain-specific models extract scientific evidence, connect it into an evidence map and ontology knowledge base, and applications expose that foundation to scientists. The assistant retrieves structured evidence and uses generative AI for summarization and conversation while showing reasoning back to skeptical scientific users.

Why this architecture

BenchSci chose structured evidence plus RAG because scientists need explainable, scientifically accurate answers rather than generic generation. Domain-specific models were necessary to interpret biomedical evidence, resolve nomenclature, and outperform generic foundation models for biomarker identification.

Models used: domain-specific LLMs · vision models · text models · smaller derivative-insight models · Google Med-LM

Scale

• BenchSci served over 50,000 scientists.
  “we work with the majority of big Pharma today and we have over 50,000 scientists that are using our platform”transcript, stated
• BenchSci worked with over 75% of the top 20 pharmaceutical companies.
  “they already working with over 75% of the top 20 pharmaceutical companies”transcript, stated
• Drug-discovery projects were described as taking 8-14 years and costing over $2 billion.
  “it takes absolutely forever to develop new drugs we're talking between 8 and 14 years to bring a drug to Market and it costs absolutely a fortune which is over $2 billion”transcript, stated
• Disease biology was framed as involving four trillion relationships.
  “there are four trillion relationships in our body and we don't know yet how all of them work and how they connect with one another”transcript, stated
• The biomarker use case reportedly increased scientist productivity by over 40%.
  “this actually increased their productivity by over 40% like I I don't know about a lot of software that actually increase productivity by 40%”transcript, stated

Alternatives considered (roads not taken)

Use generic foundation models directly for biomedical reasoning.

Rejected because: BenchSci expected medical/domain-specific pretraining to outperform generic foundation models on its scientific tasks.

Build an AI-first biotech company around one drug pipeline.

Rejected because: BenchSci argues that approach can be faster for individual drugs but does not scale as broadly as assisting many scientists across pharma.

Answer scientists without exposing evidence and reasoning.

Rejected because: Scientific users are skeptical, and adoption depends on showing evidence and reasoning behind generated answers.

Patterns

rag-retrieval (retrieval)

ASCEND combines cleaned structured biomedical evidence with generative summarization and conversation in a RAG architecture.

fine-tuning-adaptation (model-operations)

BenchSci built domain-specific vision and text models to interpret biomedical evidence the way scientists do.

structured-output-contract (data)

Scientific evidence is extracted, connected, and structured into the evidence map and ontology knowledge base before generation.

Tradeoffs

Use a structured evidence foundation plus RAG.

Chose Evidence map, ontology, and RAG-backed assistant experience. over A purely generative assistant over unstructured biomedical text..

Rationale: Scientific users require explainability, limited hallucination, and evidence-backed reasoning.

Invest in domain-specific medical models.

Chose Medical/domain-specific model capability. over Generic foundation models alone..

Rationale: BenchSci believed domain expertise was required for problems generic foundation models could not solve.

Gotchas (what bit them)

Scientific entity names and relationships are too ambiguous for naive text retrieval.

Root cause: Biomedical entities can have 20-30 names and many relationships across genes, diseases, pathways, and processes.

Resolution: BenchSci built an ontology knowledge base alongside the evidence map.

Scaling engineering alone was insufficient for biomedical assistants.

Root cause: Scientific veracity required machine-learning outputs to be biologically and scientifically correct.

Resolution: BenchSci embedded scientists into the development process and paired engineers with scientists.

Generic generation would not satisfy skeptical scientific users.

Root cause: Scientists needed to see evidence and scientific reasoning behind machine-generated answers.

Resolution: BenchSci fused RAG with explainability and limited hallucination over structured evidence.

Evaluation

BenchSci evaluated the biomarker assistant through scientist-reported productivity and the ability to reduce a month-scale biomarker workflow to days.

Metrics: Over 40% productivity increase reported by scientists at a major pharma company · Process time reduced from months to a few days · 50,000 scientists using the platform · Over 75% of top 20 pharmaceutical companies as customers

Outcomes

• Created an evidence-based assistant experience for scientists.
• Applied domain-specific medical/foundation models to biomarker identification.
• Reported over 40% productivity increase and months-to-days workflow compression.
• Built reusable evidence-map and ontology assets that drive multiple scientist workflows.

BenchSci describes 50,000 scientists, over 75% of top 20 pharma companies, over 100 scientists on staff, hundreds of data sources, and nearly a decade of evidence-foundation work.