MOAt-ClinicalTrials-MoA-256

Find every trial targeting the same biological pathway — across every disease it's been tested in.

MOAt (Mechanism-Of-Action Transformer) is the mechanism-similarity model in the OntologerMed suite. It embeds any clinical trial into a 256-dimensional vector shaped by biological mechanism of action, trained on 563,845 ClinicalTrials.gov entries grouped by MeSH intervention terms.

A keyword search for "semaglutide" finds semaglutide trials. MOAt finds liraglutide, tirzepatide, exenatide, and dulaglutide trials too — because it understands they all act on the GLP-1 pathway. And it finds them across obesity, diabetes, NASH, cardiovascular disease, and Alzheimer's, not just the indication you started with.

Also known as OntologerMed TargetLens in the Commercializer.ai platform.

---

Model Overview

Property	Value
Model name	MOAt-ClinicalTrials-MoA-256
Product name	OntologerMed TargetLens
Base model	NeuML/pubmedbert-base-embeddings (PubMedBERT)
Output dimension	256
Training method	Contrastive learning — Triplet Loss
Training corpus	563,845 ClinicalTrials.gov entries
Grouping signal	10,609 standardised MeSH intervention terms
License	Apache 2.0
HuggingFace	Ontologer/MOAt-ClinicalTrials-MoA-256

---

How It Works

MOAt is trained to cluster trials by what the drug does biologically, not what disease it treats or what words appear in the text.

• Same mechanism, different diseases → close vectors. A semaglutide obesity trial and a liraglutide diabetes trial use different language and treat different conditions — but both activate the GLP-1 receptor. MOAt places them close together.
• Same disease, different mechanisms → far apart. A semaglutide obesity trial and a phentermine obesity trial both target obesity, but through completely different pathways. MOAt separates them.
• Class nuance is preserved: MOAt correctly distinguishes pure GLP-1 agonists (semaglutide, liraglutide, similarity 0.81) from dual GIP/GLP-1 agonists (tirzepatide, similarity 0.35–0.44) — a pharmacological distinction keyword search cannot make.

---

Architecture

Trial text → PubMedBERT (384-dim) → Dense projection (384 → 256) → L2 normalize → vector

• Base: PubMedBERT — pre-trained on 30M+ PubMed abstracts
• Projection: Single dense layer reducing 384 → 256 dimensions
• Normalisation: L2, enabling cosine similarity via dot product
• Training signal: Triplet loss — anchor and positive share a MeSH intervention term; negative has a different intervention term
• Triplets: 10,000 generated from 563K trials across 10,609 distinct MeSH intervention categories

---

Validation

Similarity matrix on five representative trials:

	GLP-1 obesity (sema)	GLP-1 obesity (lira)	PD-1 lung cancer	GIP/GLP-1 obesity (tirz)	KRAS solid tumours
GLP-1 obesity (sema)	1.00	0.81	—	0.44	—
GLP-1 obesity (lira)	0.81	1.00	—	0.35	—
PD-1 lung cancer	—	—	1.00	—	0.69
GIP/GLP-1 obesity (tirz)	0.44	0.35	—	1.00	—
KRAS solid tumours	—	—	0.69	—	1.00

Semaglutide and liraglutide (both pure GLP-1 agonists) cluster at 0.81. Tirzepatide sits further away (0.35–0.44) because its dual GIP/GLP-1 mechanism is genuinely distinct. PD-1 checkpoint inhibitors cluster with KRAS-targeted therapies (both immuno-oncology, 0.69). Cross-class similarity is low. The model learned mechanism, not text.

---

Training Details

Parameter	Value
Epochs	10
Batch size	32
Learning rate	2e-5
Loss margin	0.5
Hardware	NVIDIA DGX Spark (Blackwell GB10)
Triplet grouping	MeSH intervention term (clintrials_intervention_mesh_terms.term)
Distinct MeSH terms	10,609 intervention terms (each with ≥2 trials)

---

Business Use & Applications

MOAt is purpose-built for mechanism-level intelligence — finding the full competitive and cross-indication landscape of a drug class or biological target.

Pharmaceutical & Biotech R&D

• Competitive intelligence by mechanism — map every trial in your drug class, not just your indication

  • Find all GLP-1 agonist trials across obesity, diabetes, NASH, cardiovascular disease, and Alzheimer's — not just the ones that share your drug name
  • Identify new indications being explored for your mechanism before they reach the public domain
  • Monitor which competitors are active in your mechanism space across all therapeutic areas simultaneously

• Target landscape mapping — understand the full clinical history of a biological target

  • Retrieve all trials ever run against a receptor, enzyme, or pathway — regardless of the drug used
  • Identify which indications have been explored for a mechanism and which remain untested
  • Find failed trials in your mechanism class to understand what populations and endpoints to avoid

• Cross-indication drug repurposing — surface new applications for a validated mechanism

  • Embed a lead compound's mechanism and find populations where similar mechanisms have been tested
  • Identify approved drugs in your mechanism class being tested in early-phase trials in new indications
  • Find mechanism overlap between your pipeline and external assets to identify licensing or co-development opportunities

Business Development & Licensing

• Licensing opportunity identification — find assets with mechanisms complementary to your pipeline

  • Identify early-phase assets in your mechanism class that could be in-licensed before competitors see them
  • Assess whether an acquisition target's mechanism has been validated in indications adjacent to your focus
  • Benchmark a target company's mechanism against the historical trial landscape for that pathway

• Partnership scouting — identify academic and biotech groups actively working in a mechanism area

  • Find all active trials in a mechanism class to map the competitive landscape for a partnership search
  • Identify sponsors with multiple mechanism-similar trials — indicating deep target expertise

Investment & Due Diligence

• Class-level diligence — understand how a mechanism class has historically performed
  • Retrieve all historical trials in a drug class and assess the distribution of phase, outcomes, and safety profiles
  • Identify whether a biotech's lead compound mechanism is validated, crowded, or early-stage
  • Assess mechanism differentiation: does the company's asset sit in a distinct mechanistic neighbourhood, or in a commodity class?

Drug Discovery & Academic Research

• Mechanism de-risking — find the clinical precedent for a novel target before expensive development
  • Retrieve all trials that have targeted a pathway, including those with negative or inconclusive outcomes
  • Identify which patient populations have been tested for a mechanism and which have been excluded
  • Surface the endpoint and biomarker precedents that have been used to measure mechanism engagement

---

Example Queries

MOAt is used via embedding + nearest-neighbor lookup. Below are illustrative examples of the output.

---

Example 1: Mechanism Landscape Query

Query: "What other trials are targeting the same mechanism as our GLP-1 agonist programme?"

Input text:

Phase 3 RCT of semaglutide 2.4mg SC weekly vs placebo in adults with BMI ≥ 30. Primary endpoint: % body weight loss at 68 weeks.

Top 10 MOAt nearest neighbors:

Trial	Mechanism	Indication	Cosine Sim
Liraglutide 3.0mg obesity RCT	GLP-1 agonist	Obesity	0.92
Liraglutide 1.8mg LEADER trial	GLP-1 agonist	T2D + CV risk	0.88
Semaglutide SUSTAIN-6	GLP-1 agonist	T2D + CV risk	0.87
Exenatide weekly DURATION-1	GLP-1 agonist	T2D	0.84
Dulaglutide AWARD-3	GLP-1 agonist	T2D	0.81
Tirzepatide SURMOUNT-1	GIP/GLP-1 dual	Obesity	0.44

Insight: The model correctly separates pure GLP-1 agonists (0.81–0.92) from tirzepatide's dual mechanism (0.44). Automatically surfaces the cross-indication trials (T2D, CV risk) that keyword search on "obesity" would miss.

---

Example 2: Cross-Indication Discovery

Query: "Where else has PD-1 checkpoint inhibition been tested besides lung cancer?"

Input: PD-1 inhibitor lung cancer trial text

MOAt neighbors (sample):

• PD-1 inhibitor in melanoma (sim 0.91)
• PD-1 inhibitor in head and neck squamous cell carcinoma (sim 0.88)
• PD-1 inhibitor in triple-negative breast cancer (sim 0.85)
• PD-L1 inhibitor in urothelial carcinoma (sim 0.82)
• CTLA-4 inhibitor + PD-1 combo in renal cell carcinoma (sim 0.74)

Insight: Surfaces the full cross-indication landscape of checkpoint inhibition — including PD-L1 and CTLA-4 combination approaches — from a single query.

---

Usage

from sentence_transformers import SentenceTransformer
import numpy as np

# Load model
model = SentenceTransformer("Ontologer/MOAt-ClinicalTrials-MoA-256")

# Embed a trial
trial_text = """
Phase 3 RCT of semaglutide 2.4mg SC weekly vs placebo in 1,961 adults
with obesity (BMI ≥ 30). Primary endpoint: % change in body weight at 68 weeks.
"""
query_vector = model.encode([trial_text])  # shape: (1, 256)

# Find mechanism-similar trials
similarities = np.dot(query_vector, trial_index_vectors.T)
top_k = np.argsort(similarities[0])[::-1][:20]

# Inspect nearest neighbors
for i in top_k:
    print(f"sim={similarities[0][i]:.3f} | {trial_metadata[i]['title']}")

Index size: 256 dims × 4 bytes × 563K trials ≈ 550MB.

---

Part of the OntologerMed Suite

Model	Role
OntologerMed-ClinicalTrials-Instruct	Generative LM — reasoning, extraction, and summarisation over trial text
FATE-ClinicalTrials-Outcome-256 (TrialPulse)	Outcome-shaped embedding — similarity by historical success/failure pattern
MOAt-ClinicalTrials-MoA-256 (TargetLens)	Mechanism-of-action embedding — similarity by biological pathway
PACT-ClinicalTrials-Pop-256 (PathFinder)	Population embedding — similarity by patient demographics and disease
ORACLE-ClinicalTrials-SuccessProb-v1	Classifier — probability estimate combining all three embedding dimensions

ORACLE uses MOAt embeddings as one of three inputs — alongside FATE and PACT — to generate a combined probability-of-success score.

---

Limitations

• Mechanism, not efficacy: Nearby vectors indicate mechanism similarity, not that the treatments are equally effective or safe. Always validate top neighbors with metadata filters.
• MeSH grouping limitations: Trials with non-standard, experimental, or combination mechanisms may not map cleanly to MeSH terms used during training.
• English only: Trained on English-language trial records.
• Text length limits: PubMedBERT has a 512-token limit. Very long trial protocols are truncated. Ensure the most mechanism-relevant text (interventions, summary) is in the first 512 tokens.
• Not medical or regulatory advice: Mechanism similarity does not imply safety, efficacy, or regulatory substitutability.

---

Citation

@misc{moat-clinicaltrials-2026,
  title        = {MOAt-ClinicalTrials-MoA-256: Mechanism-of-Action Embeddings for Clinical Trial Intelligence},
  author       = {Mishra, Sid},
  year         = {2026},
  note         = {Contrastive triplet embedding model trained on 563,845 ClinicalTrials.gov entries, grouped by MeSH intervention terms.},
  howpublished = {\url{https://huggingface.co/Ontologer/MOAt-ClinicalTrials-MoA-256}}
}

---

About the Author

Sid Mishra — Founder, Ontologer · Convixion AI

Sid is the founder of several AI-native and AI-powered startups and initiatives, based in Singapore. He founded Ontologer as the dedicated AI research arm of Convixion AI, with a focus on building domain-specific language models from the ground up — including data pipelines, training infrastructure, evaluation frameworks, and production deployment.

Ontologer generates novel LLM and embedding models purpose-built for use within Convixion AI's Commercializer.ai platform. MOAt is part of the OntologerMed suite — a family of purpose-built models for clinical trial intelligence. Ontologer performs every step of model development — dataset curation, training infrastructure, evaluation, and production deployment — in-house.

Collaboration & Custom Work

Sid is open to collaborating on:

• Custom domain-adapted embedding models — contrastive/triplet training on proprietary datasets for specialised retrieval tasks
• End-to-end LLM and embedding pipelines — from data curation to training to production deployment
• Evaluation framework design — task-specific benchmarks and retrieval evaluation pipelines
• RAG + embedding system design — pairing domain-adapted models with retrieval systems for production use
• Custom model architecture consulting — base model selection, training strategy, hardware planning

Site	ontologer.com
Email	sid@ontologer.com · sid@convixion.ai
LinkedIn	linkedin.com/in/sid-m-427b9865