experiment_results.md

NegBioDB Experiment Results

Last updated: 2026-03-24

Cross-Domain Summary

Domain	Negatives	ML Status	LLM L1 (acc)	LLM L4 (MCC)
DTI (Drug-Target Interaction)	30,459,583	24/24 complete	Gemini 1.000 (3-shot)	≤ 0.18 (near random)
CT (Clinical Trial Failure)	132,925	108/108 complete	Gemini 0.68	Gemini 0.56
PPI (Protein-Protein Interaction)	2,229,670	54/54 complete	~1.000 (3-shot artifact)	Llama 0.441
GE (Gene Essentiality / DepMap)	28,759,256	Seed 42 complete	4/5 models complete	Pending (Llama)

Key insight: LLM discrimination ability (L4 MCC) increases with task complexity: DTI (0.05–0.18) < PPI (0.33–0.44) < CT (~0.48–0.56). All domains show evidence of training data contamination in PPI/CT.

---

DTI Domain (Drug-Target Interaction)

Status as of 2026-03-24: ML 24/24 complete, LLM 81/81 complete.

Database

• 30,459,583 negative results, 5 splits (random/cold_compound/cold_target/scaffold/temporal)
• 1,725,446 export rows (862,723 pos + 862,723 neg)
• Source tiers: gold=818,611 / silver=198 / bronze=28,845,632 (PubChem dynamic tiers)

ML Results Summary

• 3 models: DeepDTA, GraphDTA, DrugBAN × 5 splits × 2 negative types (NegBioDB + uniform_random)
• NegBioDB negatives: Random split near-perfect AUROC (trivially separable)
• Control negatives (uniform_random): Harder splits show meaningful degradation

LLM Results (81/81 complete)

Task	Best Model	Metric	Value
L1 (MCQ)	Gemini	Accuracy (3-shot)	1.000
L4 (Discrim)	Best model	MCC	≤ 0.18 (all near random)

Key finding: DTI LLMs show near-random discrimination (L4 MCC ≤ 0.18). The binding/non-binding decision is too nuanced for zero-shot or few-shot LLMs. This is the hardest domain.

---

CT Domain (Clinical Trial Failure)

Status as of 2026-03-20: ML 108/108 complete, LLM 80/80 complete, L3 judge complete.

Database

• 132,925 failure results from 216,987 trials
• Tiers: gold 23,570 / silver 28,505 / bronze 60,223 / copper 20,627
• 8 failure categories: safety > efficacy > enrollment > strategic > regulatory > design > other (PK=0)

ML Results (CT-M1: Binary Failure Prediction)

Aggregated over 3 seeds (source: results/ct_table_m1_aggregated.csv):

Model	Split	Negatives	AUROC	MCC
XGBoost	random	NegBioDB	1.000	1.000
GNN	random	NegBioDB	1.000	1.000
MLP	random	NegBioDB	1.000	0.992
XGBoost	random	degree_matched	0.844	0.553
MLP	random	degree_matched	0.801	0.454
GNN	random	degree_matched	0.758	0.440
XGBoost	cold_condition	NegBioDB	1.000	1.000
XGBoost	cold_drug	NegBioDB	1.000	0.999

Key finding: CT-M1 is trivially separable on NegBioDB negatives (AUROC=1.0 random/cold splits). Control negatives (degree_matched) reveal meaningful discrimination AUROC ~0.76–0.84.

ML Results (CT-M2: 7-way Failure Category)

Aggregated over 3 seeds (source: results/ct_table_m2_aggregated.csv):

Model	Split	Macro-F1	Weighted-F1	MCC
XGBoost	random	0.510	0.751	0.637
XGBoost	degree_balanced	0.521	0.758	0.645
XGBoost	cold_condition	0.338	0.686	0.570
XGBoost	cold_drug	0.414	0.683	0.555
XGBoost	scaffold	0.193	0.567	0.374
XGBoost	temporal	0.193	0.602	0.454
GNN	random	0.468	0.672	0.526
MLP	random	0.358	0.619	0.432

Key finding: XGBoost best for M2. Scaffold/temporal splits are hardest (macro-F1 ~0.19). Degree-balanced helps.

LLM Results (80/80 complete)

Task	Gemini	GPT-4o-mini	Haiku	Qwen-7B	Llama-8B
L1 (5-way MCQ, acc)	0.68	0.64	0.66	0.65	0.63
L2 (extraction, field_f1)	0.75	0.73	0.48	0.81	0.77
L3 (reasoning, /5.0)	~4.7	~4.6	~4.4	~4.5	~4.6
L4 (discrim, MCC)	0.56	0.49	0.50	0.48	0.50

Key finding: CT LLMs show meaningful discrimination (MCC ~0.5). L3 ceiling effect — all models score 4.4–4.7/5.0 (judge too lenient; scores not discriminative). L2 Qwen/Llama outperform API models on field F1.

---

PPI Domain (Protein-Protein Interaction)

Status as of 2026-03-23: ML 54/54 complete (seeds 42/43/44), LLM 80/80 complete, L3 judged.

Database

• 2,229,670 negative results in DB (IntAct 779 / HuRI 500K / hu.MAP 1.23M / STRING 500K); 2,220,786 export rows after split filtering
• 61,728 positive pairs (HuRI Y2H), 18,412 proteins

ML Results (PPI-M1: Binary Non-interaction Prediction)

Aggregated over 3 seeds (source: results/ppi/table1_aggregated.md):

Model	Split	Negatives	AUROC	MCC	LogAUC
SiameseCNN	random	NegBioDB	0.963 ± 0.000	0.794 ± 0.012	0.517 ± 0.018
PIPR	random	NegBioDB	0.964 ± 0.001	0.812 ± 0.006	0.519 ± 0.009
MLPFeatures	random	NegBioDB	0.962 ± 0.001	0.788 ± 0.003	0.567 ± 0.005
SiameseCNN	cold_protein	NegBioDB	0.873 ± 0.002	0.568 ± 0.019	0.314 ± 0.014
PIPR	cold_protein	NegBioDB	0.859 ± 0.008	0.565 ± 0.019	0.288 ± 0.010
MLPFeatures	cold_protein	NegBioDB	0.931 ± 0.001	0.706 ± 0.005	0.476 ± 0.005
SiameseCNN	cold_both	NegBioDB	0.585 ± 0.040	0.070 ± 0.004	0.037 ± 0.010
PIPR	cold_both	NegBioDB	0.409 ± 0.077	−0.018 ± 0.044	0.031 ± 0.019
MLPFeatures	cold_both	NegBioDB	0.950 ± 0.021	0.749 ± 0.043	0.595 ± 0.051
SiameseCNN	random	uniform_random	0.965 ± 0.001	0.806 ± 0.007	0.552 ± 0.002
PIPR	random	uniform_random	0.966 ± 0.000	0.810 ± 0.002	0.565 ± 0.005

Key findings:

• MLPFeatures dominates cold splits: AUROC 0.95 on cold_both (vs PIPR collapse to 0.41)
• PIPR catastrophic failure on cold_both: AUROC 0.41 (below random!)
• Control negatives (uniform_random, degree_matched) inflate LogAUC by +0.04–0.05

LLM Results (80/80 complete, post-hoc fixes applied 2026-03-23)

Task	Gemini	GPT-4o-mini	Haiku	Qwen-7B	Llama-8B
L1 (4-way MCQ, acc) zero-shot	0.75	0.75	0.75	0.75	0.75
L1 (4-way MCQ, acc) 3-shot	1.000	0.997	0.998	0.998	0.997
L2 (method_accuracy) 3-shot	1.000	~0.94	~0.08	~0.94	~0.08
L3 (reasoning, /5.0) zero-shot	4.4–4.7	4.4–4.7	4.3–4.7	4.3–4.7	4.3–4.7
L3 (reasoning, /5.0) 3-shot	3.1–3.7	3.1–3.7	3.1–3.7	3.1–3.7	3.1–3.7
L4 (discrim, MCC) zero-shot	0.43	0.430	0.38	0.36	0.441

Key findings:

• L1: 3-shot near-perfect is an artifact (in-context examples reveal pattern format)
• L2: Gemini/GPT/Qwen correctly identify interaction methods, Haiku/Llama fail
• L3: zero-shot > 3-shot (gold reasoning examples degrade performance); structural reasoning collapses to ~1.2/5 in 3-shot
• L4: All models show contamination (acc_pre_2015 >> acc_post_2020)

---

GE Domain (Gene Essentiality / DepMap)

Status as of 2026-03-24: ETL+ML complete, LLM 4/5 models done (Llama pending), L3 judged.

Database

• 28,759,256 negative results (CRISPR 19.7M + RNAi 9.1M)
• Final tiers: Gold 753,878 / Silver 18,608,686 / Bronze 9,396,692
• 22,549,910 aggregated pairs (19,554 genes × 2,132 cell lines)

ML Results (Seed 42 complete; seeds 43/44 in progress)

5 splits: random/cold_gene/cold_cell_line/cold_both/degree_balanced Models: XGBoost, MLPFeatures Aggregated results (3 seeds) pending; individual seed 42 results available in results/ge/.

LLM Results (4/5 models complete — Llama pending)

Task	Models Done	Key Finding
L1 (4-way MCQ, 1,200 items)	Haiku, Gemini, GPT, Qwen	Results pending Llama
L2 (field extraction, 500 items)	Haiku, Gemini, GPT, Qwen	Results pending Llama
L3 (reasoning, 200 items, judged)	Haiku, Gemini, GPT, Qwen	zero-shot >> 3-shot (4.5 vs 2.5 overall mean)
L4 (discrimination, 475 items)	Haiku, Gemini, GPT, Qwen	Results pending Llama

Expected key finding: GE L4 MCC likely intermediate between PPI and DTI given DepMap is a public widely-studied dataset with high training data overlap.

---

Methodology Notes

ML Evaluation Protocol

• Metrics: AUROC (primary), LogAUC (early enrichment), AUPRC, MCC, BEDROC
• Seeds: 3 seeds (42, 43, 44) for statistical robustness (except GE seed 42 only so far)
• Negative types: NegBioDB (structured negatives) vs control negatives (uniform_random, degree_matched)
• Splits: Random → Cold (one entity unseen) → Cold-Both (both entities unseen, hardest)

LLM Evaluation Protocol

• Models: Claude Haiku-4.5, Gemini 2.5-Flash, GPT-4o-mini, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct
• Configs: Zero-shot × 1 + 3-shot × 3 fewshot sets = 4 configs per (model × task)
• L3 judge: Gemini 2.5-Flash LLM-as-judge, 4 dimensions × 5-point scale
• L4 contamination test: Older vs newer entity pairs (pre-cutoff vs post-cutoff)