ROADMAP.md

NegBioDB -- Roadmap

Last updated: 2026-03-30

Completed (Phase 1)

DTI Domain (Drug-Target Interaction)

• 30.5M negative results from 4 sources (ChEMBL, PubChem, BindingDB, DAVIS)
• ML baselines: DeepDTA, GraphDTA, DrugBAN across 5 splits + 2 negative controls
• LLM benchmark: L1-L4 tasks, 5 models, zero-shot and 3-shot configs
• Key finding: degree-matched negatives inflate LogAUC by +0.112

CT Domain (Clinical Trial Failure)

• 132,925 failure results from 216,987 trials (AACT, CTO, Open Targets, Shi & Du)
• ML baselines: XGBoost, MLP, GNN across M1 (binary) and M2 (7-way) tasks
• LLM benchmark: L1-L4, 5 models
• Key finding: NegBioDB negatives trivially separable (AUROC=1.0); M2 macro-F1=0.51

PPI Domain (Protein-Protein Interaction)

• 2.2M negative results from 4 sources (IntAct, HuRI, hu.MAP, STRING)
• ML baselines: Siamese CNN, PIPR, MLPFeatures across 4 splits
• LLM benchmark: L1-L4, 5 models
• Key finding: PIPR AUROC drops to 0.41 on cold_both; temporal contamination detected

GE Domain (Gene Essentiality / DepMap)

• 28.8M negative results from DepMap CRISPR and RNAi screens
• ML baselines: XGBoost, MLPFeatures (seed 42 complete)
• LLM benchmark: 4/5 models complete
• Key finding: cold-gene splits reveal severe generalization gaps

In Progress

• GE domain: remaining ML seeds (43/44) and Llama LLM runs
• Paper preparation for NeurIPS 2026 Evaluations & Datasets Track

Planned

Phase 2: Community & Platform

• Web interface for search, browse, and download
• Python library: pip install negbiodb
• REST API with tiered access
• Community submission portal
• Public leaderboard

Phase 3: Scale

• Expand curated entries via literature mining
• Specialized bio-LLM evaluations
• Cross-track ensemble evaluation (ML + LLM)