EL-CAP

Evidence-Linked, Context-Aware Molecular Prioritization

EL-CAP is an intelligent platform that transforms the way pharmaceutical and chemical companies discover and prioritize new molecules.
By combining graph neural networks (GNN) for molecular structure analysis with transformer-based NLP for understanding scientific context, EL-CAP delivers evidence-linked, context-aware rankings of potential drug candidates.

Whether you start with a research question, a candidate molecule, or both, EL-CAP can:

• 🚀 Search millions of compounds in seconds using a vector database (FAISS / ChromaDB)
• 📚 Link every result to real-world evidence: assays, publications, patents
• 🧪 Predict affinity, ADMET, novelty, and safety risks through multi-task learning
• 📊 Generate a single composite score to guide decision-making

From massive chemical libraries to ranked, evidence-backed shortlists — in one step. EL-CAP is a two-stage retrieval → scoring system with multimodal encoders (GNN for molecules, Transformer for text), a vector store for fast candidate recall, and multi-task heads plus a meta-model for final prioritization.

Architecture review

flowchart TD
subgraph A[Data Sources]
A1[ChEMBL / BindingDB] 
A2[PubMed / Patents] 
A3[ELN / Internal Data]
end

A --> B[Ingestion & Curation]
B --> C1[Molecule Standartization]
B --> C2[Assay Normalization]

subgraph D[Feature Extraction Encoders]
  D1[GNN Encoder + SSL Pretraining + ECFP fusion Output: h_mol + atom tokens]
  D2[Transformer Encoder + Domain MLM tuning + Slot Extraction]
end

C1 --> D1
C2 --> D2

D1 --> E1[Molecular Embeddings h_mol]
D2 --> E2[Text Embeddings h_txt]

%% ========== VECTOR STORES ==========
subgraph F[Storage Layers]
  F1[Mol Vector Store]
  F2[Passage Vector Store]
  F3[Tabular Store.Assay schema, labels, hard filters]
end

E1 --> F1
E2 --> F2
C2 --> F3

%% ========== QUERY PIPELINE ==========
subgraph Q[Query Handling]
  Q1[User Query: text and/or molecule]
  Q2[Mol Preproc → h_mol*]
  Q3[Text Encode → h_query]
  Q4[Mol+Text Fusion. Linear proj / light MLP]
end

Q1 --> Q2
Q1 --> Q3
Q2 --> Q4
Q3 --> Q4

%% Retrieval
Q3 --> R1[Text-only Retrieval. ANN on Mol Index]
Q2 --> R2[Mol-only Retrieval. ANN on Mol Index]
Q4 --> R3[Mol+Text Retrieval. ANN on Mol Index]

R1 --> S1[Top-K Candidates]
R2 --> S1
R3 --> S1

%% ========== RERANK / SCORING ==========
subgraph S[Scoring Pipeline]
  S1 --> S2[Cross-Encoder. Token match: text tokens ↔ atom tokens]
  S2 --> S3[Fusion Layer → MLP]
  S3 --> S4[Multi-Task Heads. Affinity, ADMET, Risk, Novelty. Masked Loss for missing labels]
  S4 --> S5[Meta-Model. Composite Priority Score + Confidence]
end

%% ========== OUTPUT ==========
S5 --> O1[Ranked Molecule List+ flags]
S5 --> O2[Evidence Linking. Retrieve passages from Text Index]
F2 --> O2
F3 --> O1

O1 --> U1[User Dashboard / API]
O2 --> U1

Loading

Search Modes

EL-CAP exposes three complementary retrieval modes that feed the same scoring pipeline.

---

1) Text-Only (context → molecules)

Input: free-form requirements (target, assay conditions, TPP constraints)
Query vector: h_query = NLP(text)
Index queried: molecular vectors h_mol (vector store)

Use cases: early triage from literature/TPP; no starting structure
Pros: broad recall from context; fastest ideation
Cons: semantic gap (text ↔ structure) without good contrastive pretraining

Flow:
text → NLP → h_query → ANN(h_query, {h_mol}) → top-K molecules → scoring → rank

---

2) Mol-Only (molecule → analogs/evidence)

Input: candidate SMILES/SDF
Query vector: h_mol* = GNN(mol)
Index queried: molecular vectors (analogs) and/or text vectors (evidence)

Use cases: lead optimization, IP proximity, explainability
Pros: robust for scaffold-centric workflows
Cons: lacks explicit context; mitigated at scoring stage via fusion with optional text

Flow:
mol → GNN → h_mol* → ANN(h_mol*, {h_mol}) → top-K → (optional ANN to texts) → scoring → rank

---

3) Mol+Text (joint context)

Input: molecule + requirements (target/assay/ADMET rules)
Query vector: h_qry = g(h_mol*, h_query) via concat/MLP or light cross-attention
Index queried: molecular (and optionally text) vectors using h_qry

Use cases: context-aware analog search; best Recall@K under constraints
Pros: highest relevance; aligns structure with requirements before scoring
Cons: most complex training (needs positive (mol, context) pairs and hard negatives)

Flow:
mol → GNN → h_mol* text → NLP → h_query h_qry = g(h_mol*, h_query) → ANN(h_qry) → top-K → scoring → rank

---

Retrieval → Scoring bridge (common to all modes)

• Retrieval output: top-K mol_id (+ cached h_mol, metadata, candidate evidence)
• Scoring input: [h_mol, h_query, structured features] → Fusion → Heads → Meta-Model
• Output: composite score + evidence cards (linked assay snippets, patents, PAINS/Lipinski/hERG/CYP flags)

---

KPIs by stage

• Retrieval: Recall@K, NDCG@K, latency p95
• Scoring: RMSE (affinity), ROC-AUC/PR-AUC (ADMET/risk), calibration (ECE), cost/1k queries

---

Practical defaults

• Embedding dim: 384–768; cosine similarity with L2-normalized vectors
• ANN: HNSW or IVF-PQ (GPU) depending on scale; K=200–1000 for rerank window
• Fusion: start with concat + MLP; upgrade to cross-attention for localized effects
• Meta-model: linear regression for interpretability; LightGBM for nonlinearity if needed