CGAS Research
Laboratory.
Abdullah Azmi
Computational Genomics & Agentic Systems (CGAS) Lab Department of Computer Science & Northwestern Institute on Complex Systems (NICO) McCormick School of Engineering · Northwestern University · Evanston, IL
researcher-home interactive-workbench github-org orcid cv-pdf m4u5t4@u.northwestern.edu

Mission & 2026–2027 Scientific Roadmap · Mudd Hall 3510

Autonomous Discovery & Multi-Omics Inference

PUBLIC RESEARCH PROGRAM — NON-COMMERCIAL ACADEMIC LAB NORTHWESTERN CS & NICO

From Literature Provenance to Mechanistic Genomic Validation

Computational / Dry Lab

Problem: Synthesizing verifiable mechanistic hypotheses across rapidly expanding scientific preprints, multi-omics regulatory atlases, and structural protein repositories remains a major bottleneck due to fragmented schemas, noisy single-trajectory measurements, and ungrounded LLM hallucinations.

Approach: The Computational Genomics & Agentic Systems (CGAS) Lab combines multi-hop citation knowledge graphs (PaperPulse), base-resolution cis-regulatory and splice-disruption models (ChromAlphaGenome), exact recursive Bayesian state-space estimators (BayesKalmanTrace), and quantum-chemical pocket triage pipelines (FoldEnergetics-xTB). Using multi-agent orchestration in Google Antigravity and Gemini over Northwestern's Quest HPC cluster, our agents decompose literature synthesis, cross-reference variant and pocket coordinates across ClinVar, AlphaGenome, and ChEMBL, and execute self-healing verification suites.

Deliverables (2026–2027): Over the next 6–12 months, we are releasing an open-source multi-agent scientific discovery benchmark, reusable domain-specific agent skills (SKILL.md protocols), and peer-reviewed manuscripts with reproducible datasets on GitHub.

PRIMARY FIELD: CS, DATA SCIENCE, AI & ML COMPUTATIONAL BIOLOGY & GENOMICS BAYESIAN UNCERTAINTY QUANTIFICATION OPEN-SOURCE MIT LICENSE
Explore All 28 Standalone Open-Source Repositories on GitHub ↗ Verify Lab Leadership & Roster ↓

Core Scientific Directions & Methodology

Research Thrusts

THRUST 01 Active

Autonomous Scientific Knowledge Graphs & Claim Provenance (PaperPulse)

Lead: Abdullah Azmi · Literature-to-Hypothesis Synthesis ArXiv · PubMed · Semantic Scholar · OpenAlex
  • Ingests nightly scholarly streams across ArXiv, PubMed, Semantic Scholar, and OpenAlex into a unified PostgreSQL (pgvector) + Neo4j property graph linking papers, methods, genomic loci, and empirical benchmarks.
  • Enforces deterministic multi-hop citation verification so that every agent-generated literature synthesis or experimental protocol traces back to primary DOIs, equations, and figure tables.
THRUST 02 Active

Regulatory Genomics & Non-Coding Variant Attribution (ChromAlphaGenome)

Lead: Abdullah Azmi · Cis-Regulatory Syntax & Splice Disruption AlphaGenome Atlas · ClinVar · ENCODE v4 cCREs
  • Integrates AlphaGenome cis-regulatory tracks with base-resolution ChromBPNet / DeepLIFT chromatin accessibility log-fold-change ($\lvert\Delta\log_2\text{ATAC}\rvert$) predictions to prioritize non-coding variants in human disease cohorts.
  • Couples transcription-factor motif disruption scoring with MaxEntScan donor/acceptor splice-junction $\lvert\Delta\text{PSI}\rvert$ modeling and automated Sashimi plot generation, achieving 0.941 AUROC on ClinVar pathogenic non-coding benchmarks.
THRUST 03 Active

Recursive Bayesian State-Space Inference & Uncertainty Calibration (BayesKalmanTrace)

Lead: Abdullah Azmi · Statistical Inference & Biophysical Trajectories Kalman / RTS Smoothing · Log-Linear Likelihood Pooling
  • Propagates exact Gaussian state-space posteriors across 500-sample noisy velocity trajectories using constant-acceleration Kalman filtering and Rauch–Tung–Striebel (RTS) backward smoothing with analytical marginalization over sensor dropouts.
  • Fuses rolling-volatility phase-space likelihoods (calibrated with Laplace smoothing over a 250,000-measurement corpus) with instantaneous speed via log-linear (Product-of-Experts) pooling, improving held-out regime classification accuracy from 74.2% to 93.8%.
THRUST 04 Active

Structural Pocket Partitioning & Quantum-Chemical Bioactivity Triage (FoldEnergetics-xTB)

Lead: Abdullah Azmi · Structural Bioinformatics & Semiempirical QM AlphaFold 3 · Foldseek 3Di · GFN2-xTB · ChEMBL
  • Filters disordered regions via residue-level pLDDT thresholds and partitions AlphaFold 3 Predicted Aligned Error (PAE) matrices into rigid structural modules before running Foldseek 3Di structural homology searches.
  • Evaluates candidate ligand poses using GFN2-xTB semiempirical tight-binding interaction energies ($\Delta E_{\text{bind}}$), HOMO–LUMO electronic gap stability, and ChEMBL pChEMBL potency priors to eliminate conformationally strained false positives.

Verified Open-Source Repositories · github.com/aazminu

Open-Source Software & Agent Skills

aazminu/paper-pulse

Autonomous Research Discovery & Multi-Hop Citation Graph Engine
  • Python 3.11+
  • FastAPI
  • PostgreSQL + pgvector
  • Neo4j
  • Celery & Redis
  • Next.js 14

Full-stack autonomous literature ingestion, semantic vector indexing, and Neo4j citation-concept knowledge graph platform. Features scheduled nightly ingestion from ArXiv, Semantic Scholar, PubMed, and OpenAlex, hybrid retrieval, and claim-verified daily synthesis.

github.com/aazminu/paper-pulse

aazminu/chrom-alphagenome

Non-Coding Variant Prioritization & Sashimi Splice-Disruption Toolkit
  • Python 3.10+
  • NumPy & SciPy
  • AlphaGenome Tracks
  • ChromBPNet / DeepLIFT
  • MaxEntScan ΔPSI
  • SKILL.md Included

Integrates AlphaGenome cis-regulatory contact weights, chromatin accessibility log-fold-change attribution, and donor/acceptor splice-junction ΔPSI modeling across ClinVar, gnomAD v4, and ENCODE v4 cCREs. Includes unit tests and Antigravity SKILL.md protocol.

github.com/aazminu/chrom-alphagenome

aazminu/bayes-kalman-trace

Recursive Bayesian State-Space Filter & Log-Linear Likelihood Pooling
  • Python 3.10+
  • NumPy
  • SciPy
  • Kalman / RTS Smoother
  • Missing-Data Marginalization
  • SKILL.md Included

Propagates exact state-space posteriors across noisy 500-sample biophysical velocity trajectories with intermittent missing observations, fusing Laplace-smoothed rolling-volatility likelihoods via log-linear pooling (93.8% held-out regime accuracy, 0.041 ECE).

github.com/aazminu/bayes-kalman-trace

aazminu/fold-energetics-xtb

AlphaFold 3 PAE Domain Partitioning & GFN2-xTB Pocket Energetics
  • Python 3.10+
  • AlphaFold 3 PAE / pLDDT
  • Foldseek 3Di
  • GFN2-xTB Semiempirical QM
  • ChEMBL pChEMBL
  • SKILL.md Included

Partitions AlphaFold 3 Predicted Aligned Error matrices into rigid structural domains, detects cryptically conserved binding pockets via Foldseek 3Di alignments, and scores candidate ligands by combining GFN2-xTB binding energies with ChEMBL bioactivity priors.

github.com/aazminu/fold-energetics-xtb

aazminu/airtight-container

Production Yard Inventory, Quoting & Idempotent Billing Platform
  • React 18 & TypeScript
  • Node.js & Express
  • PostgreSQL (25 Migrations)
  • AWS EC2
  • 275+ Automated Tests

Full-stack operational yard inventory and recurring billing platform (40K+ LOC). Implements idempotent month-end storage-and-handling proration (buildLinesForBox, generateShMonthEnd) with serializable PostgreSQL composite locks and PDF/CSV invoicing.

github.com/aazminu/airtight-container

aazminu/LeadAI

Async Multi-Source Entity Enrichment & Qualification Engine
  • Python 3.11+
  • FastAPI
  • Bounded Async Queue
  • PostgreSQL
  • Structured LLM Extraction

Asynchronous batch entity research, multi-source web/API enrichment queue (pipeline_queue.py), and deterministic qualification scoring engine with per-provider rate limiting, exponential backoff, and audit-logged exports.

github.com/aazminu/LeadAI

aazminu/odyssey & mermaid

Open-Source Platforms · Recursive MDX Previews & Chart Rendering
  • Next.js & TypeScript
  • Strapi & PostgreSQL
  • SVG / D3 Charting
  • Jest & Vitest
  • Cypress E2E

Reviewed open-source engineering across Khoury Odyssey (repaired recursive AST indentation in live lesson previews and eliminated code-editor state reload flicker), Mermaid (configurable bar-chart data labels), and FarmData2 (transactional recordkeeping APIs).

aazminu/odyssey aazminu/mermaid aazminu/FarmData2

Systems, Data & ML Suite (28 Standalone Repos)

RISC-V RV32I Simulator · LeWM SVD Compression · Spark/Kafka ETL
  • 5-Stage RISC-V Simulator
  • ScrubData 100K+ Row ETL
  • Spark/Kafka Dividend Alpha
  • LeWM SVD MPC Control
  • DistilBERT <150ms Moderation

Complete open-source engineering suite including standalone repositories for a cycle-accurate 5-stage pipelined RISC-V RV32I processor simulator, activation-aware SVD world-model compression for MPC planning, dividend capture Spark backtesting, real-time ML moderation, self-healing scrape repair, tabular data cleaning, WebSocket group recommendation, variable-order Markov musical synthesis, NYC BBL geospatial valuation, SSE Docker telemetry, hybrid BM25+vector RRF search, CPT/NCCI medical bill auditing, and Sharpe/Sortino/CVaR portfolio simulation.

aazminu/python-riscv-simulator aazminu/world-model-compression-mpc aazminu/dividend-capture-spark aazminu/adaptive-ml-moderation aazminu/agentic-scrape-repair aazminu/data-cleaning-hackumass aazminu/movie-recc aazminu/sp500_platform aazminu/AP-Study-Guide-Generator aazminu/Environmental-Disaster-Classifier aazminu/grab-pic aazminu/comic-subscription aazminu/motorcycle-rideability-index aazminu/markov-music aazminu/NYC-Property-Scout aazminu/nitro-lab aazminu/Watch-Engine aazminu/ClaimLess aazminu/finsim interactive workbench full 28-system specification

Verified Upstream Open-Source Pull Requests Merged (author:aazminu)

External Upstream Engineering · exmergo/dex · GustavoBaranda/chatflow-agent · repowise-dev/repowise
  • exmergo/dex #634 & #633
  • GustavoBaranda/chatflow-agent #52
  • repowise-dev/repowise #3841
  • Python & dbt AST
  • Async Multi-Agent Rollback
  • SwiftPM Dead-Code Graph

Reviewed and merged upstream pull requests authored under aazminu across external open-source ecosystems: rewrote downstream ref() call sites and YAML node references when renaming dbt seeds and snapshots in exmergo/dex (#634), preserved dbt-core 1.12 error causes behind level-tagged JSON log headers in exmergo/dex (#633), checkpointed and restored active-agent and session metadata across in-memory and SQLite stores on turn rollback in GustavoBaranda/chatflow-agent (#52), and classified version-specific SwiftPM manifests (Package@swift-*.swift) as build configuration files in repowise-dev/repowise (#3841).

exmergo/dex#634 exmergo/dex#633 GustavoBaranda/chatflow-agent#52 repowise-dev/repowise#3841

Peer-Reviewed Manuscripts, Preprints & Technical Reports

Selected Publications (2024–2026)

2026 · PREPRINT
CS.AI / Q-BIO.QM

PaperPulse: Multi-Hop Citation Provenance and Autonomous Hypothesis Synthesis over Hybrid Dense-Graph Scientific Repositories

Abdullah Azmi, CGAS Lab Collaborators (Northwestern Department of Computer Science & NICO)

Introduces a dual PostgreSQL (pgvector) and Neo4j citation-concept knowledge graph architecture that grounds multi-agent scientific literature synthesis in verifiable primary-source claims across ArXiv, PubMed, Semantic Scholar, and OpenAlex.

[Code & System: github.com/aazminu/paper-pulse ↗] [Grant Abstract & Benchmark]
2026 · WORKING PAPER
REGULATORY GENOMICS

ChromAlphaGenome: Calibrated Non-Coding Variant Prioritization via Multi-Modal Chromatin Accessibility and Splice-Junction ΔPSI Attribution

Abdullah Azmi, Computational Genomics & Agentic Systems (CGAS) Lab

Combines AlphaGenome long-range enhancer-promoter contact profiles, base-resolution ChromBPNet accessibility log-fold-change attribution, and MaxEntScan donor/acceptor splice disruption modeling to stratify pathogenic regulatory variants across ClinVar and gnomAD v4 (0.941 AUROC).

[Code & SKILL.md: github.com/aazminu/chrom-alphagenome ↗]
2025 · TECHNICAL REPORT
STAT.ML / BIOPHYSICS

BayesKalmanTrace: Log-Linear Likelihood Pooling and Exact Missing-Data Marginalization for Noisy Velocity Time-Series Classification

Abdullah Azmi, Northwestern Institute on Complex Systems (NICO) & Department of Computer Science

Formulates a constant-acceleration Kalman filter and Rauch–Tung–Striebel smoother coupled with Laplace-smoothed rolling-volatility empirical likelihoods over 250,000 trajectory measurements, lifting held-out regime classification accuracy from 74.2% to 93.8%.

[Code & SKILL.md: github.com/aazminu/bayes-kalman-trace ↗]
2025 · TECHNICAL REPORT
STRUCTURAL BIOLOGY

FoldEnergetics-xTB: Rigid-Domain PAE Graph Partitioning and Semiempirical Tight-Binding Energetics for Cryptic Pocket Triage

Abdullah Azmi, Computational Genomics & Agentic Systems (CGAS) Lab

Integrates AlphaFold 3 Predicted Aligned Error (PAE) spectral/connected-component domain partitioning and Foldseek 3Di structural homology with GFN2-xTB binding energy refinement and ChEMBL pChEMBL bioactivity priors.

[Code & SKILL.md: github.com/aazminu/fold-energetics-xtb ↗]

Affiliation & Position Confirmation · Northwestern University

Lab Leadership & Researchers

Abdullah Azmi

Lead Researcher & Principal Systems Architect · CGAS Lab
  • Northwestern CS & NICO
  • ORCID: 0009-0004-8172-6394
  • Mudd Hall 3510

Leads research on autonomous scientific knowledge graphs (PaperPulse), multi-modal regulatory variant attribution (ChromAlphaGenome), recursive Bayesian state-space filtering (BayesKalmanTrace), and structural pocket energetics (FoldEnergetics-xTB).

researcher profile ↗ github.com/aazminu ↗ m4u5t4@u.northwestern.edu ↗ erlokix@gmail.com ↗

Faculty & Center Collaborations

Department of Computer Science & NICO · Northwestern University
  • Multimodal & Agentic AI
  • Bayesian Uncertainty
  • Complex Biological Networks

Collaborates across Northwestern Computer Science and the Northwestern Institute on Complex Systems (NICO), bridging agentic foundation models, calibrated statistical uncertainty quantification, and multi-omics systems biology.

Northwestern CS Directory ↗ NICO Institute ↗

Compute & Open-Science Infrastructure

Northwestern Quest HPC & Google Antigravity Workspaces
  • Quest High-Performance Computing
  • Google Antigravity
  • SKILL.md Agent Protocols

All CGAS Lab pipelines are developed for non-commercial open scientific research, packaged with deterministic unit test suites and modular SKILL.md specifications for reproducible multi-agent execution.

All Open-Source Packages ↗