Portfolio classification: original research implementation. Aishwarya Anand developed AAPred as her NTU Final Year Project, covering the model architecture, leakage-aware evaluation, reproducibility work, and research communication.

AAPred

Antibody-antigen binding affinity prediction with protein language models, 3D residue graphs, and Graph-Mamba.

AAPred was developed as Aishwarya Anand's Final Year Project at Nanyang Technological University (NTU), supervised by Prof. Jagath C. Rajapakse. The project received an A+ grade and contributed to the GEPBind manuscript submitted to ECCB 2026.

The implementation predicts binding free energy (delta_g, in kcal/mol). This is Delta G, rather than mutation-induced Delta Delta G.

Why It Matters

Measuring antibody-antigen affinity experimentally is expensive and slow. AAPred explores computational pre-screening by combining sequence, physicochemical, and structural information in one regression model.

The implemented AAPred pipeline combines:

• ESM-2 650M residue embeddings for antibody light chain, antibody heavy chain, and antigen sequence context.
• AAindex PCA features for residue-level physicochemical information.
• 3D residue graphs built from C-alpha radius neighborhoods, with rotation-invariant radial basis encodings of pairwise distance.
• GraphGPS / Graph-Mamba layers using local GatedGCN message passing and a global Mamba state-space component.
• Cross-attention fusion between sequence modalities and between sequence and graph representations.
• Leakage-aware evaluation using GroupKFold by PDB complex identifier.

The main implementation is in aapred/.

Results

The strongest directly traceable AAPred result in this archive is the grouped 10-fold cross-validation run:

Model	Protocol	RMSE (kcal/mol)	Pearson r
AAPred HybridGraphMamba	10-fold GroupKFold by PDB ID	1.824 +/- 0.139	0.507 +/- 0.078
Published MVSF-AB reference	Paper-reported SAbDab result	1.839	0.491

The split protocols and preprocessing are not guaranteed to be identical, so the baseline row is contextual rather than a strict head-to-head reproduction. See the RESULTS_EVIDENCE_LEDGER.md for the complete provenance and caveats.

The repository also contains 45 Graph-Mamba ablation configurations. The configuration files are preserved, but results for all 45 runs are not present in the supplied archive.

AAPred And GEPBind

AAPred is the Graph-Mamba FYP implementation preserved here. GEPBind is the later manuscript system that evolved from this work and uses a different final graph configuration. They are related, but they should not be treated as the same evaluated model.

The manuscript draft is preserved at docs/manuscript/GEPBind_ECCB_2026.md.

Repository Map

Path	Contents
aapred/	Canonical, cleaned AAPred source and run scripts
research/final-submission/	Final FYP evidence package, original AAPred copy, baselines, plots, and thesis notes
research/development/	Historical model branches and exploratory experiments
research/baselines/	Earlier baselines and comparative research code
research/results/	Consolidated result tables
docs/	Thesis, defense slides, manuscript, and provenance
auxiliary/inference/	Auxiliary inference configuration and input lists
notebooks/	Exploratory notebooks
environment/	Reference environment configuration

The historical folders are retained intentionally. They contain evidence, failed experiments, alternative model families, and intermediate results that are useful for auditability.

Quick Validation

The source-level validation does not download models or run training:

cd aapred
bash scripts/run_smoke_checks.sh

The supplied archive does not contain the processed graph tensor, source FASTA, or cleaned structure files needed for a new end-to-end training run. The smoke check reports this as a data-readiness warning while still validating the available source. See aapred/DATA_AVAILABILITY.md.

Full Environment

The audited training environment used Python 3.10 on a CUDA system. Install PyTorch and PyTorch Geometric wheels that match the target CUDA runtime, then:

python -m pip install -r aapred/requirements.txt

The full ESM-2 model is downloaded from Hugging Face on first use.

Preservation

This repository is a lossless reorganization of the supplied research workspace: original research files were moved into clear ownership areas, not discarded. The source ZIP identity, omissions declared by the original handoff, and publication exclusions are documented in docs/provenance/.

Machine-specific Claude session telemetry is retained locally under .private/ but intentionally excluded from GitHub. A hard-coded Jupyter token was removed from the launcher.

Citation

Use CITATION.cff for the software project. Refer to the manuscript draft for the GEPBind author list and paper-specific citation.