8000
Skip to content
View abinittio's full-sized avatar

Highlights

  • Pro

Block or report abinittio

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
abinittio/README.md

Nabil Yasini-Ardekani — computational chemistry × machine learning

I build machine-learning models for molecules and materials, and I spend most of my effort on the part that decides whether they mean anything: the evaluation. Leakage-controlled splits, ablations that isolate one variable, pre-registered decision rules, and negative results reported as negative.


Selected work

Project What it is The result that matters
FiltraNex · live demo Five-stage screening cascade for CNS stimulant candidates, with an applicability-domain gate Predictive accuracy is a function of distance from training chemistry — so every result ships with a confidence tier instead of a bare number
StereoAwareGNN Protocol-controlled BBB permeability benchmark: 8 architectures × 3 evaluation protocols ChemBERTa tops the random split at 0.958 and drops to 0.746 external; a fingerprint random forest gives up 0.005 and finishes first. Protocol moves the score more than architecture does
bbb-honest-eval The leakage audit behind that correction Reproduces an inflated 0.96 "external" AUC, isolates the dataset overlap that caused it, and publishes the corrected figures
In-silico drug discovery toolkit Multi-endpoint ADMET platform: transporters, abuse liability, hERG, CYP450, BBB Transporter ROC-AUC 0.968 (scaffold split, in-dataset), plus a SMARTS rules engine for known failure modes
CYP450 metabolism predictor Multi-task GNN over five major CYP isoforms Drug–drug interaction liability from structure alone
DoseTrack PK/PD simulation for lisdexamfetamine: prodrug conversion, RK4 integration, tolerance dynamics A continuous physiological state estimator behind a minimal logging interface
MGT for CO₂ capture (private until examination) Ablation study of the Molecular Graph Transformer's long-range attention channel on 25,000 metal–organic frameworks The channel is worth 24% lower error — but the gain appears in only 2 of 3 random initialisations, survives a capacity-matched control, and is not improved by feeding it real partial charges

How I work

  • Controls before conclusions. If two models differ in more than one way, the comparison cannot attribute anything. Ablate one variable, hold the rest fixed.
  • Pre-register the decision rule. Minimum effect size and pass/fail criteria go into the repository before the runs, so a negative result cannot be reinterpreted later.
  • Publish the correction. When a headline number turns out to be leakage, the retraction stays in the record next to the corrected figure.
  • Say what a model cannot do. Applicability domains, failure modes and out-of-scope chemistry are part of the deliverable, not an appendix.

Toolchain

PyTorch · PyTorch Geometric · DGL · RDKit · pymatgen · scikit-learn · pandas · Flask / Streamlit · GPU work on Kaggle and Colab

Elsewhere

nabil.engineer · LinkedIn

Popular repositories Loading

  1. Insilico-Drug-Discovery-Toolkit Insilico-Drug-Discovery-Toolkit Public

    Integrated ADMET platform: monoamine transporter (DAT/NET/SERT) classification, abuse liability, hERG, CYP450 and BBB, with a SMARTS pharmacology rules engine. Transporter ROC-AUC 0.968, scaffold s…

    Python 2

  2. StereoAwareGNN StereoAwareGNN Public

    Protocol-controlled BBB permeability benchmark: 8 architectures across random, scaffold and leakage-controlled external splits. Evaluation protocol moves the score more than architecture does.

    Python 1

  3. CYP450-Metabolism-Predictor CYP450-Metabolism-Predictor Public

    Multi-task GNN predicting inhibition of 5 major cytochrome P450 enzymes (CYP1A2, CYP2C9, CYP2C19, CYP2D6, CYP3A4) for drug-drug interaction assessment.

    Python 1

  4. abinittio abinittio Public

    Profile README

  5. dosetrack-v4 dosetrack-v4 Public

    Minimal PK/PD dose tracker. A continuous physiological state estimator with a minimal logging interface. Log a dose, see the curve.

    Python

  6. bbb-honest-eval bbb-honest-eval Public

    Leakage-controlled audit of a BBB permeability GNN: reproduces the inflated external AUC, isolates the overlap that caused it, and reports the corrected figures.

    Python

0