01 / Overview
Status: stages 1–5 complete · stage 6 partial · weight coverage 0.70MemoryRx
An open, zero-lab computational pipeline for identifying candidate small molecules against Alzheimer's disease
An open computational pipeline that prioritised Alzheimer's disease targets from public evidence, trained a QSAR model on CDK5 bioactivity data, designed ten derivative compounds, docked them against a CDK5 crystal structure, and profiled them for predicted brain penetration — with the honest result that none of the ten meets accepted CNS-permeability heuristics.
Every value on this page is a computational prediction. No compound here is a drug; all require experimental validation.
02 / Research question
Can computational screening identify blood-brain-barrier-permeable small molecules with predicted activity against several molecular pathways linked to Alzheimer's disease?
Context and gap
Disease-modifying anti-amyloid therapies exist, but Alzheimer's disease remains complex, and tau-related pathways, neuroinflammation and small-molecule approaches remain of high interest. That is the gap this project was designed around: two targets in two different pathways, approached with public data and free tooling only.
Answer, plainly
For this scaffold series, no. The modifications that improved predicted CDK5 affinity were polar groups that pushed polarity past CNS limits. All ten candidate compounds fail the CNS-permeability heuristic.
03 / Pipeline
Six stages, conceptually
Each stage used public resources only. Stage 6 is partial, and stages 3 and 5 cover CDK5 alone.
-
STEP 1
Alzheimer's Target Atlas
CompleteAssembled a candidate target list for Alzheimer's disease from public association evidence, grouped by molecular pathway. Scored each gene against weighted criteria to produce a transparent, reproducible ranking rather than a literature impression.
Open Targets Platform
-
STEP 2
Compound Database & Target Structures
CompleteCollected measured bioactivity data for the selected targets and retrieved experimentally determined human protein structures. Structures were required to have a defined ligand-binding site usable for docking.
ChEMBL · PubChem · RCSB PDB
-
STEP 3
QSAR Model
Complete — CDK5 onlyTrained classification and regression models relating molecular structure to measured CDK5 inhibition. Validated by held-out test set and cross-validation. No equivalent model exists for TREM2.
RDKit · scikit-learn · XGBoost
-
STEP 4
De novo Design
CompleteTook the most active compound in the CDK5 dataset as a parent scaffold and attached small functional groups at open positions. Variants were filtered through the trained classifier and oral drug-likeness rules.
RDKit · in-house QSAR classifier
-
STEP 5
Molecular Docking
Complete — CDK5 onlyDocked every designed candidate into the CDK5 ATP pocket under one fixed protocol, with the structure's own co-crystallised reference ligand re-docked identically so scores have an internal comparator.
AutoDock Vina / PyRx · RCSB PDB 1UNG
-
STEP 6
BBB & ADME Profiling
PartialComputed physicochemical descriptors and compared them with accepted CNS-permeability heuristics. Toxicity predictions and several ADME fields were not run, so this stage remains incomplete.
SwissADME · RDKit descriptors
04 / Target selection
Fifty genes, seven pathways, two targets
Fifty genes were drawn from Open Targets across seven pathways, narrowed to sixteen with usable human PDB structures, then scored on five weighted criteria.
The seven pathways
Scoring criteria and weights
Selection logic
TREM2 and CD33 both score 4.15, but both sit in neuroinflammation, so they cannot be paired under the two-different-pathways constraint. The selected pair is TREM2 (neuroinflammation, 4.15) plus CDK5 (tau pathology, 3.80).
Stage-3 criterion scores are expert judgement on a 1–5 scale and were revised during verification against Open Targets. Weights: G 0.25, D 0.25, S 0.20, M 0.15, N 0.15.
Table scrolls horizontally · column headers sort
Target structures
CDK5 — PDB 1UNG, X-ray 2.2 Å, (R)-roscovitine bound in the ATP pocket.
TREM2 — PDB 6XDS, X-ray 2.0 Å, fragment-bound Ig domain.
The TREM2 arm did not complete
TREM2 has far fewer characterised small-molecule modulators than a kinase. Public queries returned no usable human bioactivity set, so no QSAR model and no docking were possible: no model, insufficient public ligand data. Every downstream result on this page concerns CDK5 only, and the multi-target premise is unrealised.
05 / QSAR model
A usable classifier, a weak regressor
Models were trained on measured CDK5 bioactivity. Actives were defined as IC50 at or below 1000 nM.
TREM2: no model, insufficient public ligand data.
Interpretation
Classification performs well, separating actives from inactives at ROC-AUC 0.915 on the test set and 0.898 cross-validated, above the 0.80 usually treated as acceptable. Regression is weak, R² 0.45–0.47, so the model ranks compounds usefully but is not a quantitative potency predictor and is used as a classifier only.
06 / De novo design
Ten derivatives of one parent scaffold
The parent scaffold is CHEMBL384304, the most active compound against CDK5 in the dataset, designated M1. Small functional groups were attached at open positions and the variants filtered through the classifier and Lipinski's Rule of Five.
P(active) is the classifier's predicted probability of CDK5 activity, not a measured potency. Safety: not yet assessed.
Oral drug-likeness is not brain penetration
All ten candidate compounds pass the Rule of Five with zero violations and carry predicted activity 0.974–0.986 — yet the same ten fail CNS heuristics in stage 6. Oral drug-likeness and brain penetration are different filters.
07 / Molecular docking
Scores that sit inside the method's error
AutoDock Vina against 1UNG, a 20 × 20 × 20 Å box on the ATP pocket, exhaustiveness 8, with the reference ligand (R)-roscovitine re-docked under the identical protocol.
Docking score in kcal/mol; Δ is relative to the re-docked reference at −7.54. A gain of 0.03–0.65 kcal/mol is inside AutoDock Vina's typical error of roughly 1–2 kcal/mol.
Docking score magnitude, zero baseline
What this does not show
Eight of ten candidates score more negatively than the reference, but the margins — 0.03–0.65 kcal/mol — are inside AutoDock Vina's typical error of roughly 1–2 kcal/mol, so they do not establish stronger binding. Docking scores are estimates, not measurements, and only one identical protocol may be compared with itself.
08 / MemoryRx Score and ranking
A provisional ranking at 0.70 weight coverage
Weight allocation
CNS-permeability panel — Fail 10 of 10
Heuristics: MW under 450, TPSA under 90 Ų, HBD at most 3, LogP 1–4.
Observed: MW 456–496, TPSA 127.5–168.0, HBD 2–4, LogP 0.05–1.71.
TPSA is the decisive failure: every candidate exceeds 90 Ų by at least 37 Ų. CNS flag: Fail 10 / 10
What the BBB sub-score rests on
SwissADME BBB-permeant and P-gp fields were not populated, and no toxicity predictions were run — hepatotoxicity, mutagenicity, hERG and ProTox values are not yet assessed. The BBB sub-score therefore came from physicochemical descriptors alone.
TPSA per candidate against the 90 Ų CNS threshold
Naming: both identifiers, always
Final names were assigned by rank after scoring. Final MRX-01 is candidate MRX-06; final MRX-02 is candidate MRX-09; final MRX-03 is candidate MRX-08. The original MRX-01, MRX-02 and MRX-03 placed 9th, 8th and 4th. Both identifiers are shown together everywhere; a final name never appears alone.
Weight coverage is 0.70 for every row, so every score in this table carries that caveat. S_TREM2 and S_SA are empty in every row. CNS flag is Fail for all ten.
Table scrolls horizontally · first column stays fixed · headers sort
09 / Candidate passports
Ranks 1–3, each failing the CNS heuristic
Every card is headed with both identifiers and the CNS flag. All values are predictions at weight coverage 0.70.
10 / Methods, limitations, conclusions
Everything ran on free tools and a laptop
No licensed software, no cluster, no laboratory. Public databases, open-source libraries and a free notebook environment.
Target–disease association evidence used to assemble the 50-gene list.
Measured CDK5 bioactivity data for QSAR training.
Compound records and structure cross-checking.
Experimental structures 1UNG (CDK5) and 6XDS (TREM2).
Descriptors, fingerprints, structural alerts and derivative enumeration.
Random Forest models, cross-validation and metrics.
Gradient-boosted comparison models for classification and regression.
Docking into the CDK5 ATP pocket under one fixed protocol.
Physicochemical and ADME descriptors; several fields were not populated.
Free notebook environment in which the whole pipeline ran.
Limitations
- No wet-lab, in vivo, preclinical or clinical work was performed. Every value on this page is a prediction.
- The TREM2 arm did not complete, leaving 20 per cent of the scoring weight unassigned.
- Toxicity and synthetic accessibility were not run, leaving a further 10 per cent unassigned — total coverage 0.70.
- QSAR regression is weak (R² 0.45–0.47), so the model is used as a classifier only.
- Docking margins (0.03–0.65 kcal/mol) are within the method's typical error of roughly 1–2 kcal/mol.
- BBB models are qualitative, so the wording throughout is predicted permeability.
- Novelty is low by design: maximum Tanimoto similarity 0.59–0.85 to known CDK5 inhibitors.
- Stage-3 target scores are expert judgement and were revised during verification against Open Targets.
- These are candidate compounds, not drugs. Nothing here is medical advice.
Conclusions
01 Reproducibility
The pipeline is reproducible end to end on free tools, and the prioritisation and classification stages met their intended standard.
02 The answer is negative
For this series the answer is negative: affinity gains came from polar groups that broke permeability — a textbook potency-versus-permeability trade-off that was not anticipated at design.
03 Next steps
Complete or formally drop the TREM2 arm and redistribute its weight; run toxicity and synthetic-accessibility predictions; rebuild the library with TPSA as a hard design constraint rather than a post-hoc filter.
About the researcher
Zamira Bonu Ibragimova — pharmacist and pharmaceutical biotechnology researcher, Uzbekistan.
Trained in industrial pharmacy at the Tashkent Pharmaceutical Institute, with experience as a production technologist, a medical representative, and a first-category specialist in the inspection department of Uzbekistan's Agency for the Development of the Pharmaceutical Industry. Research interests are in pharmaceutical biotechnology and medicinal-plant in vitro culture, and she intends to pursue a PhD in pharmaceutical biotechnology.
Contact
banuibragimova8888@gmail.com