About
Built by a physician, open to anyone who needs it.
Every promising molecule begins as a question: will it bind, will it last, will it be safe. This exists to answer that question before a laboratory has to — and to say plainly when computation cannot.
Dr Souriv Chakraborty, an experienced physician and computational pharmacologist, has developed this site for free public use by academicians, industry leaders, scientists and investors. Certain additional services require human intervention, and that is where pricing starts.
Who this is for
Academics and students who need to look a molecule up properly without an institutional licence. Scientists and development teams holding a compound or running a trial, with no dedicated computational pharmacology function in-house. Industry leaders weighing a candidate against what has already been reported about its class. And investors needing a technical read on an asset outside their in-house therapeutic expertise, ahead of a decision rather than after it.
Engagements are currently accepted from Europe, the United Kingdom, the United States, Canada, Japan and Australia. The practice is based in Kolkata, India; contracting and invoicing are handled directly, and NDAs are standard on new chemical entity work.
Why the platform is free
Gating a database lookup behind a signup wall would be a worse business than being useful. The value sits in the interpretation, not in the query, and the quickest way to demonstrate that is to let people run the query themselves. What is charged for is the part a machine cannot do: reading the combined evidence, deciding what it means for a specific decision, and putting a name to the conclusion.
How the work is actually done
If a method cannot be stated plainly enough to be checked, it is not a method worth selling. Everything below is what runs, named at version where the version matters.
Languages and runtime
The platform is Python throughout: a Gradio interface over a SQLite catalogue, containerised and served on Google Cloud Run, with the static site on nginx alongside it. Heavier work — model training, enumeration, docking campaigns, molecular dynamics — runs as documented notebook pipelines on Google Colab CPU and GPU instances, with every run writing its model artefact, its metrics file and its predictions to disk so a result can be regenerated rather than retold.
Databases and sources
| Layer | Source | Use |
|---|---|---|
| Compound universe | ChEMBL-derived | About 17,721 curated entities — small molecules, biologics and gene therapies — with canonical SMILES, modality and highest clinical phase |
| Adverse effects | PubMed, Scopus, US label, spontaneous reports | Drug-induced effects indexed in MEDLINE, cited by PMID, first; label text second; the five most reported reactions per molecule last, as a signal only |
| Indications | Label and trial | Approved labelled indications kept separate from off-label and pipeline use |
| Bioactivity | ChEMBL target assays | Training data for liability models, filtered to exact-relation values and de-duplicated on canonical structure |
| Literature | NCBI PubMed | Live retrieval through E-utilities at query time; citations given as resolvable PMIDs |
| Structures | RCSB PDB | Receptor structures fetched by accession code, never modelled silently |
| Descriptors | RDKit | Physicochemical properties, scaffolds and fingerprints computed locally |
Cheminformatics
Physicochemical descriptors and Lipinski parameters are computed with RDKit from canonical SMILES, alongside Murcko scaffolds for series work. Structural neighbourhood is assessed by Tanimoto similarity over 2048-bit Morgan fingerprints at radius 2, with the threshold reported alongside every result — a similarity figure without its radius and threshold is not interpretable.
Machine learning
Liability models are built per target rather than bought off the shelf. A typical build pulls the bioactivity set for a named ChEMBL target, keeps only exact-relation measurements, de-duplicates on canonical structure, and trains a random-forest regressor on Morgan fingerprints, reporting five-fold cross-validated R-squared, RMSE and MAE in log units so the error bar on any single prediction is visible.
Two disciplines matter more than the algorithm. Applicability domain is checked explicitly: maximum Tanimoto similarity to the training set is reported per molecule, and predictions outside the cut-off are labelled as extrapolation rather than quietly averaged in. Spread is compared against model error: when a series of analogues varies by less than the cross-validation RMSE, that is reported as the model being unable to rank them, not as a winner. Tree ensembles regress toward the training mean on unfamiliar chemistry, and a comfortable number can simply be the model saying it does not know.
Alongside the purpose-built models, published ADMET predictors cover the standard panel — cardiac channel liability, blood–brain barrier penetration, efflux substrate status, mutagenicity, hepatotoxicity and clinical toxicity — used as corroboration, never as the single basis for a claim.
Design and enumeration
Analogue work enumerates a combinatorial neighbourhood around the feature that has to be retained, then filters on the properties that drive the liability being designed out — basicity, lipophilicity, molecular weight, and the number of edits from the parent. Survivors are ranked by multi-objective Pareto selection across predicted liability, CNS exposure score, similarity to the parent, edit count, synthetic accessibility and structural alerts. The output is a prioritised list of synthesis hypotheses; it is never presented as a set of de-risked molecules.
Docking and simulation
Docking uses AutoDock Vina 1.2.7, with GNINA convolutional rescoring where a deep-learned pose score adds something. Parameters below are the defaults; any deviation is stated in the report.
| Parameter | Default |
|---|---|
| Receptor preparation | Single chain extracted; waters, cryoprotectants, lipids and crystallisation fusion domains stripped; catalytic cofactors such as HEM retained |
| Box placement | Centroid of a co-crystallised reference ligand in the same structure |
| Box size | 22 × 22 × 22 Å |
| Exhaustiveness | 16 |
| Modes returned | 9 |
| Ligand preparation | SMILES → Open Babel 3.1.1, --gen3d --best, protonated at pH 7.4 |
| Controls | Property-matched decoys, with retrieval reported as an ROC AUC for the target in question |
Where the question needs more than a static pose, molecular dynamics and MM/GBSA give a relative energetic ranking with per-residue decomposition. Both are reported as comparative measures within a single target, with the single-trajectory and missing-entropy caveats stated.
On scores. A docking affinity is a ranking heuristic, not a measured binding constant, and a decoy control that returns an ROC AUC near chance means the score is not separating actives on that target. Reports give poses, interactions and comparisons; they do not convert a score into a dissociation constant.
Language models, and what they are allowed to do
The platform's clinical agent is retrieval-grounded. A PubMed search runs first through NCBI E-utilities; the retrieved records, with their PMIDs, are handed to the model as the evidence it must cite, and the answer comes back in a fixed structure — mechanism, clinical evidence, safety considerations. Gemini 3.6 Flash is the primary model, with Llama 3.1 8B on Groq as a fallback when it is unavailable.
The boundary is deliberate. A language model is used to synthesise retrieved literature into readable prose, never as the source of a pharmacological fact, and never as the last step before something is delivered. Anything that leaves as a deliverable has been read by a physician first.
The toolchain, named
Not exhaustive, and versions move. This is what the work is actually done with.
| Layer | What is used |
|---|---|
| Languages | Python for everything substantive; R and Julia occasionally for analysis; shell for pipeline orchestration |
| Chemistry | RDKit (fingerprints, Lipinski and property filters, scaffolds, analogue enumeration), Open Babel, PyMOL |
| Docking | AutoDock Vina 1.2.7; GNINA for convolutional rescoring on GPU; property-matched decoy controls |
| Simulation | OpenMM on CUDA for molecular dynamics; tau-RAMD for unbinding kinetics; MM/GBSA and MM/PBSA for relative binding energy |
| Quantum and electronic structure | xtb for semiempirical Fukui descriptors; DFT at PBE0/def2-SVP where reactivity needs it |
| Machine learning | scikit-learn with joblib persistence for QSAR (random forests over Morgan fingerprints, five-fold cross-validated); admet-ai for ADMET and toxicity panels; AlphaFold for structure modelling where no experimental structure exists |
| Databases | ChEMBL for bioactivity and the compound universe; RCSB PDB for structures; UniProt for sequence and residue numbering; FDA FAERS for adverse-event reporting; NCBI Entrez for literature |
| Statistics and sequence work | MedCalc for clinical statistics; MEGA for sequence and phylogenetic analysis |
| Compute | Google Colab CPU and GPU runtimes for the heavy work, with every run writing its model artefact, metrics and predictions to storage |
| This platform | Gradio and SQLite in Docker on Google Cloud Run, with nginx serving the static site; Gemini 3.6 Flash and Llama 3.1 8B behind the retrieval-grounded agent |
Language-model output is treated as drafting, not evidence: anything it produces is checked against the primary source before it reaches a deliverable, and a deliverable is read by a physician before it is sent.
Stated limitations
- Static docking cannot represent receptor flexibility, explicit solvent or induced fit.
- Loss-of-function variants that truncate or fail to express a protein cannot be demonstrated by docking the wild-type pocket.
- A model asked to predict outside its applicability domain returns the training mean in a confident voice; this is reported, not smoothed over.
- Estimated physicochemical values such as topological pKa are heuristics, and any compound taken forward should have them computed properly or measured.
- Spontaneous adverse-event reporting carries reporting bias; raw counts are not incidence rates and are never presented as such.
- Fingerprint similarity is a structural measure, not a measure of shared pharmacology.
- Database contents reflect a snapshot; the version and retrieval date appear in every report.