Three scoring axes. One ranked shortlist.
Alkira evaluates each compound on binding affinity, toxicity risk, and synthesizability simultaneously, returning a ranked output your team can act on without assembling scores from separate tools.
Scoring axes
What each score measures and how to use it
Each axis addresses a distinct failure mode. Using them together surfaces the candidates most likely to progress, not just the ones with the highest binding number.
Binding affinity score
Structure-based docking predicts free energy of binding (kcal/mol) between each candidate and your specified target pocket. Higher magnitude numbers indicate stronger predicted binding. The score does not account for off-target interactions, so use it alongside selectivity judgment. Values above 8 kcal/mol in absolute terms are generally worth prioritizing; values below 6 usually indicate weak binding geometry.
What to do: sort your series by binding score descending, then check tox and synth for the top 30 percent.
Toxicity risk prediction
A multi-model ensemble trained on public ADMET datasets evaluates each compound for hERG channel inhibition (cardiac risk), reactive functional group alerts, CYP450 metabolic liabilities, and broad cytotoxicity flags. The output is a three-level classification: Low, Medium, or High risk. Medium-risk compounds are not automatically disqualified, but they warrant closer attention before committing bench time.
What to do: filter High-risk compounds out of your priority queue unless the target class has known tolerance for the flagged liability.
Synthesizability score
Retrosynthesis-based scoring estimates the feasibility of preparing each compound from commercially available building blocks. The score (0-100%) reflects the model's estimate of route availability, step count, and reagent accessibility. A score above 70% typically corresponds to a 3-6 step synthesis from commercial starting materials. Below 40% suggests significant synthetic complexity that should be weighed against the binding advantage before committing to the route.
What to do: a high synth score alone does not confirm a compound is worth making, but a very low score is a meaningful reason to deprioritize it in a large series.
Input to output
From a SMILES file to a ranked shortlist in three steps
Upload or paste your SMILES
Submit a CSV with a SMILES column, or paste SMILES strings directly into the input panel. Specify your target pocket by PDB ID or upload a binding site definition.
CC1=CC(=CC=C1)NC(=O)C2=CC=C(Cl)C=C2
COC1=CC=C(NC(=O)C2=CN=CC=C2)C=C1
C1CC2=CC=CC=C2CC1NC(=O)C3=CC=CN=C3
FC1=CC=C(NC(=O)C2=NC=CC=C2)C=C1
CC(C)CC1=CC=C(C=C1)C(C)C(=O)OAlkira runs all three models
Binding, tox, and synthesizability models run in parallel. Most runs complete within 90 seconds for series up to 500 compounds. Larger batches on the Team plan take proportionally longer but remain in the same session.
Download your ranked shortlist
The output is a ranked CSV with all three scores, confidence bands, and a composite rank. Filter, sort, or export directly. Your run history is saved to your account for comparison across series.
| Compound | Binding | Tox | Synth | Conf | Rank |
|---|---|---|---|---|---|
| KB-4821 | 8.2 | Low | 85% | High | 1 |
| AL-0934 | 7.9 | Low | 78% | High | 2 |
| MB-2217 | 8.5 | Medium | 54% | Medium | 3 |
| CK-1055 | 7.1 | High | 71% | High | filtered |
Under the hood
How the scoring models work
Binding affinity uses a structure-based docking pipeline extended with a graph neural network re-scoring step. The GNN was trained on a curated set of experimentally validated binding data from public crystallographic databases. It corrects for known docking artefacts in flexible binding sites.
Toxicity prediction uses an ensemble of independent classifiers trained on ChEMBL ADMET data, a curated hERG dataset, and literature-sourced reactive group alerts. Ensemble voting reduces single-model overconfidence and provides the three-level confidence band.
Synthesizability is estimated using a retrosynthesis graph search over a commercial building block catalog, combined with a reaction score model that weights step count and reaction yield history. The percent score reflects the estimated probability of finding a viable route under realistic laboratory conditions.
All models operate on SMILES input. No 3D structure preparation is required. For binding scoring, Alkira handles pocket preparation internally from the PDB ID or uploaded structure you provide.
Alkira is a computational research tool for early-stage compound prioritization. Scores are predictive model outputs intended to guide synthesis decisions, not clinical recommendations. The platform does not make regulatory submissions or clinical efficacy claims.
Access options
Three ways to submit a compound series
CSV or SMILES file upload
Upload a CSV file with a SMILES column. Alkira auto-detects the column header. Supported: .csv, .smi, .txt. Maximum 500 compounds on Researcher plan, 2000 on Team.
Paste inline
Paste SMILES strings directly into the submission panel, one per line. Useful for quick ad-hoc scoring without creating a file. Supports up to 50 compounds per paste.
REST API (Team plan)
Automate submission from your existing workflow. Send a POST with your SMILES array and receive a JSON response with ranked scores. API key issued from your account dashboard.
# Submit compounds via REST API
curl -X POST https://api.alkirabio.com/v1/score \
-H "Authorization: Bearer $ALKIRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"target_pdb": "3ERK",
"smiles": [
"CC1=CC(=CC=C1)NC(=O)C2=CC=C(Cl)C=C2",
"COC1=CC=C(NC(=O)C2=CN=CC=C2)C=C1"
]
}'
# Response
{
"run_id": "run_20260602_093841",
"compounds": [
{ "smiles": "CC1=CC...", "binding": 8.2, "tox": "Low", "synth": 85, "rank": 1 },
{ "smiles": "COC1=CC...", "binding": 7.1, "tox": "Low", "synth": 72, "rank": 2 }
]
}Ready to score your current compound series?
Submit your SMILES file and get a ranked shortlist in 90 seconds. No installation required. One free run to start.