What we are starting with
A 200-compound SMILES file is a common starting point for a first computational screening pass. It might come from a small virtual library enumerated around a fragment hit, a set of commercially available analogues around a known scaffold, or a mix of chemist-nominated structures from several sources. The format is a flat text file or CSV, one SMILES string per row, optionally with a compound identifier column.
The question this walkthrough answers is: what happens from the moment you submit that file to the moment you have a ranked shortlist you can use to decide which compounds to request for synthesis? The full cycle, including model run time, takes around 90 seconds for a 200-compound batch. But most of the real work happens in the ten minutes before and after that 90 seconds, when you prepare the input and interpret the output.
Preparing the input file
The most common preparation issues are worth addressing upfront, because a malformed input file will produce errors or silent failures that are harder to diagnose after the run.
Check that your SMILES strings are valid before submitting. The fastest way is to round-trip through RDKit: parse each SMILES and re-canonicalise it. If any string fails to parse, it will either be dropped from the run or return null scores. A SMILES file assembled from multiple sources often contains encoding artifacts, especially if some strings came from a sketcher export and others from a database query.
Assign unique compound identifiers if your file does not already have them. The output ranking is returned with the same identifiers you submitted. If you submit with no identifier column, the ranked list comes back with index positions only, which makes it harder to trace back to your original set. A short alphanumeric ID, for example LIB-001 through LIB-200, is sufficient.
Decide whether to include explicit hydrogens. The scoring engine handles both representations, but canonicalising to implicit hydrogens before submission is cleaner and reduces file size. If your SMILES were exported from a 3D structure tool with explicit Hs, strip them first.
Submitting the batch and what happens during the run
Upload the CSV through the platform interface or POST it to the API endpoint if you are working programmatically. The request accepts a file upload or an inline JSON payload with a smiles_list array. For 200 compounds, either works fine; the file upload is simpler for one-off runs.
During the 90-second processing window, the platform is running three parallel scoring pipelines. The binding affinity model, a graph neural network operating on the molecular graph, generates a predicted binding score in kcal/mol and a confidence interval. The ADMET pipeline runs the tox models, returning per-property flags and confidence levels for the subset of properties most relevant to early hit evaluation. The retrosynthesis engine analyses each SMILES for synthesizability, traversing the retrosynthetic tree to a specified search depth and returning a synthesizability score between 0 and 1.
The three scores are computed independently. There is no cascade where a low binding score suppresses tox analysis, or vice versa. Every compound in the input receives scores on all three axes, or an explicit null if the model cannot process a particular structure.
Reading the ranked output
The output CSV has one row per input compound, sorted by the composite rank. The default ranking weights binding score most heavily, with tox flags applied as hard filters at configurable thresholds: compounds with high-confidence flags for specific liabilities are moved to the bottom of the ranked list rather than being silently removed.
The columns to check first are: Rank, Binding (kcal/mol), Binding CI lower/upper, Synth Score, and Tox Flag summary. The confidence interval on the binding score is the first signal about how much to trust that row's rank. A binding prediction of 7.8 kcal/mol with a 95% confidence interval of 7.1 to 8.5 is a tighter prediction than one with a CI of 5.9 to 9.7, even though both report the same central estimate. Compounds with wide confidence intervals should be treated as uncertain: the model is telling you it lacks strong signal for that structure.
The Synth Score column reports the synthesizability value and, in the annotated output, a brief summary of what drove the score. A score of 0.82 might note "commercial building blocks identified for both synthetic branches." A score of 0.38 might note "chiral centre density, no direct route found under 6 steps." These annotations are brief but useful for understanding whether the synthesizability assessment is driven by genuine route difficulty or by a gap in the retrosynthetic training data.
Identifying the first synthesis batch
For a 200-compound input, a first synthesis batch of 8 to 15 compounds is typical. The selection logic moves through three passes.
First pass: take the top 30 by composite rank and examine the binding confidence intervals. Drop any compound whose CI lower bound is below 6.5 kcal/mol (roughly 15 micromolar), regardless of the central estimate. You are left with compounds where the model has reasonable confidence that the binding is meaningful.
Second pass: from the remaining set, apply your synthesizability threshold. If you have a capable synthetic chemist and some tolerance for difficulty, a threshold of 0.55 is defensible. If you need compounds that can be purchased as building blocks or made in two to three steps, set the threshold at 0.70. Compounds below the threshold move to a "design target" list, not the synthesis queue.
Third pass: review the tox flag column. Compounds with any high-confidence flag for hERG inhibition, genotoxicity, or reactive metabolite formation are deprioritised. Medium-confidence flags are documented but do not automatically exclude a compound from the first batch. You want to see those flags measured experimentally, not assumed to be true positives.
What you are left with after three passes is your first synthesis batch. For a 200-compound input, this is typically 8 to 15 compounds. If it comes out fewer than 5, consider loosening the synthesizability threshold or expanding the initial binding confidence criterion. If it comes out more than 20, your input set was well-curated upfront, which is a good sign.
What to record before you send the synthesis request
Before the first batch goes to synthesis, document three things for each compound being queued: the binding rank and CI, the synthesizability score with any flag notes, and the decision logic if the compound is borderline (for example, "medium hERG flag acknowledged, proceeding to measure"). This documentation matters for program history, particularly when results come back and you are calibrating the scoring model against your first round of data.
A compound that ranked 2nd in the Alkira output and comes back with IC50 of 340 nM is a useful calibration point: the prediction was in the right region. A compound that ranked 7th and comes back at 12 nM is equally useful: the model may have underestimated binding for this scaffold class, or the confidence interval was wide enough to accommodate the result. Either way, the comparison between predicted rank and measured activity is the input to improving scoring accuracy for your next round.
The 90-second run produces a starting point, not a final answer. The compounds you choose to synthesize from that starting point, and the results you get, are what turn a ranked hit list into real SAR.