Every stage of a discovery programme is well served by mature software. Almost none of it was written to talk to the stage on either side.
A target arrives as a gene symbol in a spreadsheet. A chemist needs a structure, a pocket and a reason to believe the site can be bound. Rebuilding that context by hand is where the evidence trail breaks.
Alignment, differential expression, docking, dynamics and free energy were each written by and for specialists. Running them in sequence is a different skill from running any one of them well.
Assay outcomes seldom reach the models that proposed the molecule. Without that return path, the next round begins from the same literature as the last one did.
The same job identifier follows a programme from the first FASTQ file to the last free energy calculation. Each stage below names the platform that runs it and the modules involved.
Sequencing runs arrive as an accession list or as raw files. Quality control, trimming, alignment and quantification turn them into a count matrix, with every parameter and engine version written into the job record. None of this is novel science, and that is the point. It should be routine, and it should be reproducible by anyone who opens the record six months later.
Differential expression separates what changed from what did not, and functional enrichment says what those changes mean together. In our study of thirty nine lung transcriptomes (a manuscript, not yet published), 3,464 of 10,484 tested features were differentially expressed, and extracellular matrix organisation dominated both over representation and preranked GSEA. The optional modules added 38,909 differential splice junctions and 271 candidate fusion events.
A gene list is not a target. Candidates are ranked against the evidence behind them, mapped onto a structure, and checked for a pocket worth binding before anyone commits chemistry to them. This is the handoff that usually breaks, because biology hands over a symbol and chemistry needs a site. Here the ranked target carries its evidence into the next stage instead of being retyped into it.
Generative models propose structures the literature has not seen, a graph network triages them against the target, and pharmacokinetic and physicochemical filters remove what would fail later rather than at the end. In the published NDM-1 campaign a generative run produced 752,436 molecules from 455 known actives, and ten stages of filtering carried four of them forward.
The survivors are docked into the holo structure, carried through replicate molecular dynamics and scored by binding free energy. Redocking reproduced the crystallographic pose of the reference ligand at 0.840 angstrom, within the conventional acceptance threshold. What leaves the platform is a short list with the full parameter trail behind every decision that produced it.
Nominated molecules go to synthesis and assay. Those results return to the platform as training data, models are retrained inside their applicability domain, and the next round begins from measured evidence rather than from the same starting point as the last.
This is the part that a chain of separate tools cannot do. A result can only improve a model it can reach, and it can only reach a model that shares its data representation. Two halves of one system is not a marketing arrangement. It is what makes the return path possible at all.
The campaigns below were run on the platforms. Most are written up in the peer reviewed literature, and the pulmonary fibrosis study is a manuscript. They were chosen because they are difficult, not because they were likely to work.
NDM-1 hydrolyses almost the whole beta lactam class, including the carbapenems that form the last line of defence against multidrug resistant infections. No inhibitor has reached the clinic. The open, solvent exposed site makes it a hard test rather than a convenient one.
Published, 2026Thirty nine lung transcriptomes from patients and control donors, carried from raw reads to functional enrichment in one run. Matrix metalloproteinases, fibrillar collagens and basal keratins led the differentially expressed set.
ManuscriptTwo separate generative campaigns against kinase targets, each taken through machine learning triage, docking and replicate molecular dynamics, with alchemical free energy for the checkpoint kinase work.
PublishedA generative and physics based campaign against a mycobacterial toxin, reported with redocking, replicate dynamics and end point free energy for every nominated molecule.
PublishedFour candidates nominated against a fungal cell wall target, following the same cascade of generative design, machine learning screening, pharmacokinetic filtering, docking and dynamics.
PublishedThe cascade is not specific to these targets. If you have sequencing data and a disease, the same path applies, and the parameter trail behind every decision travels with the result.
OpenThe detail behind the headline numbers, one campaign for each platform.
752,436→4Molecules generated, then nominated
New Delhi metallo-β-lactamase-1 hydrolyses almost the whole β-lactam class and no inhibitor has reached the clinic. From 455 curated actives, a generative run produced 752,436 structures. A ten-stage cascade, run entirely within one session, carried four molecules into holo docking, triplicate 100 ns molecular dynamics and binding free energy.
Mangal P, Shinde O, Chikhale S, Bhowmick S, Islam MA. Journal of Molecular Graphics and Modelling 149 (2026) 109566.
3,464Differentially expressed features, of 10,484 tested
Thirty-nine lung transcriptomes from patients with idiopathic pulmonary fibrosis and control donors were taken from raw reads through to functional enrichment in a single guided run. Matrix metalloproteinases, fibrillar collagens and basal keratins led the differentially expressed set, and extracellular matrix organisation dominated both over-representation and gene set enrichment analysis.
The two optional modules ran in the same session, reporting differential splice junctions and candidate fusion events together with their reading frames and retained protein domains.
An independent re-run from the exported counts reproduced identical significance calls. Battula M, Mangal P, Shinde O, Chikhale S, Bhowmick S, Islam MA. Manuscript, not yet published.
Every number on this page is computational. Computational results are hypotheses, and they require experimental validation. The platforms exist to make that validation faster to reach and easier to trust, by carrying the evidence and the parameters forward with the result. They do not replace it, and nothing here should be read as a claim that they do.
Tell us what you have sequenced, or what you are trying to hit. A computational scientist will walk you through what the two platforms would do with it, where the evidence would be strong, and where it would not.
