SYNTHEX / robotic synthesis & validation
SYNTHEX is the physical execution layer of the SoftMining discovery loop. It miniaturises chemistry before scaling automation: tiny reagent volumes support rapid preparation, high parallelism and lower-cost robotic architectures while reducing reagent inventory, waste and exposure. Functionalised magnetic supports, AI-assisted mixture tracking and integrated qualification convert ADAPT-selected hypotheses into machine-readable experimental evidence.
01 / Why physical discovery stalls
The physical bottleneck is not only reaction speed. Miniaturisation creates the throughput advantage, but it also breaks assumptions inherited from flask-scale chemistry: transfers and classical purification become disproportionately expensive, while every loss of material matters. SYNTHEX treats execution, separation, mixture composition, qualification and biological readout as one controlled process.
Computational workflows can rank chemical space faster than conventional synthesis can test it. The consequence is not simply delay: model uncertainty survives because too few mechanistically distinct alternatives reach experiment.
EXPERIMENTAL REFERENCE: [S3] systematic physical libraries with measured structure–property relations.
At very small scale, conventional isolation and purification can consume more time and material than the reaction itself and may become physically impractical. SYNTHEX therefore designs capture, washing, separation and mixture-state tracking into the synthesis instead of assuming that flask-scale work-up can simply be shrunk.
PLATFORM-SPECIFIC ENGINEERING: [S2] and [S3] support the materials/miniaturised experimental lineage; the exact SYNTHEX purification and tracking architecture is proprietary.
Failed reaction, impure product, invalid assay, inactive compound and inconclusive measurement are different scientific states. Collapsing them into one “negative” destroys the information required for learning.
EXPERIMENTAL REFERENCE: [S1] design–synthesis–microbiological assay workflow; failure semantics remain a SYNTHEX-specific implementation requirement.
A result without route, conditions, controls, quality and provenance cannot update a computational hypothesis correctly. SYNTHEX returns the experiment as a structured object, not a spreadsheet endpoint.
EXPERIMENTAL REFERENCE: [S5] measured fabrication space linked to a predictive model.
Tiny reaction volumes reduce reagent inventory and waste, so more chemical alternatives can be tested within the same material budget.
Short, standardised robotic operations and dense parallel scheduling increase the number of preparations completed per cycle.
Small payloads and reagent volumes reduce mechanical and containment demands, enabling more economical robotic architectures and easier replication of parallel units.
Using very small quantities reduces chemical inventory, waste generation and the consequence of spills or operator exposure; it does not remove the need for standard safety controls.
02 / Execution contract
ADAPT defines the informative batch and its constraints. SYNTHEX converts that batch into physical operations, qualifies what was actually produced and returns every outcome with enough context to alter the next programme state.
Receive the candidate batch, required controls, assay objectives and decision context from ADAPT.
Translate route graphs, reagents, conditions and controls into available robotic and analytical operations.
Execute miniaturised parallel reaction sequences, magnetic capture, washing, transfer and release, while AI-assisted mixture tracking follows composition where classical purification becomes impractical.
Separate identity, purity and yield from biological response, assay validity and replicate uncertainty.
Send complete status codes, measurements, uncertainty and failures back to the persistent programme memory.
03 / Execution subsystem specifications
SYNTHEX is not generic laboratory automation. Its architecture is built around miniaturised chemistry: inexpensive parallel execution, recoverable material handling, AI-assisted mixture-state inference, qualification and data return remain coupled throughout the synthesis.
Programme-specific modules execute miniaturised liquid handling, timed reactions, transfers and instrument hand-offs with position-level timestamps, deviations and alarms. Small payloads and reagent volumes reduce hardware demands, supporting lower-cost robots that can be replicated for high parallelism.
ENGINEERING LINEAGE: [S5]. This supports instrumented experimental-process optimisation, not SYNTHEX robotic hardware performance.
Recoverable magnetic supports provide a physical handle for immobilisation, sequential reaction, washing, separation and release at miniaturised scale.
MATERIALS REFERENCES: [S2], [S3]. These document solid/material synthesis and characterisation; the supplied bibliography does not directly validate the magnetic-support subsystem.
SYNTHEX increases reaction density by shrinking the amount used per experiment. The advantage is simultaneous: faster preparation, higher parallelism, lower reagent cost and lower waste/exposure. Throughput is counted as qualified products—not crude reaction attempts—so speed cannot hide poor chemistry.
EXPERIMENTAL REFERENCE: [S1].
Miniaturisation makes repeated isolation and purification difficult because every transfer consumes material and time. A proprietary genetic-algorithm tool therefore tracks the evolving composition of the reaction mixture along the synthesis trajectory, allowing SYNTHEX to reason about a dynamic mixture instead of assuming a perfectly purified intermediate after every step.
PROPRIETARY SYNTHEX MODULE: this capability is supplied from the founders' internal technology stack; no peer-reviewed methods reference for the specific implementation was provided in the current bibliography.
A biological measurement is admitted into the learning loop only after chemical identity and purity are qualified and assay controls establish that the measurement itself is interpretable.
04 / The point where the physical loop becomes algorithmic
A low-purity sample cannot be learned as an inactive molecule. A qualified inactive molecule can. Preserving that distinction is the practical mechanism by which robotic execution changes the next computational decision.
Every result must retain candidate identity, executed route, conditions, material history, mixture-state history where applicable, QC, assay context, uncertainty and status. This prevents chemistry failures—or unresolved mixtures—from contaminating biological learning.
Minimum public return object
status ∈ {qualified-active, qualified-inactive, chemical-failure, analytical-failure, assay-invalid, inconclusive}.
u carries uncertainty and replicate state; it is not discarded before ADAPT receives the result.
05 / ADAPT interface
The two pages intentionally do not repeat each other. SYNTHEX exposes physical feasibility and returns qualified experimental states; target reasoning, pocket analysis, molecular similarity, screening and multi-objective selection remain on the ADAPT layer.
Candidate identities, decision objective, controls, permissible chemistry and requested measurements.
Route, material, handling, qualification, assay context, deviations and failure semantics.
Qualified measurements and structured failures update the persistent programme state.
06 / Protected engineering layer
Reaction classes, nanoparticle composition and loading, robot inventory, operating volumes, field gradients, recovery, cycle times, QC thresholds, assay stack and prospective performance are programme-specific technical disclosures—not marketing filler.
Subsystem roles, physical equations, data semantics and design–make–test logic.
Enough to understand the architectureHardware configuration, materials, reaction scope, process windows, validation thresholds and reproducibility data.
Discussed under technical disclosureThe strongest demonstration is a registered ADAPT-selected batch compared with a predefined baseline on information gained per cycle.
Physical evidence, not a slide metric07 / Experimental references
SYNTHEX is deliberately not documented with the ADAPT computational bibliography. The references below establish SoftMining's design–make–characterise, materials, synthesis, assay and experimental-ML lineage. Where the supplied portfolio does not directly validate a SYNTHEX subsystem—robotic throughput or magnetic-support performance—the page says so explicitly.
Synthesis and Antimicrobial Studies of New Antibacterial Azo-Compounds Active against S. aureus and L. monocytogenes. Molecules 22, 1372 (2017). DOI: 10.3390/molecules22081372.
Cationic Azobenzenes as Light-Responsive Crosslinkers for Alginate-Based Supramolecular Hydrogels. Polymers 16, 1233 (2024). DOI: 10.3390/polym16091233.
Tailored Thermoresponsive Polyurethane Hydrogels: Structure-Property Relationships for Injectable Biomedical Applications. Polymers 17, 2350 (2025). DOI: 10.3390/polym17172350.
Photomodulation of Vesicle Dynamics Using Fluorescent Photoswitchable Amphiphiles. Journal of Materials Chemistry B 14, 3093–3108 (2026). DOI: 10.1039/D5TB01894C.
Superhydrophobic Coatings and Artificial Neural Networks: Design, Development and Optimization. In Advances in Bionanomaterials II, pp. 32–40 (2020). DOI: 10.1007/978-3-030-47705-9_4.
A New Serotonin 2A Receptor Antagonist with Potential Benefits in Non-Alcoholic Fatty Liver Disease. Life Sciences 314, 121315 (2023). DOI: 10.1016/j.lfs.2022.121315.