In Silico Drug Discovery Today: How Structure Prediction, Virtual Screening, AI, and Experiments Work Together

Researchers now have access to accurate protein-structure predictions, ultra-large purchasable libraries, molecular-docking engines, generative models, learned property predictors, and increasingly mature free-energy methods. That abundance creates a new problem: a workflow can look technically advanced while its outputs answer different questions, rest on incompatible assumptions, or lose the provenance needed to reproduce a decision.

This guide maps the present-day workflow from biological question to experimental feedback. It is intended for individual computational researchers choosing tools and for laboratories evaluating whether a workflow can be reviewed, repeated, and handed from one specialist to another.

The first computational input should not be a protein file. It should be a decision question. Are you looking for ligands that occupy a known orthosteric site, stabilize a particular conformation, disrupt an interaction, explain a resistance mutation, or create a chemically diverse set for an assay?

That question determines the relevant target state, controls, library composition, computational methods, and experimental endpoint. Without it, a workflow can optimize an attractive score against a biologically irrelevant state. Record the target rationale, intended mechanism, assay context, known liabilities, and the criterion that would justify the next experiment before screening begins.

AlphaFold 2 changed access to protein-structure hypotheses by achieving high accuracy in the CASP14 assessment and reporting confidence estimates that help users identify uncertain regions [1]. AlphaFold 3 broadened structure prediction to complexes that can include proteins, nucleic acids, small molecules, ions, and modified residues [2]. These advances reduce a historic barrier: many projects no longer have to begin with no structural model at all.

Structure availability is not the same as structure suitability. A docking receptor still needs a relevant biological state, credible binding-site geometry, appropriate cofactors or waters, repaired chemistry, and an explicit preparation policy. Confidence can vary locally even when the global fold looks convincing. A predicted complex is also not interchangeable with an AutoDock Vina-style prospective screen of a user-defined ligand library.

Use the AlphaFold and molecular-docking guide when deciding how a predicted receptor can enter a small-molecule workflow without overstating what structure prediction has established.

A database row or SMILES string is not yet a docking-ready scientific object. Salt handling, protonation, tautomerism, stereochemistry, valence, explicit hydrogens, three-dimensional geometry, atom typing, and charge assignment can change the object evaluated by the next method.

Keep the source identity separate from every prepared state and retain failures rather than allowing unsupported molecules to disappear from the denominator. The protein and ligand preparation guide explains the scientific choices; the ligand preflight guide shows how to detect structural problems before they become ambiguous docking failures.

Structure-based virtual screening asks whether a search method can generate plausible poses and rank compounds usefully enough to reduce a much larger library to a smaller set for review or testing. Prospective studies have shown that docking can operate at enormous scale and recover previously unrecognized chemotypes, as demonstrated in screens of more than 170 million make-on-demand compounds [3] and in the open-source VirtualFlow platform for ultra-large screens [4].

Scale is not proof of accuracy. A defensible screen preserves the receptor, ligand states, search box, scoring function, exhaustiveness, random seed policy, software version, failures, ranked poses, and review decisions. It validates the protocol with target-relevant evidence before exposing the production library to it.

Machine learning can prioritize compounds, estimate properties, generate molecular proposals, predict poses, and identify patterns across data that are difficult to encode manually. Its role is strongest when the training data, representation, endpoint, and deployment domain match the decision being made.

There are persuasive examples of computational prediction connected to experiments. Stokes and colleagues trained a deep learning model on antibacterial measurements, applied it to other chemical collections, and experimentally validated prioritized compounds [8]. More recently, Scalia and colleagues combined a primary screen of roughly two million compounds with deep learning and prospective testing after screening more than 1.4 billion synthesizable compounds in silico [7]. These are integrated discovery studies, not evidence that a generic model works for every target, assay, or chemical space.

Molecular dynamics and free-energy calculations address questions that a docking score was not designed to answer. They can examine conformational stability, interactions over time, or energetic differences among a narrower set of compounds. In an industry-scale assessment, Schindler and colleagues evaluated relative binding free-energy calculations across active drug discovery projects and documented both prospective usefulness and practical constraints [5].

These methods are not automatic confirmation layers. Their reliability depends on the structural model, chemical mapping, force field, sampling, protocol, diagnostics, and the type of chemical change. They are usually most defensible when the question has narrowed enough to justify their greater setup and compute cost.

A compound can have an attractive predicted pose and still be a poor experimental candidate. Solubility, permeability, chemical stability, aggregation risk, reactivity, metabolism, selectivity, synthesis, and assay compatibility can change the value of a shortlist.

Use predicted properties to expose trade-offs and prioritize measurements, not to collapse every objective into an unexplained composite score. Preserve the individual values, prediction methods, confidence or domain indicators, decision thresholds, and the reason a candidate advanced. The best candidate is the one that fits the current decision, not necessarily the molecule with the most extreme value in any single column.

Experiments do more than approve the final shortlist. They expose assay interference, target dependence, chemical-state errors, inactive regions of chemical space, and model blind spots. Those results should return to the computational workflow as new evidence rather than being reduced to a success/failure headline.

Grisoni and colleagues demonstrated this loop by coupling generative molecular design with automated synthesis and in vitro testing [9]. The antibacterial studies above likewise connected computational prioritization to prospective measurements [7,8]. The transferable principle is not a particular model: design the computational and experimental records so that each tested compound can improve the next decision cycle.

1. Frame the decision. Define the biological question, target state, assay, controls, and advancement rule.
2. Choose structural evidence. Select an experimental structure or predicted model and document local confidence and state relevance.
3. Curate the chemical set. Preserve source identities, enumerate intended states, and retain preparation failures.
4. Validate before screening. Test whether the docking protocol is fit for the intended prioritization task.
5. Generate and review hypotheses. Dock under a frozen protocol, inspect poses, and keep score, geometry, and provenance together.
6. Add orthogonal evidence selectively. Use learned property models, similarity, physics-based refinement, or other methods when each answers a defined unresolved question.
7. Select a diverse, testable set. Balance predicted evidence, chemical diversity, liabilities, synthesis or availability, and assay capacity.
8. Measure and learn. Return experimental outcomes, including failures and controls, to the next cycle.

This sequence is a framework, not a mandatory software stack. Some projects begin with phenotypic data, fragment hits, known ligands, or an experimental structure. The invariant requirement is that each method has an explicit role and a traceable handoff.

An individual researcher may prioritize rapid setup, local control, transparent files, and a manageable one-time cost. A laboratory or company may additionally need standardized protocols, reviewable histories, deployment rules, purchasing clarity, continuity, and reliable handoffs. The scientific core should remain the same; the operational proof required by each buyer is different.

MolNexus is focused on one part of this system: reproducible protein-small-molecule docking on a local Windows desktop. Its current product specification connects receptor and ligand preparation review, interaction-box definition, AutoDock Vina execution with Vina or Vinardo scoring, pose inspection, local job history, and exports.

That scope matters. MolNexus should be evaluated as a structured docking workflow, not as target discovery, a generative chemistry system, a free-energy platform, or experimental validation. It can make the docking layer easier to operate and audit; it does not turn a docking result into a biological conclusion. The commercial release is coming soon, and purchase and download are not yet open.

The present era is defined less by one winning algorithm than by better connections among biology, structures, chemical data, docking, learned models, physical simulations, and experiments. Structure prediction has made credible starting models available to many more projects. Virtual screening can explore chemical spaces at remarkable scale. AI can learn useful patterns and propose candidates. Physics-based methods can refine narrower comparisons. None removes the need to state the question, preserve provenance, test applicability, and measure the result.

A strong in silico workflow makes each handoff explicit. That is what turns a collection of modern tools into a reproducible discovery process.