Technical Whitepaper

Unified Protein Engineering Platform™

From Gene Synthesis to Functional Protein in One Continuous Workflow

An in-silico-first protein engineering service that unifies Design, Predict, Validate and Synthesize into a single computational pipeline

Table of Contents

Note: page numbers refresh automatically when you select the table of contents in Word and press F9 (update field).

1. Overview - The Problem This Platform Solves

1.1 The Structural Bottleneck Facing the Industry

In protein therapeutics and industrial enzyme development, the dominant cost is not the cost of synthesis but the cost of failed synthesis. Gene synthesis prices have fallen dramatically over the past fifteen years, and physically making a gene is no longer the bottleneck. The bottleneck has moved entirely to deciding what to synthesise. In a typical enzyme improvement project, a team clones, expresses and purifies several hundred variants, then discovers six months later that only a single-digit number show improved activity. What is consumed in this process is not reagent budget but researcher time and project opportunity cost.

The root of the problem is the non-linearity of the sequence-structure-function mapping. A single amino acid substitution 8 Å away from the active site can change the catalytic turnover number (k_cat) by an order of magnitude if that residue sits on a hinge governing the flexibility of the substrate entry channel. Conversely, substitutions immediately adjacent to the active site frequently have no effect at all. Because of this counter-intuitiveness, rational design has often lost to directed evolution; yet directed evolution faces a search space that explodes as 20^N, so the region that can be covered experimentally is an infinitesimal fraction of the whole.

1.2 The Platform's Approach

This platform does not claim to replace experiments. Instead it sets a far more defensible goal: to computationally redistribute the prior probability of the candidates entering the laboratory. Concretely, the aim is to convert a candidate pool with a 1-3% success rate under random selection into an enriched pool with a 25-45% success rate after computational filtering, which means achieving the same expected outcome with an order of magnitude less wet-lab work.

Three principles underpin the design. First, no single predictive model is trusted; instead independent evidence from three families - physics-based, evolution-based and learning-based - is fused. Second, every prediction carries an uncertainty that is propagated forward, and axes with large uncertainty are down-weighted arithmetically rather than by arbitrary if/else thresholds (inverse-variance weighted fusion). Third, a funnel architecture places cheap filters early and expensive filters late, securing throughput and accuracy simultaneously.

1.3 Quantitative Summary

Table 1-1. Platform performance summary. Evidence grades are elaborated in Chapters 5 and 6.

2. Service Scope: Continuity from Gene Synthesis to Protein Synthesis

2.1 Why Continuity Matters Technically

Many organisations run protein design and gene synthesis ordering as separate processes. A design team fixes an amino acid sequence, a separate tool codon-optimises it, and the result is ordered from a synthesis vendor. This separation looks reasonable but produces three distinct losses of information.

First, two designs judged equivalent at the amino acid level are not equivalent at all at the DNA level. Certain sequences fail synthesis outright, or become far more expensive, because of skewed GC content, repeats or hairpin structures. If this is unknown at design time, the situation 'our best candidate cannot be synthesised' surfaces far too late. Second, codon usage bias does not merely govern expression level; through the translation rate profile it shapes co-translational folding. Ribosome pausing caused by rare codons plays a functional role in buying folding time at domain boundaries, and 'full optimisation' that ignores this can actually increase inclusion body formation. Third, the choice of expression host determines post-translational modification (PTM), disulfide bond forming capacity and the chaperone environment, so it must already be considered at the protein design stage.

This platform treats all three layers - nucleotide, translation dynamics and folded protein - within a single objective function. Protein design scores and synthesisability scores compete on the same candidate ranking table, so an optimal design that cannot be synthesised never reaches the top in the first place.

2.2 Service Boundary Definition

Table 2-1. Service boundary in RACI terms. The platform fully owns stages 2-5 and provides specifications and criteria for stages 6-8.

2.3 Closed-Loop Operation

The real value of this service arises from iteration, not from one-shot design. As soon as round-one wet-lab data returns, it is used to calibrate the predictive models. Specifically, the prediction-measurement correlation of each evaluation axis is re-estimated, and axes with weak correlation are automatically down-weighted according to the inverse-variance principle. This is superior to adding ad hoc rules because axis reliability is determined by data rather than by human bias. Typically, predictive accuracy rises significantly from round two onward thanks to target-specific calibration, and the rate of reaching the target specification improves most sharply in round three.

3. Unified Platform Architecture - A Single 17-Stage Pipeline

The platform supports two entry modes: an improvement mode that engineers an existing protein, and a de novo mode that creates a scaffold not found in nature. The two modes differ only in the first three stages and share every subsequent evaluation and validation stage. From the customer's perspective this is one service; only the entry point is chosen automatically according to the nature of the target.

3.1 Stage Composition

Table 3-1. Unified pipeline stages. Candidate counts are measured values from a representative enzyme improvement project and vary by project. [internal]

Figure 3-1. The candidate funnel. Each stage costs more per candidate than the one before, so low-cost filters are placed early to ensure expensive computation reaches only a small number of candidates. Figures are illustrative of a representative enzyme improvement project. [internal]

3.2 The Information-Theoretic Basis of the Funnel

The most common misconception about funnel architectures is the assumption that increasing the initial candidate count improves the final outcome. This holds only when each filter stage has sufficient discriminative power (selection correlation rho). If a filter's rho is low, raising the initial pool from 1,000 to 10,000 barely changes the quality of the final selected set, because the filter is effectively equivalent to random sampling. Internally we call this the 'constant-count trap', and it is a phenomenon we confirmed during actual development.

The platform therefore prioritises filter discriminative power over initial scale. For example, a structural prescreen is inserted before sequence design (S1), filtering scaffolds first by geometric measures such as the RMS of Cα distances within the catalytic triad, the radius of gyration and the residue contact count. Only after discriminative power is secured is the initial candidate count expanded.

3.3 Uncertainty Propagation

Every computational axis produces not only a point estimate but also a standard error. Structure prediction, for instance, is repeated across multiple random seeds, and the inter-seed variance is used as a confidence measure. A single-seed prediction may look good by chance, but a candidate with large seed-to-seed deviation is classified as 'unstably predicted' and automatically penalised. This matters greatly in practice, because mistaking a low-confidence candidate for a high-scoring one and sending it into the laboratory is the most expensive error available.

Fusion across axes uses inverse-variance weighting. High-variance axes automatically contribute less, so if an axis carries no signal for a given project its influence disappears without any manual intervention, while reliable axes naturally dominate the ranking. A materiality gate is also applied: if an axis fails to produce a significant difference between candidates, it is excluded from ranking altogether. This prevents spurious signals - such as a score that merely decreases in proportion to mutation count - from contaminating the ranking.

4. Internal Algorithm Module Reference (GC-Series)

The platform's computational capability comprises 30 internal modules (4 design generation, 7 structural and physical, 5 evolution and learning, 6 risk and developability, 2 target modulation, 6 translational and modality-specific). Each module is defined by a functional contract, and the implementing backend may be replaced as performance improves. It is a design principle that the customer interface and deliverable specifications remain unchanged regardless of backend replacement. Not every module is active in every project: only the modules required by the nature of the target and modality are switched on automatically (for example, the target modulation family only for genome editing projects, and the translational and modality-specific family only when the relevant modality is chosen).

4.1 Design Generation Family

4.2 Structural and Physical Prediction Family

4.3 Evolution- and Learning-Based Family

4.4 Risk and Developability Family

4.5 Target Modulation Family (Nucleic Acid Targeting Systems)

The GC-GUIDE and GC-OFFTARGET modules are activated only for projects involving genome editing enzymes. In ordinary enzyme improvement or antigen design projects they remain inactive and consume no compute budget.

4.6 Translational and Modality-Specific Family

Where the previous five families are general-purpose engines reused across every protein project, this family consists of translational-stage modules that activate only when a specific modality is chosen. They do not generate new candidates; instead they evaluate already-generated, already-validated candidates against the developability, pharmacokinetic and regulatory axes unique to that modality. They are therefore mostly analysis- and assessment-only, distinct from the upper families that own a generative pipeline.

5. Biological Meaning of Each Stage - What the Numbers Actually Mean

This is the most important chapter of the whitepaper. A computational score has no meaning in itself; it can be used for decisions only when it is clear which biological phenomenon it proxies. Each section follows the order: what is computed, what it means biologically, and how it can be wrong.

5.1 Scaffold Generation - Does Function Precede Form?

The starting point of de novo design is the question: to catalyse this chemical reaction, which atoms must sit where in space? Taking a serine hydrolase as an example, the Ser-His-Asp catalytic triad demands a specific geometric arrangement. The distance between the Oγ of Ser and the Nε2 of His must fall in the 2.6-3.2 Å range where a hydrogen bond forms; the Nδ1 of His must sit at a comparable distance from the Asp carboxylate; and the angles formed by the three atoms also have a narrow tolerance. When this arrangement holds, His can act as a general base to abstract the proton from Ser, and the activated alkoxide can perform the classic nucleophilic attack on the substrate carbonyl carbon.

What GC-SCAFFOLD does is fill in the remaining protein backbone plausibly while satisfying these geometric constraints. Biologically this amounts to computationally re-exploring the fold space that natural evolution searched over hundreds of millions of years, and the key insight is that more than one fold can support the same catalytic geometry. Serine hydrolase activity in fact arose independently on several evolutionarily unrelated scaffolds - the α/β-hydrolase fold, the chymotrypsin fold, the subtilisin fold - a case of convergent evolution. This is the theoretical justification for de novo design: since the form that hosts a function is not unique, forms that nature never found may also exist.

Failure mode: even when the motif geometry is reproduced exactly, catalysis will not work if the electrostatic environment created by surrounding residues is wrong. The pKa of His can shift anywhere from 4 to 9 depending on its environment, yet to act as a base at physiological pH it must sit near 6-7. GC-ELEC computes this pKa shift and thereby screens out designs whose geometry is right but whose chemistry is wrong.

5.2 Sequence Design - Reading the Folding Code Backwards

Given backbone coordinates, asking which amino acid sequence will fold into that structure is the inverse folding problem. It is fundamentally easier than forward folding because very many sequences satisfy a single structure. The biological basis is the common observation that natural proteins sharing under 30% sequence identity frequently adopt essentially the same fold.

GC-SEQDESIGN produces an amino acid probability distribution conditioned on the local chemical environment of each position - solvent exposure, neighbouring residues, backbone dihedral angles. Biologically this distribution corresponds to the residues evolution would plausibly have permitted at that position. Probability concentrates on Leu/Ile/Val/Phe at hydrophobic core positions and on polar residues such as Glu/Lys/Gln at solvent-exposed positions, matching the hydrophobicity distribution of real proteins.

Two control points matter in practice. First, cysteine exclusion. An unintended Cys in a designed sequence can form incorrect disulfide bonds leading to aggregation or misfolding; in cytoplasmic expression in particular the reducing environment prevents bond formation, leaving free thiols that damage long-term storage stability. Cys should therefore be permitted only when explicitly intended, and must be forbidden at generation time rather than stripped afterwards. Second, prevention of wild-type reversion. Inverse design models often tend to reproduce the original sequence verbatim, which is a good benchmark metric but a failure for novel design purposes. Temperature parameter control that enforces design diversity, together with an identity ceiling relative to the original, is required.

5.3 Structure Prediction - Biological Interpretation of Confidence Metrics

Structure prediction models output confidence metrics alongside coordinates. Regions of high local confidence denote well-defined structure the model is sure about; regions of low confidence can mean two entirely different things. One is model ignorance (lack of information); the other is the biological fact that the region is genuinely intrinsically disordered. Distinguishing between the two is the heart of interpretation.

The way to distinguish them is comparison against sequence-based disorder prediction. If the composition shows classic disorder signatures - low hydrophobicity, high net charge, enrichment in Pro/Gly/Ser - and local confidence is low, genuine disorder is likely and this is not a design defect; the region may serve a functional role as a linker or flexible loop. If, however, the composition is hydrophobic yet confidence is low, that is a design failure signal: hydrophobic residues have failed to find a stable position, which implies aggregation risk.

Interface confidence in complex prediction requires separate interpretation. It is the confidence the model assigns to its own predicted contact geometry; it is not a measure of binding specificity nor of binding affinity (Kd). Conflating these three is a common but incorrect interpretation. High-affinity binding does not necessarily yield high interface confidence, and high interface confidence does not prove that binding occurs - a model can be confidently wrong.

The correct usage is therefore as follows. Interface confidence is used as a TRIAGE metric rather than a binary decision criterion. A low value is a de-prioritisation signal meaning 'the structural basis for this binding hypothesis is weak, so do not proceed without further validation'; a high value merely signals that the candidate is worth committing to physics-based validation (GC-DOCK, GC-MD, GC-FEP), and is not itself a conclusion. Quantitative claims about affinity must rest on GC-FEP results and measured data; we do not assert affinity or specificity from interface confidence alone.

The rationale for the multi-seed strategy and the limits of its interpretation: a single-seed prediction may look good by chance, so the deviation across repeated predictions with different random seeds is used to estimate confidence. What matters is what this deviation actually measures - it is a measure of MODEL UNCERTAINTY, not a physically sampled conformational ensemble. Equating the two is incorrect. Regions with large inter-seed deviation are sometimes observed to coincide with genuinely flexible regions, but this is a correlation rather than an equivalence, and the same deviation can arise from sparse training data or model limitations. Where a genuine conformational ensemble is required, physics-based sampling via GC-MD or enhanced sampling must be performed separately.

5.4 Stability Prediction - What ΔΔG Means at the Bench

Sign convention: throughout this document ΔΔG is defined as ΔΔG = ΔG_variant - ΔG_WT with respect to folding free energy, so negative values denote stabilisation and positive values destabilisation. Because some tools adopt the opposite convention, this must always be checked when comparing against other literature or other tool outputs. Platform deliverables record the sign convention used for each ΔΔG column as metadata.

There is frequent demand to convert ΔΔG into a change in melting temperature (Tm), but no fixed conversion factor exists. The relationship between the two depends on the unfolding enthalpy (ΔH_unfold) and heat capacity change (ΔCp) of the protein concerned; these vary greatly between proteins and even for the same protein under different conditions. This document therefore offers no general conversion of the form 'ΔΔG of 1 kcal/mol equals a Tm shift of X °C'. A target-specific conversion is derived only where thermodynamic parameters have been obtained experimentally for the customer's target; otherwise ΔΔG is used solely as a relative ranking metric between candidates.

There is, however, a well-known trade-off between stability and activity. Enzyme active sites generally adopt thermodynamically unfavourable arrangements: charged residues buried in hydrophobic environments, strained dihedral angles, dense polar networks. This 'strategic instability' supplies the energy needed for substrate binding and transition state stabilisation. Indiscriminate stabilisation around the active site therefore abolishes activity.

The platform handles this trade-off explicitly. Candidate stabilising mutations are stratified by distance from the active site, overlap with the substrate access route, and coevolutionary coupling strength with catalytic residues. Only stabilising mutations at positions sufficiently distant from the active site and with weak coevolutionary signal are classified as 'safe stabilisation' and prioritised. This corresponds to the peripheral stabilisation strategy, an approach demonstrated in industrial enzyme engineering.

5.5 Molecular Dynamics - What a Static Structure Misses

Structure prediction yields a single static set of coordinates, yet most protein function is dynamic. Gate residues must open for a substrate to enter the active site, a different conformational transition is needed for product release, and allosteric regulation is by definition long-range dynamic coupling. GC-MD samples these dynamics by numerically integrating Newton's equations of motion in an environment containing explicit water molecules and ions.

The interpretive metrics and their biological meaning are as follows. If the backbone RMSD trajectory rises initially and then reaches a plateau, the structure has settled into a stable local minimum; continued rise signals that the fold is unravelling. Per-residue RMSF gives a flexibility profile: if active site residues are excessively flexible, catalytic geometry is not maintained and low activity is expected. Secondary structure retention shows whether the designed helices and sheets actually persist, and serves as the final gate on whether the topology intended at design time is physically self-consistent.

Stating the limitation honestly: typical simulations of tens to hundreds of nanoseconds are far shorter than the actual functional transition timescales of proteins (microseconds to milliseconds). MD is therefore powerful as a negative filter confirming that a structure does not collapse, but it cannot prove that a structure functions. Claims that blur this distinction are not scientifically defensible, and the platform uses MD results solely as an exclusion filter.

5.6 Free Energy Perturbation - The Most Expensive and Most Accurate Axis

GC-FEP computes ΔΔG by continuously transforming the wild type into the variant through non-physical intermediate states and integrating the energy change over each interval. This is a statistically rigorous method, and when sufficiently converged agreement with experiment within 1 kcal/mol has been reported. The price is computational cost: hours to tens of hours of GPU time per mutation.

FEP is therefore placed at the very end of the funnel and applied to only the final 20-35 candidates. The biological question this stage answers is whether a mutation genuinely strengthens binding or is an artefact of the earlier approximate models.

Cases genuinely arise where earlier-stage scores and FEP results disagree. FEP is not given unconditional priority in such cases. High confidence is assigned to an FEP result only when all of the following preconditions hold: (a) the starting binding pose is supported by an experimental structure or by independent docking/MD evidence; (b) protonation states have been explicitly determined at the relevant pH; (c) ligand force field parameters have been validated; (d) forward/backward hysteresis lies within tolerance and λ-window overlap is sufficient; and (e) thermodynamic cycle closure error is within the acceptance criterion. If any one of these is unmet, the FEP value is downgraded to indicative status and that fact is flagged in the deliverables. FEP is thus the axis that CAN be the most accurate, not the axis that is always right.

5.7 Immunogenicity - The Leading Cause of Failure in Protein Therapeutics

A substantial share of clinical failure in protein therapeutics arises not from insufficient efficacy but from anti-drug antibody (ADA) formation. The mechanism is as follows: the administered protein is taken up by antigen-presenting cells and degraded in the lysosome; the resulting peptide fragments are loaded onto MHC class II and presented at the surface; CD4+ T cells recognise them and provide B cell help; and neutralising antibodies are ultimately produced. Once antibodies form, the drug is cleared rapidly or its activity is neutralised, and in severe cases cross-reactivity with endogenous proteins creates safety problems.

GC-IMMUNO scans the full designed sequence, predicting per-allele MHC class II binding strength to build a T-cell epitope map. To be technically precise: the MHC class II binding groove is open at both ends, so peptides typically 13-25 residues long are loaded, and what actually determines binding affinity is the 9-residue binding core seated within them. Prediction is therefore performed over 13-25-mer peptide windows while simultaneously estimating the position of the binding core inside each, and the contribution of flanking residues outside the core to affinity is also accounted for. What matters most in biological interpretation is not the absolute number of epitopes but their foreignness. Epitopes identical to sequences present in human proteins are generally ignored due to central tolerance, whereas sequences absent from the human proteome - as in de novo designed proteins - are not subject to tolerance and therefore carry higher risk. De novo design consequently carries intrinsically greater immunogenicity risk than engineering of an existing protein.

Mitigation strategies are also supported computationally. Once epitope hotspots are identified, substitutions are searched that weaken MHC binding at those positions while preserving structure and function - deimmunisation. GC-STAB and GC-FOLD are used concurrently to confirm that the deimmunising mutations do not damage stability or activity, because single-axis optimisation invariably sacrifices another axis.

5.8 Solubility and Aggregation - The Physical Basis of Manufacturability

Protein aggregation begins when partially unfolded intermediates associate through exposed hydrophobic surfaces. Once a nucleus forms, growth accelerates autocatalytically, and aggregates are known to raise immunogenicity substantially, making them a key control item from a regulatory standpoint as well.

GC-SOL predicts aggregation-prone hotspots from sequence and structure. The physical basis is that aggregation-inducing sequences have a characteristic composition: stretches combining consecutive hydrophobic residues, β-sheet propensity and low net charge are dangerous. Charged residues, conversely, suppress aggregation through electrostatic repulsion and act as charge gatekeepers. The computational result leads to two practical actions. First, if a hotspot is surface-exposed, it is mitigated by introducing charged residues. Second, if a hotspot is structurally essential, the recommendation is to address it through formulation (pH, ionic strength, surfactant) rather than sequence change. Because net charge approaches zero and aggregation rises sharply when the isoelectric point approaches the formulation pH, the design stage adjusts to keep at least 1.5 units of separation between the pI computed by GC-ELEC and the target formulation pH.

5.9 Codon Optimisation - The Biology of Translation Dynamics

Countless DNA sequences encode the same amino acid sequence, and the choice among them can swing expression level by tens of fold. The simple view - use the codons most frequent in the host - is only partly correct.

The sophisticated view is as follows. First, in the roughly 30-50 codons at the 5' end, rare codons and weak mRNA secondary structure are actually advantageous, because strong hairpins in this region physically impede ribosome entry and reduce initiation efficiency; indeed, codon usage at the 5' end of E. coli genes is observed to be less optimised than the genome average. Second, moderate translational slowdown at domain boundaries can benefit co-translational folding, buying time for the nascent chain to fold after it leaves the ribosome exit tunnel and before the next domain is synthesised. Third, the tRNA pool varies with growth conditions, so induction conditions and medium influence the optimal codon choice.

GC-GENE handles these factors together while simultaneously checking synthesis feasibility. Specifically it flags regions where local GC content falls outside 30-70%, direct repeats of 8 bp or more, restriction enzyme recognition sites, strong hairpins (below a ΔG threshold) and homopolymer runs of six or more identical bases - all leading causes of synthesis failure or cost escalation. Because this check is integrated into the design stage, redesign triggered by a 'cannot be synthesised' verdict does not structurally occur.

6. Performance Comparison Against Industry Standards

This chapter compares the platform against baselines common in the industry. The comparison is made at three levels: (a) accuracy of individual prediction axes, (b) enrichment efficiency at the pipeline level, and (c) operational and regulatory maturity. For fairness, items where the platform is inferior are also stated explicitly.

6.1 Definition of Comparison Baselines

6.2 Accuracy Comparison by Prediction Axis

Table 6-1. Accuracy by axis. Ranges for public tools and commercial suites are typical values reported in benchmark literature for the field [literature]. Because specific figures depend on dataset and evaluation protocol, this should be read as a grade comparison rather than an absolute one.

A caution on interpretation: the table above is not a superiority ranking. Platform figures come from internal data while comparator figures come from external reports using different datasets and evaluation protocols, so a direct comparison does not hold. Claiming superiority from the relative magnitude of two numbers measured on different datasets is not methodologically valid; the 'advantage/comparable' verdicts used in the previous edition have therefore been withdrawn and replaced with a 'comparison status' column. Head-to-head comparison on a common benchmark is registered as a remediation item in section 10.5.

The table does nonetheless convey valid information: the performance levels of individual prediction tools have largely converged, and no axis shows an overwhelming gap. This suggests that differentiation arises not from the accuracy of individual axes but from how the axes are combined and in what order they are applied. The next section addresses exactly that.

6.3 Pipeline-Level Enrichment Efficiency - The Real Differentiator

Enrichment efficiency is defined as the ratio of the success rate among N candidates that passed the computational filters to the success rate among N drawn at random. This is the metric most directly tied to the value the customer is paying for.

Table 6-2. Enrichment efficiency by scenario - all figures graded [internal observation, unaudited]. The validation caveat below must be read together with this table.

6.4 Functional Coverage Matrix

Table 6-3. Y = fully supported, P = partial or requires a separate module, N = unsupported. [internal assessment]

The matrix reveals four areas unique to this platform: integration of freedom-to-operate review into the design stage, end-to-end connection through to gene synthesisability, automatic uncertainty-based weighting, and closed-loop recalibration. These are not individual-prediction-accuracy features but functions at the level of turning research output into a commercialisable asset - an area generally not addressed by either academic pipelines or commercial suites.

6.5 Operational Maturity Comparison

Table 6-4. Operational maturity. The test count is a measured value from the internal regression suite. [internal]

Operational maturity is an undervalued factor determining the real-world utility of a computational platform. If a molecular dynamics run lasting several days fails midway and cannot be resumed, no amount of theoretical accuracy will move the project forward. This platform attaches a computation fingerprint to each stage cache so that only stages whose code changed are re-executed while the rest are retrieved instantly from cache. This structurally prevents the silent failure in which a deployment changes nothing at all.

6.6 Areas Where This Platform Is Inferior (Honest Disclosure)

• Graphical interface: interactive visualisation tooling is limited compared with commercial suites. The platform is optimised for batch processing and report generation; for hands-on structural exploration we recommend a separate viewer.

• Membrane proteins and large complexes: prediction confidence is currently low for membrane proteins in a lipid bilayer environment and for supramolecular complexes above 500 kDa. Such targets are explicitly flagged during the feasibility scan.

• PTM-dependent function: for proteins whose function requires glycosylation, phosphorylation and the like, current computational models do not adequately capture these modifications. This is a constraint for projects targeting eukaryotic expression systems.

• Intrinsically disordered proteins (IDPs): for targets without a static structure, most of the structure-centric evaluation axes in this pipeline become meaningless. A separate ensemble-based approach is required.

• Absolute activity values: the platform is reliable for relative ranking among candidates but does not predict absolute values of k_cat or K_M. This is a limitation of the field as a whole; no tool currently provides reliable absolute values.

• Long-timescale dynamics: processes taking microseconds or longer, such as allosteric transitions or domain rearrangements, cannot be simulated directly and are only partially accessible even with enhanced sampling.

• Cellular environment not represented: computations generally assume dilute solution, whereas the cytoplasm is a crowded environment exceeding 300 mg/mL protein concentration, and this crowding effect is not adequately modelled.

7. Application to Drug Development - Strategy by Modality

Protein engineering plays two roles in drug development: cases where a protein (or its derivative) is itself the medicine (biologics), and cases where the protein is a tool that makes or delivers a medicine. The platform supports both, reconfiguring a single computational skeleton - the 17-stage funnel of Chapter 3 and the GC modules of Chapter 4 - by modality to produce different deliverables. This chapter enumerates concretely what can be produced, modality by modality, and states honestly the central design challenge of each, the modules that address it, and its maturity.

7.1 The Landscape of Producible Modalities

The modalities the platform can generate or evaluate for a single target are listed below. 'Entry mode' refers to the two entry points of Chapter 3 (improvement mode, which engineers an existing molecule, and de novo mode, which creates a scaffold not found in nature) or an assessment mode that only evaluates a supplied design. 'Maturity' is an honesty grade, distinguishing whether a generative pipeline is fully in place, to avoid overstatement.

Table 7-1. Producible modalities. Maturity: ** full generative pipeline / * improvement-mode plugin / ~ assessment and gating only / (open) latent, needs user data. Each modality is elaborated in sections 7.2-7.6 and Chapter 8. The 'representative example' column lists external public references showing each modality category is real (mixed approved / clinical / research stages), not platform outputs. [internal assessment]

7.2 Protein and Antibody Therapeutics

7.2.1 Enzyme Replacement Therapy

In diseases caused by a specific enzyme deficiency, such as lysosomal storage disorders, recombinant enzyme is administered intravenously (public examples: imiglucerase for Gaucher, agalsidase beta for Fabry, alglucosidase/avalglucosidase alfa for Pompe, olipudase alfa for ASMD). The central challenge of this modality is not potency but in vivo stability, delivery to the target tissue, and immunogenicity. Intracellular lysosomal uptake in particular depends on the mannose-6-phosphate (M6P) receptor pathway, so glycan design governs efficacy, and in a deficient patient the protein is readily recognised as foreign, promoting anti-drug antibodies (ADA).

Platform application: (a) GC-STAB and GC-REMD design peripheral stabilising mutations that simultaneously raise plasma protease resistance and thermal stability. (b) GC-IMMUNO assesses ADA risk in advance and proposes deimmunising substitutions. (c) GC-ELEC evaluates activity and stability at the acidic pH of the lysosome (approximately 4.5-5.0); an enzyme optimal at neutral pH may be inactivated there, so retention of activity under acidic conditions is part of the design objective. (d) GC-KCAT ranks relative catalytic efficiency among candidates and GC-PK evaluates a plasma half-life proxy. M6P glycosylation itself depends strongly on the expression host and process and is therefore optimised at the wet-lab/process stage rather than computationally, linking to the expression design of Chapter 5.

7.2.2 Antibodies and Antibody-Derived Molecules

Antibodies account for most of the biologics market and are the most mature area of protein engineering (historical milestones: the first fully human antibody, adalimumab, came from phage display, and humanisation was established via CDR grafting). The platform designs not only full-length IgG but also Fab/scFv fragments and camelid single-domain nanobodies (VHH; public approved example caplacizumab) within one pipeline. Recently, de novo antibody design generating the binding site directly from structure has been demonstrated (public literature: RFdiffusion family, Nature 2025), and the platform's de novo entry mode corresponds to this trend.

Platform application: complementarity-determining region (CDR) libraries are designed with GC-LIBRARY, antigen-antibody complexes are predicted with GC-FOLD, and binding delta-delta-G for top candidates is computed precisely with GC-FEP. In de novo entry, GC-SCAFFOLD generates a new binding scaffold anchored on the antigen epitope. Optimising affinity alone promotes candidates that aggregate readily, have extreme isoelectric points, or contain chemically labile motifs (deamidation-prone Asn-Gly, isomerisation-prone Asp-Gly, oxidation-prone exposed Met), so evaluating developability axes in parallel is essential. GC-DEVELOP scans these motifs automatically, GC-PK assesses viscosity and half-life, and GC-IMMUNO evaluates humanness and T-cell epitopes, pulling CMC-stage problems forward into the design stage.

7.2.3 Bi/Multispecific Antibodies and T-Cell Engagers

Bispecific antibodies bind two targets simultaneously (or one target and CD3 on T cells) to create a novel mechanism (public approved examples: CD19xCD3 blinatumomab, gp100xCD3 tebentafusp). Two design challenges dominate this modality. First, the 'chain-pairing problem' - different heavy and light chains must pair correctly - solved by format engineering such as knobs-into-holes, CrossMab and common light chains. Second, for T-cell engagers, CD3 affinity must be tuned to balance efficacy against the risk of cytokine release syndrome (CRS).

Platform application: each binding arm is generated with the antibody pipeline of 7.2.2, and GC-FORMAT evaluates and gates chain-pairing fidelity (mispairing risk) and CRS risk in the chosen format. GC-FORMAT currently operates as developability assessment and gating over a supplied combination (maturity triangle), while generation of the candidate arms themselves is handled by the upper antibody family (maturity circle-double). GC-FOLD checks the geometric viability of the ternary complex (two antibody arms plus target/T-cell), and GC-PK evaluates format-dependent half-life (for example the short half-life of small engagers).

7.2.4 Antibody-Drug Conjugates (ADC)

ADCs combine the target selectivity of an antibody with the killing power of a cytotoxic payload (public approved examples: trastuzumab emtansine with a non-cleavable linker and DAR of about 3.5; trastuzumab deruxtecan with a cleavable linker, DAR of about 8 and a strong bystander effect). Four factors dominate clinical failure: (1) drug-to-antibody ratio (DAR) - excessive DAR causes aggregation and fast clearance; (2) linker stability - risk of premature release in systemic circulation versus intratumoural bystander killing; (3) total hydrophobic load from payload hydrophobicity; and (4) conjugation homogeneity - the predictability of site-specific conjugation (THIOMAB, unnatural amino acids) versus random attachment.

Platform application: the antibody portion is designed with the 7.2.2 pipeline, and GC-FORMAT's conjugation developability assessment quantifies the optimal DAR window (Gaussian, favouring a clinically stable mid-range DAR), payload-linker hydrophobic load, site-specific conjugation homogeneity, and linker release profile. GC-PK evaluates clearance and hepatic exposure risk from rising DAR, and GC-DEVELOP evaluates post-conjugation aggregation risk. Because the bystander effect is a benefit in solid tumours with low target-antigen density but a toxicity in normal tissue, GC-FORMAT treats it as a context-dependent weight.

7.2.5 De Novo Binders / Mini-Binders

Mini-binders are small proteins of 60-100 amino acids designed entirely de novo to bind a target surface with high affinity. This modality is a flagship demonstration of computational design itself (public literature: SARS-CoV-2 spike miniprotein inhibitors, Science 2020; design of binders from the target structure alone, Nature 2022; RFdiffusion, Nature 2023). Being smaller than antibodies, they are advantageous in tissue penetration, manufacturability and stability, and can be designed from a target-surface hotspot alone without a known binding partner. The central challenge is the low experimental hit rate: many designs fail to express or bind, so enriching successful candidates with computational filters is the real value of this modality.

Platform application: GC-SCAFFOLD generates a binding scaffold anchored on hotspot residues of the target surface, GC-SEQDESIGN inverse-designs a sequence that folds into it, and the multi-seed consensus of GC-FOLD removes candidates with low predicted interface confidence. GC-DOCK and (when convergence conditions are met) GC-FEP evaluate interface binding precisely, and GC-MD validates persistence of the complex under physiological conditions. The funnel of Chapter 3 works most dramatically for this modality, because compensating for a low hit rate with multi-axis filters is the whole point (see enrichment efficiency in section 6.3).

7.3 Modalities Beyond Antibodies

7.3.1 Targeted Protein Degraders (PROTAC / Molecular Glue)

Targeted protein degraders, instead of inhibiting a target, route it to the cell's ubiquitin-proteasome machinery for degradation. Molecular glues glue a neosubstrate to an E3 ligase (for example cereblon; public examples: the IMiD class such as lenalidomide), while PROTACs are bifunctional molecules linking a target ligand and an E3 ligand (public late-clinical examples: the ER degrader vepdegestrant and the AR degrader bavdegalutamide - approval status to be confirmed at publication). The central design challenge is whether the target-degrader-E3 'ternary complex' forms with productive geometry, with PROTAC linker length and composition governing cooperativity and the 'hook effect'.

Platform application: GC-TERNARY evaluates and gates the cooperativity (alpha), linker geometry fit, and DC50/E3 proxies of the ternary complex (maturity triangle - assessment of a supplied degrader design, not generation). GC-DOCK generates ligand-binding pose ensembles for the target and E3 separately, and GC-PK evaluates the 'beyond rule of 5' permeability and oral-absorption risk characteristic of bifunctional molecules. Where the target protein itself needs a novel scaffold, a partner protein can be designed with the upper generative families.

7.3.2 Peptide and Macrocyclic Peptide Therapeutics

Peptide therapeutics fill the space between small molecules and antibodies - more specific than small molecules and smaller than antibodies, so they can reach intracellular targets and protein-protein interactions (public examples: the GLP-1 class, oral macrocyclic PCSK9 inhibitors). Macrocyclisation, N-methylation and backbone rigidification enter the 'beyond rule of 5' chemical space, and the dominant challenges are membrane permeability and oral bioavailability, protease stability, and short half-life.

Platform application: GC-SCAFFOLD generates cyclic and linear peptide scaffolds anchored on a target-surface motif, GC-PK evaluates membrane permeability and oral-absorption proxies, and GC-IMMUNO evaluates MHC-binding immunogenicity (both the low immunogenicity of short sequences and the risk upon carrier conjugation). The co-folding of GC-FOLD validates target-peptide complex geometry, and protease stability is addressed by scanning cleavage-prone motifs. This modality runs largely on CPU, allowing large candidate sets to be screened at low cost.

7.3.3 Engineered Cytokines and Biobetters

Cytokines are potent but have a narrow therapeutic window. The representative challenge is receptor-subunit bias: IL-2, for example, can be biased towards oncology (effector T cells) or autoimmunity (regulatory T cells) by decoupling CD25 (Treg-biasing) engagement from IL-2R-beta-gamma engagement (public examples: the no-alpha IL-2 class, PEGylated Treg-selective IL-2, and the fully computationally designed IL-2 mimetic Neo-2/15, Nature 2019). Biobetters improve existing biologics through extended half-life, better stability, or elimination of cold-chain requirements.

Platform application: GC-STAB and GC-REMD design stabilising and half-life-extending mutations, GC-FORMAT handles Fc-fusion and dimerisation formats, GC-IMMUNO assesses the risk of new T-cell epitopes from engineered substitutions, and GC-PK evaluates half-life and exposure. Receptor-bias design is approached by predicting the cytokine-receptor complex with GC-FOLD and precisely computing per-subunit binding delta-delta-G differences with GC-FEP, searching for mutations that retain binding to one subunit while weakening binding to another. In de novo entry, GC-SCAFFOLD can generate an entirely novel mimetic reproducing the receptor-binding interface.

7.3.4 Small-Molecule Hit Discovery (Auxiliary Modality)

The platform's only non-protein modality, generating and evaluating small-molecule hits against a target pocket. Uses include a standalone small-molecule candidate, the warhead of a degrader, or a combination small molecule. A generative chemistry model produces candidates, which are filtered by docking, interaction fingerprinting, quantum cross-checking and retrosynthesis analysis. This modality runs on CPU, with heavy steps such as docking, QM and retrosynthesis activated optionally.

Platform application: GC-DOCK handles binding poses, GC-QM the quantum-chemical validity of interactions, and GC-FF force-field parameters for non-standard ligands. Synthesisability (existence of a retrosynthetic route) and drug-likeness (ADMET, toxicity, BBB proxies) act as quality gates, removing non-synthesisable or strongly toxic candidates early. The same uncertainty-propagation and inverse-variance fusion principles as the protein modalities apply.

7.3.5 TCR / pMHC Design (TCR-T)

T-cell receptor (TCR)-based therapies can recognise peptide-MHC (pMHC) complexes derived from intracellular as well as surface proteins, targeting tumour antigens that antibodies cannot reach (public example: the soluble TCR-based bispecific tebentafusp). The central challenge is raising specificity and affinity for the pMHC while suppressing cross-reactivity (alloreactivity) against normal-tissue self-peptides - excessive affinity maturation can cause fatal off-target toxicity.

Platform application: GC-FOLD predicts TCR-pMHC co-folding, GC-MD validates interface stability, and a dedicated alloreactivity assessment scores the binding risk against similar self-peptides. GC-IMMUNO evaluates per-allele MHC presentation and GC-FEP precisely computes binding delta-delta-G for top candidates. The specificity-affinity balance is set as an explicit objective so that high-affinity-only risky candidates do not rise to the top.

7.4 Genome and Transcriptome Editing Tools

A genome editing tool is a complex of protein (enzyme) and nucleic acid (guide), so both components must be optimised together. Three challenges are common from a clinical perspective: minimising off-target action, deliverable size, and immunogenicity (the high risk characteristic of microbially derived proteins). The subsections below specify the tools the platform can generate or evaluate, by editing modality.

7.4.1 Miniature Cas Nucleases (Single AAV)

CRISPR therapies have begun entering the clinic (public approved example: the first CRISPR therapy, exa-cel, editing the BCL11A enhancer, 2023), but the flagship enzyme SpCas9 (about 1,368 aa) is too large to package into a single AAV (about 4.7 kb limit) together with a promoter and guide. This calls for compact enzymes such as SaCas9 (about 1,053 aa) or the miniature Cas12f family (AsCas12f1 about 422 aa, engineered CasMINI about 529 aa), but compact enzymes tend to have lower on-target activity that protein engineering must recover. The de novo design of exactly these miniature nucleases is the platform's flagship pipeline.

Platform application: the full 17-stage funnel is deployed for this modality. GC-SCAFFOLD generates a scaffold hosting the RuvC/HNH catalytic domains, GC-SEQDESIGN inverse-designs the sequence, GC-GUIDE predicts on-target efficiency (in internal validation, Spearman rho of about 0.78-0.79 with measured editing efficiency [internal]), and GC-OFFTARGET evaluates genome-wide off-targets. GC-FOLD and GC-MD validate the structure and stability of the nuclease-guide-DNA ternary complex, a catalytic-geometry gate checks the catalytic viability of the active site, and GC-IMMUNO assesses the immunogenicity risk of the microbially derived protein. The final deliverable carries a payload profile for LNP delivery (section 7.6.2).

7.4.2 Base Editors and Prime Editors

Base editors convert a single base without a double-strand break (cytosine CBE, adenine ABE), and prime editors perform 'search-and-replace' editing with an nCas9-reverse transcriptase fusion and pegRNA (public methods: Nature 2016/2017 and Nature 2019 respectively; public clinical programmes: in vivo base editing of PCSK9/ANGPTL3, prime editing for chronic granulomatous disease). Note: the earlier exa-cel is Cas9 double-strand-break editing, not base editing, and the two must be distinguished. The central challenges are bystander edits within the window, off-target DNA/RNA editing by the deaminase, and delivery of the large editor-plus-RT cassette.

Platform application: the platform designs base editors (deaminase-Cas fusion, default TadA-8e class) and prime editors (Cas-RT fusion, default M-MLV RT class) with dedicated fusion pipelines. GC-FOLD validates the holo structure of the fusion, GC-MD the linker and RT-loop dynamics (critical, since linker length and flexibility govern activity in fusion proteins), and GC-GUIDE/GC-OFFTARGET evaluate the editing window and off-targets. Bystander-edit risk is addressed by explicitly scoring the presence of non-target bases within the editing window.

7.4.3 Epigenetic Editors (dCas Effectors)

Epigenetic editors fuse an epigenetic effector (for example a DNA methyltransferase or demethylase) to catalytically dead dCas to switch gene expression on or off without cutting DNA (public concept: stable methylation editing such as CRISPRoff). The DNA-non-cutting nature gives a favourable safety profile, but durability of the effect and accuracy of the target position relative to the transcription start site (TSS) are what matter.

Platform application: an epigenetic-effector scoring assessment evaluates and gates the placement of the DNA-non-cutting effector and its TSS-distance suitability (maturity triangle). GC-FOLD validates the structural viability of the dCas-effector fusion and GC-OFFTARGET evaluates off-target binding of the dCas itself. Where scaffold and sequence generation for the fusion is needed, it is delegated to the upper generative families (GC-SCAFFOLD, GC-SEQDESIGN).

7.4.4 RNA Editing and Suppressor tRNA

This family acts at the RNA level without permanently altering DNA, giving reversibility and safety advantages. ADAR-recruiting editing uses a guide oligo to recruit endogenous ADAR for A-to-I editing (read biologically as G), reverting a point mutation (public example: the first-ever demonstration of therapeutic RNA editing in humans, WVE-006, in AATD). Suppressor tRNAs read through premature termination codons (PTCs) to restore full-length protein expression (public literature: engineered tRNAs, Nature 2023; reported dystrophin restoration in DMD). The central challenges are, respectively, specificity against off-target A-to-I editing, and achieving high readthrough efficiency without suppressing native stop codons.

Platform application: an ADAR-guide scoring assessment evaluates the editing window and index, and a suppressor-tRNA scoring assessment evaluates and gates PTC readthrough efficiency and the +1 context (maturity triangle). Because these two sub-modalities are nucleic-acid sequence design, they link to the guide discovery and design family of GC-GUIDE, and off-target editing/suppression risk is handled as an extension of the GC-OFFTARGET principle. With no protein component, the structural and MD families are largely deactivated, keeping compute cost low.

7.4.5 Site-Specific Recombinases

Serine recombinases (for example Bxb1, phiC31, TP901) recognise specific attachment-site (att) pairs to excise, insert or invert large DNA cassettes. Unlike CRISPR, they do not depend on host repair after a double-strand break, which is advantageous for precise large insertions (for example full-gene knock-in). The central challenges are att-site specificity and recombination efficiency, and reprogramming specificity to a new target site.

Platform application: the platform designs recombinases in de novo entry mode. GC-SCAFFOLD generates the catalytic scaffold and GC-SEQDESIGN inverse-designs the sequence; GC-MD validates the stability of the recombinase-DNA complex and GC-FOLD the geometry of the att-recognition interface. Reprogramming att specificity is approached with a combinatorial library of DNA-contacting residues (GC-LIBRARY).

7.5 Molecular-Biology and Manufacturing Enzyme Toolkit

Where the previous sections dealt with 'proteins that become medicines', this section deals with 'enzymes that make medicines and genetic material'. The platform engineers the enzyme families used in molecular biology reactions with dedicated plugins - that is, an improvement mode that improves fidelity, thermostability, catalytic efficiency and substrate range from a wild-type (or commercial-equivalent) starting point. This toolkit underpins upstream manufacturing across mRNA therapeutics, enzymatic gene synthesis, PCR/isothermal amplification, cloning, and next-generation-sequencing library preparation.

Table 7-2. Molecular-biology enzyme toolkit (summary of 14 improvement-mode plugin classes). Targets are public equivalent categories of commercial enzymes; an actual project starts from a customer-specified starting sequence. [internal assessment]

7.5.1 High-Fidelity Polymerases and Reverse Transcriptases

Polymerase fidelity is defined by error rate, which GC-FIDELITY estimates by multi-axis consensus - combining base-selection energy, polymerase-to-proofreading (exo) domain geometry, and activation barrier. These axes are reliable for relative ranking among candidates but do not produce an absolute error rate (see the notices in sections 6.6 and 4.6). Thermostability is evaluated by the transition-temperature estimate of GC-REMD, and catalytic turnover by GC-KCAT. This family is used directly in high-fidelity PCR, NGS library preparation and reverse transcription.

7.5.2 RNA Polymerases (Low-dsRNA T7)

T7 RNA polymerase is the key enzyme for producing mRNA therapeutics and vaccines by in vitro transcription (IVT). Wild-type T7 generates double-stranded RNA (dsRNA) byproducts that stimulate innate immunity, so low-dsRNA variants are decisive for mRNA purity and safety. The platform supports RNAP engineering whose objective is low dsRNA, 2'-modified nucleotide acceptance and yield (public target category: the low-dsRNA T7 class). GC-FIDELITY evaluates transcription fidelity, GC-MD the stability of the elongation complex, and GC-KCAT the transcription efficiency.

7.5.3 TdT / Enzymatic Nucleic Acid Synthesis

Terminal deoxynucleotidyl transferase (TdT) adds nucleotides to a 3' end without a template and is central to enzymatic DNA synthesis, which replaces phosphoramidite chemistry that uses hazardous reagents (public direction: evolving TdT to accept 3'-reversible-terminator nucleotides for single-base control; literature reports about 80 substitutions over 32 rounds with over 99% per-cycle addition). Natural TdT is template-independent but uncontrolled, so single-base control requires redesigning the active site to efficiently accept bulky, charged reversible-terminator substrates, along with activity on structured 3' ends (hairpins) and thermostability.

Platform application: this modality is the platform's origin and a flagship validation target. A dedicated TdT plugin evaluates selectivity across a 4-way dNTP panel and a modified-nucleotide panel, and a divalent-ion geometry gate validates catalytic coordination. GC-KCAT evaluates relative per-substrate addition efficiency, GC-DOCK the binding pose of reversible-terminator substrates, and GC-STAB/GC-REMD stability, with stabilising designs such as disulfide introduction searched alongside. Because per-cycle efficiency governs error accumulation in long oligo synthesis, the ranking of relative addition efficiency is the most important output metric of this modality.

7.5.4 Industrial Biocatalysis and De Novo Enzymes

Biocatalysts used in active pharmaceutical ingredient (API) manufacture offer higher stereoselectivity than chemical synthesis and reduce organic solvents and heavy-metal catalysts, advantageous both regulatorily and environmentally (public example: an engineered transaminase for sitagliptin manufacture). Beyond this, the fully de novo design of enzymes catalysing reactions absent from nature has been demonstrated as a flagship achievement of computational design (public literature: retro-aldolase Science 2008, Kemp eliminase Nature 2008, deep-learning luciferases Nature 2023, catalytic-motif-scaffolded high-efficiency enzymes Nature 2025). First-generation de novo enzymes had low kcat requiring directed evolution, but recent methods approach natural-enzyme efficiency by design alone.

Platform application: the objective function for industrial enzymes differs from that for therapeutic proteins - immunogenicity is irrelevant, while tolerance to high temperature, organic solvents and extreme pH, and a high turnover number (TON), matter. The thermal-stability axis of GC-REMD and the organic-solvent weighting of GC-MD are raised and GC-IMMUNO is deactivated; this reconfiguration happens at the configuration level and requires no code change. In de novo enzyme design, GC-SCAFFOLD places the active site anchored on a catalytic motif (theozyme), GC-QM evaluates transition-state stabilisation (lowering the activation barrier), and GC-KCAT evaluates relative catalytic efficiency.

7.6 Delivery-Oriented Design

7.6.1 AAV Capsid Engineering

AAV is the mainstream approved gene-therapy vector (public examples: retinal voretigene, SMA onasemnogene abeparvovec), but controlling tissue tropism and evading pre-existing neutralising antibodies are unsolved challenges. The landmark engineered variant PHP.eB crosses the blood-brain barrier via a 7-mer peptide insertion between residues 588-589 of AAV9, but this crossing is dependent on a specific mouse strain (the Ly6a receptor) and is reported not to translate to primates - a textbook case of the 'preclinical-to-clinical translation gap', which the platform flags explicitly.

Platform application (maturity open-triangle - latent): a capsid-insert library diversity and developability assessment and a neutralising-antibody epitope panel are implemented, but meaningful output requires the customer to supply AAV serotype sequences and insertion positions (no specific serotype sequence is built into the platform). GC-FORMAT evaluates the diversity of the insert-peptide library, GC-IMMUNO neutralising-antibody evasion, and GC-DEVELOP capsid-assembly-related developability. Because the mouse-to-primate translation risk is not resolved computationally, cross-species validation is stated as mandatory.

7.6.2 LNP Payload Matching

When a protein or gene-therapy payload is delivered by lipid nanoparticle (LNP), the payload's physical properties (molecular weight, isoelectric point, surface charge) govern the ionisable-lipid ratio and buffer-pH choice. The platform's design deliverables attach a payload profile carrying these properties in a standard schema, handed off losslessly to the delivery-design stage (for example when delivering the miniature Cas nuclease RNP of 7.4.1 by LNP).

Platform application: GC-ELEC produces surface charge and isoelectric point, GC-SOL solubility and aggregation, and these values are structured into the payload profile. This acts as a contract interface with a separate delivery-design pipeline (LNP optimisation), with schema stability guaranteed in the same way as the module-replaceability guarantee of section 4.6.

7.7 Regulatory Perspective

When in-silico design results are used in a regulatory submission, what reviewers require is not predictive accuracy per se but traceability of how the decision was reached. The platform provides three things for this. First, configuration snapshots and computation fingerprints are preserved for every run, so identical results can be reproduced. Second, stage-by-stage logs record at which stage and why each candidate was selected or eliminated. Third, predictions carry attached uncertainties - GC-UQ produces conformal prediction intervals, an applicability domain, and a model credibility grade, answering quantitatively how much a value can be trusted and, further, whether a target lies within the model's applicability. This applies across all modalities and links to the critical quality attribute (CQA) and regulatory-pathway mapping of ICH Q8 quality by design (QbD).

It must be stated clearly, however, that no regulatory authority currently accepts safety or efficacy on the basis of in-silico prediction alone. The regulatory role of computational results lies in justifying experimental design and providing a basis for a risk-based approach, and the platform claims value only within that scope. As modalities diversify, each has its own regulatory pathway (for example gene therapy, cell therapy, vaccine, small molecule), so the GC-UQ and QbD mapping serves to indicate the form of the required documentation early, per modality.

8. Application to Vaccine Development - Antigen Design and Immunogenicity

In vaccine development the role of protein engineering is the opposite of that in therapeutic proteins. For therapeutics the goal is to minimise immune response; for vaccines it is to maximise the desired immune response while suppressing undesired ones. This reversal of direction changes the entire design logic.

8.1 Antigen Stabilisation - The Core of Structure-Based Vaccinology

Many viral surface glycoproteins undergo large-scale structural transitions during cell entry. Prefusion and postfusion conformations exist, and the most potent epitopes recognised by neutralising antibodies generally exist only in the prefusion form. The problem is that the prefusion form is metastable: it transitions spontaneously to the postfusion form during purification or storage, losing the neutralising epitopes.

The central strategy of structure-based vaccinology is to lock the prefusion form structurally. Representative methods are (a) introducing prolines into flexible hinge regions to physically block helical extension, (b) introducing disulfide bonds spanning two domains to seal off the transition, and (c) introducing cavity-filling mutations in the hydrophobic core to stabilise the prefusion form thermodynamically. This approach is a principle demonstrated in the development of several respiratory virus vaccines.

Platform application: GC-FOLD predicts both prefusion and postfusion conformations, and GC-MD compares their relative stability. GC-STAB searches for mutations that selectively stabilise the prefusion form, and the key point is that what is optimised is the difference in stability between the two states (ΔΔΔG) - simply raising stability would stabilise the postfusion form equally and achieve nothing. Disulfide bond candidates are validated against the GC-FOLD structure to check whether Cβ-Cβ distance and dihedral geometry fall within the range where a bond can actually form. In one internal validation case, zero of six candidates were judged geometrically feasible - an example of structure-based validation catching the common error of proposing disulfide bonds from distance alone [internal].

8.2 Epitope Focusing

When a full-length antigen is used, the immune system concentrates its response on easily accessible immunodominant epitopes, and these are not necessarily neutralising epitopes. Indeed, many viruses achieve immune evasion by keeping immunodominant regions variable - a strategy that directs antibodies towards sites that confer no protection.

The countermeasure is to design a small scaffold antigen carrying only the conserved neutralising epitope. This is essentially a de novo design problem: GC-SCAFFOLD takes the epitope backbone geometry as a motif and generates a new scaffold that reproduces it precisely. Success criteria are (a) backbone RMSD of the grafted epitope within 1 Å of the original, (b) a valid predicted complex with an existing neutralising antibody that recognises the epitope, and (c) the scaffold itself not creating new immunodominant epitopes. The third condition is particularly demanding, because a newly designed scaffold is an entirely novel sequence absent from the human proteome and can therefore be a strong immunogen in its own right. GC-IMMUNO quantifies this risk by computing epitope density separately for the scaffold region and the epitope region.

8.3 Broadly Protective Antigen Design

For rapidly mutating pathogens, an antigen covering an entire lineage rather than a specific strain is required. Computational approaches include ancestral sequence reconstruction and consensus design. Ancestral reconstruction estimates the common ancestral sequence of extant variants using a phylogenetic tree, and it has been observed across several families that reconstructed ancestral proteins tend to have higher thermal stability and broader substrate range than their extant descendants. Applied to vaccine antigens, this may induce antibodies that cross-react across multiple variants.

Platform application: GC-COEVO and GC-PLM handle phylogenetic information, and the ancestral reconstruction capability of GC-LIBRARY generates candidate ancestral sequences. The resulting ancestral antigens pass through the full pipeline again for evaluation of expressibility, stability and immunogenicity. We state clearly that breadth of protection itself cannot be proven computationally and requires animal validation.

8.4 Design Constraints by Vaccine Platform

8.5 Safety Signals Specific to Vaccines

Two safety items must be checked in antigen design. First, sequence similarity to human proteins. If an antigen sequence has significant similarity to a human protein, there is a risk of autoimmune response through molecular mimicry. GC-SIMSEARCH performs a genome-wide search against the human proteome and flags similar segments. Second, epitopes associated with antibody-dependent enhancement (ADE) risk. In some pathogens non-neutralising antibodies have been reported to promote infection, and designs that avoid known risk epitopes are required. Because this depends heavily on pathogen-specific literature knowledge, the platform does not provide an automatic verdict; it reports only whether such segments are present and requests expert judgement.

8.6 Personalized Neoantigen Vaccines

Neoantigens arising from tumour-specific somatic mutations are absent from normal tissue and are therefore ideal tumour targets. A personalized neoantigen vaccine is a pipeline that, within weeks, performs patient tumour/normal sequencing, neoantigen prediction with HLA-binding and immunogenicity ranking, selection of the top epitopes (typically 20-34), and mRNA/LNP manufacture (public clinical examples: mRNA-4157/V940 plus pembrolizumab in melanoma, KEYNOTE-942, reduced recurrence risk; autogene cevumeran in pancreatic cancer, Nature 2023, neoantigen-specific T-cell responses). The bottleneck is accurate neoantigen prioritisation.

Platform application (maturity circle-double - full assessment and ranking pipeline): mutant and wild-type peptides are extracted from somatic variant calls (fixing a canonical transcript is essential - without it, variant coordinates shift and the wrong peptide is produced), HLA genotype is determined, and GC-IMMUNO scores MHC class I and II presentation and binding strength per allele. GC-UQ attaches predicted confidence and applicability domain to each candidate so that low-confidence neoantigens do not rise to the top. Clonality and expression filters are applied together to prioritise clonal neoantigens present throughout the tumour.

8.7 Self-Assembling Antigen Carriers (VLP / Nanoparticles)

Presenting an antigen multivalently on a nanoparticle surface amplifies the immune response through B-cell receptor crosslinking. Virus-like particles (VLPs) or self-assembling scaffolds such as ferritin and encapsulin serve as such carriers, and presenting the conserved neutralising epitope with accurate geometry is key. Combined with the epitope focusing of section 8.2, this enables the design of a potent antigen presenting only the conserved epitope multivalently.

Platform application (maturity circle-double - full de novo generative pipeline): GC-SCAFFOLD takes the epitope backbone geometry as a motif and generates a self-assembling symmetric scaffold, GC-FOLD predicts the assembly and the antigen-presentation interface, and GC-DOCK evaluates presentation density and orientation. Because a newly designed scaffold can itself be a strong immunogen, GC-IMMUNO quantifies this risk by computing epitope density separately for the scaffold region and the epitope region (the same principle as section 8.2).

9. Customer Guide - What You Provide and What You Receive

This chapter is written so that decision-makers without a computational biology background can evaluate and manage a project. Technical terminology is minimised and the description follows the practical workflow.

9.1 Information Required to Start a Project

9.2 Deliverable Specification

At project completion the customer receives the following deliverables.

9.3 Schedule and Milestones

Table 9-1. Standard project schedule. Computation duration varies with target size and the number of evaluation axes.

The computation window is wide (2-6 weeks) because the later pipeline stages - molecular dynamics, free energy perturbation, quantum chemistry - consume large amounts of GPU time. These stages can be selectively reduced according to budget and schedule, in which case both predictive precision and elapsed time decrease together. This trade-off is presented explicitly during the feasibility scan.

9.4 How to Read the Ranking Table

The most frequent customer question is whether it suffices to test only the top-ranked candidate. The answer is no, and understanding why is central to using this platform correctly.

The composite score is a weighted sum across axes, and candidates with similar scores are not statistically distinguishable. If the score gap between rank 1 and rank 5 is smaller than the uncertainty on each candidate, they are effectively tied. The ranking table therefore reports uncertainty alongside the composite score, and candidate groups (tiers) that are significantly distinguishable are bundled automatically.

Diversity matters even more. If all ten top candidates arise from the same design strategy - say, hydrophobic core filling in the same region - then all ten fail if that strategy is wrong. Ten candidates drawn from different strategies, by contrast, distribute the risk. The ranking table clusters candidates by design strategy and recommends selecting evenly across clusters when committing to experiments. The logic is identical to portfolio theory.

9.5 Frequently Asked Questions

Q. What happens if a prediction is wrong?

Predictions are statistical statements and are frequently wrong for individual candidates. What we guarantee is a set-level property: that the success rate of the filtered group is significantly higher than random. Contracts specify the candidate count and the expected success-rate range; we do not guarantee individual candidates without a basis.

Q. Will our data be used to train models?

Customer data is used only to calibrate models for that project and is stored in isolation at project completion. Transfer to another customer's project is never performed without explicit written consent.

Q. Who owns the intellectual property in the designs?

By default the rights to designed sequences and inventions derived from them vest in the customer, and this is confirmed in the contract. The platform provides FTO review, but this is not legal advice and the final judgement rests with the customer's legal function.

Q. Can we use the computational results in a publication?

Yes. The reproducibility package includes a methods description and computation fingerprints, supporting both the writing of a Methods section and responses to reviewer reproduction requests. We provide standard wording that replaces internal module names with functional descriptions.

Q. Can we run our own tools in parallel or cross-validate the results?

We encourage it. Deliverables are in standard formats (FASTA, PDB, TSV) and can be fed directly into customer tooling. If cross-validation reveals discrepancies, we perform root cause analysis jointly.

Q. What is the minimum project unit?

The feasibility scan (3-5 days) is the minimum unit; it states the difficulty of the target and the expected performance range. If the scan is negative, we do not recommend proceeding to a full project.

10. Compliance Matrix: Publication, Open Data and Commercial Software Requirements

This chapter assesses item by item how far the platform meets requirements arising from academic publication, open data standards and commercial software procurement. Items not met are stated explicitly together with a remediation plan.

10.1 Academic Publication Requirements

Requirements imposed by major computational biology journals converge broadly on method reproducibility, data availability, code availability, statistical rigour and control design.

10.2 Open Database and Standards Compliance

10.3 Commercial Software Procurement Requirements

This is an assessment against items commonly required in enterprise procurement review, with items important in regulated (pharmaceutical and biotech) procurement placed first.

10.4 Third-Party Component Licence Management

The platform encapsulates backend components through its internal module contract structure. The licence management policy is as follows.

1. Every backend component is licence-reviewed before adoption, with 'use' restrictions evaluated separately from 'redistribution' restrictions. Academic-only licences are never used in commercial projects.

2. Installability is unrelated to permission to use. Even if a package installs, if its licence prohibits commercial use the component is excluded; this principle admits no exceptions.

3. Where capability from a commercially restricted component is genuinely needed, we either rebuild an in-house model from public data (data licences and code licences are distinct) or route around it via an external data service.

4. Licence status is subject to periodic re-review, and if terms change the backend of the affected module is replaced. Thanks to the module contract structure, such replacement does not affect the customer interface.

5. Customer deliverables and documents use internal module codenames only, so that backend composition information is not disclosed.

10.5 Remediation Plan for Unmet Items

11. Validation, Quality Control and Reproducibility

11.1 The Triple Validation Principle

Every platform deliverable passes three layers of validation. The system is designed to prevent the most dangerous failure mode in computational pipelines: silent failure. A silent failure is a case where the program terminates without error but the result is meaningless; because there is no error message, detection is extremely difficult.

1. Structural validation - is the deliverable formally correct? We check that sequences contain no non-standard characters, that structural coordinates have no missing atoms or clashes, and that numerical values lie in physically possible ranges.

2. Mechanism validation - did the computation actually execute the intended path? A plausible result is not evidence. We confirm from the logs that the computational stage genuinely ran and did not fall back to a default value or an alternative path.

3. Contextual validation - is the result consistent with other results? If a mutation reported as greatly stabilising appears destabilising in molecular dynamics, one of the two is wrong and investigation is required.

11.2 Regression Test System

The platform holds more than 500 automated regression tests, all of which run on any code change [internal]. Tests are layered into unit level (numerical accuracy of individual functions), integration level (integrity of data handoff between modules) and system level (a reduced-scale run of the full pipeline).

A mandatory pre-deployment step is the baseline diff: deliverables for identical inputs before and after the change are compared to confirm that nothing other than the intended change has moved. This matters because tests can genuinely cement incorrect behaviour as a contract - a situation in which fixing a bug causes a test written on the assumption of that bug to fail. Deciding whether to amend the test or revert the code then requires careful judgement.

11.3 Reproducibility Mechanisms

11.4 Prediction-Measurement Comparison Protocol

Once wet-lab data returns, a standardised comparison analysis is performed. For each evaluation axis the Spearman rank correlation between predicted and measured values is computed, and axes whose correlation is not significant are automatically down-weighted in the next round. At the same time, individual candidates whose predictions were badly wrong are analysed for common patterns - a process of checking for systematic bias in particular structural regions, mutation types or property ranges.

The results of this analysis are supplied to the customer as a round report, which states honestly which axes worked well for this target and which failed. Not hiding failed axes is the foundation of long-term trust, and it is also used directly to optimise resource allocation in the next round.

12. Appendix

12.1 Glossary

12.2 Deliverable File Specification Summary

12.3 Adoption Review Checklist

Items an organisation evaluating adoption of this platform should confirm internally.

[ ] Can we express our target's improvement objective as a quantitative figure (specificity at the level of 'Tm +8 °C')?

[ ] Can we provide the target protein sequence and, if possible, structural information?

[ ] Do we have existing variant experimental data? (If so, calibration is possible from round one.)

[ ] Do we have the capability or an outsourcing route for wet-lab validation?

[ ] Do we have budget and schedule for round two if round one misses the objective?

[ ] Has legal reviewed IP ownership of the designs and the scope of FTO responsibility?

[ ] Does the customer data handling policy (security, isolation, retention) meet our internal standards?

[ ] If we plan to use this in a regulatory submission, is the required documentation level (IQ/OQ/PQ etc.) defined?

[ ] Is the target free of membrane protein, disordered protein or PTM-dependent characteristics? (Prior consultation is needed if not.)

[ ] Have success and failure criteria been agreed in writing before contracting?

12.4 Immediate Next Steps

1. Prepare the target information package - assemble the amino acid sequence (FASTA), any available structure files, existing experimental data and quantified improvement objectives into a single archive.

2. Request a feasibility scan - submit the package for a 3-5 business day scan. This stage states the expected enrichment range and the principal risk factors.

3. Hold a scan review meeting - judge whether the stated expected performance range meets your business objectives. If it does not, stopping at this point is the rational choice.

4. Agree the design specification in writing - fix the objective function, success and failure criteria, deliverable scope, schedule and cost.

5. Start round one while planning wet-lab validation in parallel - prepare expression, purification and assay plans concurrently with computation so that experiments can begin the moment candidates are confirmed.

Unified Protein Engineering Platform Whitepaper v3.0