Skip to main content
NCBI home page
Briefings in Bioinformatics logoLink to Briefings in Bioinformatics
. 2026 Jun 11;27(3):bbag284. doi: 10.1093/bib/bbag284

An intrinsic sequence-structural profile for mRNA-delivered therapeutic antibodies

Farbod Mahmoudinobar 1,#, Grace Meng 2,#, J Wade Davis 3, Sandeep Kumar 4,✉
PMCID: PMC13376147  PMID: 42281399

Abstract

The mRNA technology represents a paradigm shift for the biopharmaceutical industry because in vivo delivery of antibody-based therapeutics can expand their clinical applications. Moreover, the mRNA delivery shifts the burden of assuring CMC (chemistry, manufacturing, and control) to the mRNA molecule. However, in vivo developability issues such as intracellular expression, proper protein folding, secretion into extracellular milieu, degradation in physiological conditions, poly-specificity, immunogenicity, safety, efficacy, and pharmacology can still pose significant barriers to the success of these products. Here, we have leveraged insights from the sequence—structural properties of the variable regions (Fvs) from the marketed biotherapeutics to build a medicine-likeness profile for the mRNA-delivered biotherapeutics. Our dataset consists of 122 unique Fvs from 117 marketed biotherapeutics (as of February 2024) and was systematically evaluated using 9 nonredundant sequence and structural parameters to capture a holistic view of developability attributes relevant to in vivo performance of mRNA-delivered antibody therapeutics. To ensure robustness, 25 antibodies that failed to reach approval due to developability issues, as well as those withdrawn from the market post approval, were included as controls. Our findings highlight a complementary relationship between sequence- and structure-based characteristics, leading us to develop a combined scoring system. This integrated “medicine-likeness” profile enables early-stage in silico assessments of mRNA-delivered therapeutic antibody candidates, offering a valuable tool for researchers to predict and optimize their in vivo developability.

Keywords: mRNA, antibody, developability, drug discovery, computational biophysics

Graphical Abstract

Graphical Abstract.

Descriptors calculated from sequence and structure of antibodies are use to generate a developability score.

Open in a new tab

Introduction

Antibody-based biotherapeutics such as monoclonal antibodies [1], antibody drug conjugates (ADCs), VHHs (Variable Heavy Domain of Heavy-chain Antibodies), bi-specific/multi-specific antibodies and more, continue to be major avenues for developing treatment options in many therapeutic areas including oncology, infectious diseases, neurodegenerative diseases, and other autoimmune disorders. Since the approval of the first antibody for human use in 1986 (muromonab [2]), the number of marketed and clinical stage therapeutic antibodies has grown rapidly. Despite improved industrialization of recombinantly produced biotherapeutics over the past several decades, many challenges remain in developing antibody-based biotherapeutics. Many such drug candidates can be identified, but very few of them are able to succeed in drug product development and clinical trials to become approved medications [3], due to various challenges associated with their product and clinical development.

Several methods are currently used to generate antibodies against a desired antigen. These methods include animal immunization, hybridoma [4], B-cell sequencing [5], phage [6], and yeast display [7]. With the recent improvements in artificial intelligence and machine learning, as well as the emergence of publicly available antibody sequence resources such as Observed Antibody Space (OAS, https://opig.stats.ox.ac.uk/webapps/oas/) [8, 9] and Adaptive Immune Receptor Repertoire Community (AIRR, https://www.antibodysociety.org/the-airr-community/) [10], it has become feasible to generate libraries of fully human antibodies in silico, in both an antigen-specific and antigen-independent manner [11–13]. Concomitant with the advent of in silico methods that can potentially help alleviate many hurdles to the discovery and development of recombinantly produced antibody-based biotherapeutics, there is another exciting innovation that can further expand the use of antibody-based formats as therapeutic agents—the emergence of mRNA technology and its ability to deliver the information to produce biotherapeutics in vivo.

Developability is a major concept in the discovery and development of recombinantly produced antibodies [14]. Developability is the ability to translate newly discovered therapeutic antibody drug candidates into regulatorily approved therapeutic drug products available in the market. Developability goes beyond the CMC (chemistry, manufacturing, and control), formulation, storage, shipping, delivery devices, routes of administration, and other such in vitro considerations. It also includes in vivo aspects such as safety, efficacy, and pharmacology. In other words, it pertains to both in vitro ease of product development and in vivo performance during clinical trials, approvals, and beyond. In the case of the recombinantly produced therapeutic antibodies, both the drug substance and the active ingredient in the drug product are the same biological entity, namely, a protein. In case of mRNA-delivered biotherapeutics, the drug substance is an mRNA. Upon delivery, it produces the drug product in vivo, which is an antibody. Hence, the drug substance and the drug product are now different biological entities. Therefore, in vivo developability is separated from the in vitro developability for mRNA-delivered therapeutic antibodies. The in vitro developability considerations are relevant to the mRNA molecule, while the in vivo developability considerations such as intracellular expression, folding, secretion into extracellular spaces, stability under physiological conditions, off-target binding, pharmacology, immunogenicity, safety, and efficacy are relevant to the therapeutic antibody molecule. It is important to incorporate considerations of the in vivo developability in the discovery and design of mRNA-delivered biotherapeutics.

In this work, we asked how we can combine calculated antibody sequence and structural features indicative of developability attributes to design more effective mRNA-delivered therapeutic antibodies. We describe a statistical profile, constructed using the nonredundant sequence-structural descriptors derived from the unique variable regions (Fvs) found in Food and Drug Administration (FDA)-approved biotherapeutics currently available in the market. No approved mRNA-delivered therapeutic antibodies are available in the market yet. However, they are in multiple ongoing clinical trials [15]. Therefore, this profile is based on recombinantly produced therapeutic antibodies. However, we have accounted for considerations that are unique to mRNA delivery of therapeutic antibodies and their behavior under the physiological conditions. This profile should be treated as an early statistical framework for assessing in vivo performance of the mRNA-delivered antibody-based biotherapeutics, till further information on in vivo characteristics of mRNA-delivered therapeutic antibodies becomes available in the public domain. This profile will serve as a valuable computational tool for discovery and design of mRNA-delivered therapeutic antibodies.

Results

Dataset curation and germline coverage of marketed antibodies

In this work, we have studied sequence and structural properties of 117 FDA-approved antibody-based biotherapeutics available in the market as of February 2024. Of these 117 biotherapeutics, 10 are bispecific antibodies, and they contain a total of 122 unique variable regions (Fvs) as shown in Table S1a in the Supporting Information (SI). Nearly all (117, 96%) of the 122 Fvs are from full-length monoclonal antibodies and antibody drug conjugates, with a smaller number of Fabs (fragment antigen-binding regions—certolizumab [16, 17] and idarucizumab [18, 19]), and scFvs (single chain variable fragments), blinatumomab [a bispecific T-cell engager or BITE (bispecific T-cell engager) [20, 21]], and crizanlizumab [22, 23]. Table S1b shows the light and heavy chain isotypes of the full-length antibodies and Fabs. A total of 109 (89.3%) of these 122 Fvs contain kappa light chains and the remaining 13 (10.7%) contain lambda light chains. Heavy chains Isotypes could be determined only for the 117 full-length mAbs. Most of them are of IgG1 (75, 64.1%) and IgG4 (26, 22.2%), with only 13 (11.1%) being of IgG2 isotype. In the case of three biotherapeutics, the heavy chains are hybrids of multiple isotypes (IgG1-IgG2 for tafasitamab and IgG2-IgG4 for eculizumab and ravulizumab). These observations are consistent with those of Kumar et al. [24]. The classification of each Fv as murine, chimeric (murine variable regions with human constant regions), humanized (only the CDRs (complementarity-determining regions) are of murine origin), or fully human was determined from the World Health Organization (WHO) INN recommended lists (https://www.who.int/teams/health-product-and-policy-standards/inn/inn-lists) and confirmed with the IMGT [25] database. While many early antibody drugs were murine or chimeric, most current therapeutics are either humanized or fully human (Table 1). There is currently only one fully murine antibody in use, the bispecific T-cell engager blinatumomab, and chimeric antibodies constitute only 13% of marketed therapeutics. Six antibodies also contain one humanized chain while the other one is chimeric (Table S1c).

Table 1.

Germline pairs with ≥5 incidences in the marketed antibody biotherapeutics.

Heavy chain germline family Light chain germline family Count
IGHV3 IGKV1 29 (23.8%)
IGHV1 IGKV1 21 (17.2%)
IGHV3 IGKV3 15 (12.3%)
IGHV1 IGKV3 9 (7.4%)
IGHV4 IGKV1 5 (4.1%)
IGHV1 IGKV2 5 (4.1%)
Open in a new tab

Figure S1 in the Supplementary Information shows V-gene usage of the heavy and light chains found in the 122 variable regions, respectively. We observed that heavy chains in the marketed antibodies are predominantly from IGHV3 and IGHV1 germline loci, and most light chains are from IGKV1 and IGKV3 germline loci (Table 1).

Descriptor framework: definition, robustness, and statistical independence

To quantitatively characterize the 122 Fvs from the marketed antibodies in terms of their intrinsic sequence and structural attributes, we identified 9 nonredundant descriptors that describe different aspects of biotherapeutic developability considerations [26]. The details of these descriptors can be found in the Methods section. Briefly, four of these descriptors are sequence-based, derived using only the amino acid sequences of the FVs. These are percent humanness, Aggregation propensity (AP) normalized by the sequence length), chemical liabilities score, and fraction of an Fv’s sequence covered by MHC II binding potential of T-cell immune epitopes normalized by sequence length (IM). The remaining five structure-based descriptors, namely, BSAVH-VL (surface area buried between the light and heavy chains), pIFV_3D (structure-based isoelectric point for the Fv), HI (hydrophobic imbalance), RM (ratio of dipole to hydrophobic moments), and RP (ratio of patches), were computed using the structural models of the 122 Fv regions. The above five descriptors have been previously described by Kumar et al. [24, 27]. These descriptors capture information about electrostatics, conformational stability, and molecular surface properties of the Fvs. Because this profile is focused on mRNA delivery, we consider only the physiological condition (pH 7.4, 150 mM salt) in this report. Taken together, these nine computational descriptors refer to conformational stability, aggregation, immunogenicity, chemical degradations, hydrophobicity, and electrostatic properties of the therapeutic antibodies under physiological conditions. These properties can play important roles towards intracellular expression, proper folding, and secretion of therapeutic antibodies in the extracellular milieu along with their safety, efficacy, and pharmacology.

To understand how the choice of molecular modeling methods and software could potentially affect the values of the structure-based descriptors, the variable regions of the 122 Fvs were constructed using 2 different protocols. One of them is a homology-based method as implemented in MOE from Chemical Computing Group [28] and the other one is a deep learning based, ABodyBuilder [29]. Furthermore, the homology models generated using MOE were used as starting structures for molecular dynamics simulations to generate conformer ensembles for each Fv. Table 2 shows the statistical data comparing the five structure-based descriptors. Average values for the five structure-based descriptors are consistent across all three approaches, highlighting the robustness of our methodology. Figure 1 shows the correlation plots of the individual descriptors between the conformation ensembles from MD versus MOE energy minimized structures. While the Pearson correlation coefficient (R2) is lower for the more fluctuating variables (BSA and RP), the low P-values for all parameters demonstrate that the correlations are statistically significant. This observation implies that well-prepared static homology models are sufficient for large-scale early-stage screening of antibodies.

Table 2.

Values of the five nonredundant structure-based descriptors computed using homology-based method (MOE), MD conformer ensembles, and deep learning-based method (ABodyBuilder) of the 122 Fv regions from marketed antibody-based biotherapeutics.

Descriptor MOE energy minimized models (mean ± SD) [Min – Max] Conformer ensembles (mean ± SD) [Min – Max] ABodyBuilder (mean ± SD) [Min – Max]
Surface area buried between VH and VL domains of Fv (BSAVH-VL) (Å2) (810 ± 80) [570–1117] (813 ± 71) [648–1005] (793 ± 72)
[600–948]
Isoelectric Point of Fv (pIFV_3D) (7.6 ± 1.2) [4.2–9.5] (7.5 ± 1.3) [4.3–9.4] (7.6 ± 1.3) [4.2–9.5]
Hydrophobic Imbalance of Fv (HI) (0.92 ± 0.36) [0.23–1.76] (0.95 ± 0.39) [0.33–2.1] (0.90 ± 0.35) [0.17–1.84]
Ratio of Moments (RM) (1.04 ± 0.67) [0.04–5.11] (1.08 ± 0.74) [0.22–5.82] (0.97 ± 0.57) [0.18–3.43]
Ratio of charged and hydrophobic patches (RP) (2.05 ± 1.32) [0.75–13.56] (1.96 ± 0.80) [0.78–5.30] (1.64 ± 0.73) [0.44–5.47]
Open in a new tab

Figure 1.

Scatter plots comparing homology models and MD ensembles descriptors.

Open in a new tab

Scatter plots of structural descriptors for MOE-generated homology models and MD conformational ensembles. Linear best fit is shown in dashed line, slope, Pearson correlation co-efficient, R2, and P-value for each descriptor is reported on each plot. Note that all descriptors show statistically significant correlations (P-values < 0.05) between the descriptors computed using the homology models and the conformer ensembles from MD simulations.

In the next step of profiling sequence and structural attributes of the 122 variable regions from the marketed antibodies, we asked whether the 9 descriptors chosen in this study are correlated with each other. Figure S2 shows pairwise Pearson’s correlation coefficients for all nine descriptors. All descriptors exhibited low correlation coefficients with one another, where pairs showing |R2| ≤ 0.18 (R ≤ 0.42) are considered non-correlated [30] (see Methods section). Hence, these nine descriptors capture different sequence and structural aspects of the Fvs found among the marketed antibody therapeutics.

Construction of the medicine-likeness profile and classification of antibodies

Having established non-redundant descriptors, we next constructed an integrated scoring framework to assess in vivo developability of mRNA-delivered therapeutic antibodies based on the similarity of their intrinsic sequence-structural attributes to those of the recombinantly produced and FDA-approved therapeutic antibodies. The distribution of values for each descriptor was normalized by computing the Z-scores and interpreting them according to the scheme described in the Methods section (see Fig. 2). All Z-score values >3 and those <−3 were reassigned to 3 and −3, respectively, to mitigate the impact of outliers. For each of the 122 marketed Fvs, the Z-scores for its sequence and structure-based descriptors were summed to yield the respective sequence and structure scores. In the next step, these scores were combined to yield a single metric, called total medicine-likeness score.

Figure 2.

Flowchart for medicine-likeness profiling of antibody Fvs.

Open in a new tab

Flowchart diagram showing the calculation process of medicine-likeness profile for 122 Fvs from marketed antibody based biotherapeutics.

Figure 3 shows the distributions of the five structure-based descriptors of the marketed antibodies. Most of the marketed antibodies have their buried surface area between heavy and light chains (BSAVH-VL) within the range of 700 to 900 Å2 showing a normal distribution. This descriptor indicates conformational stability of an Fv with higher values corresponding to more stable Fv. The structure-based isoelectric point of the 122 Fvs (pIFv_3D) from approved therapeutic antibodies shows a wider distribution where majority of the Fvs have pI around 8 and few with values above 9 or below 6. Hydrophobic imbalance (HI) quantifies asymmetry in the surface distribution of hydrophobic residues. HI values range from 0.25 to 1.75, with most clustering near 1.0 and a small subset exhibiting higher imbalance for the 122 Fvs analyzed here. Ratio of electrostatic dipole moment to hydrophobic moment (RM) is a measure of imbalance between electrostatic and hydrophobic polarities of the antibody Fvs. In the 122 Fvs, the average value of RM is nearly 1, suggesting that the 2 parameters are nearly balanced, with a minority of molecules exhibiting elevated RM values. The ratio of patches (RP) parameter measures the ratio of the summed surface area of positively and negatively charged patches to that of hydrophobic patches and shows marketed antibodies predominantly possess larger electrostatic patches, with a normal distribution centered on 2 ± 1.3.

Figure 3.

Distributions of five structure-based antibody descriptors.

Open in a new tab

Distributions of the five structure-based descriptors used in developing this profile. The black dashed line shows the median value for each descriptor, green lines show mean value, and red solid lines represent 25 and 75 percentile limits. The structure-based descriptors include, (a) the buried surface area between the heavy and light chains (BSAVH-VL), (b) structure-based isoelectric point (pIFv_3D), (c) HI, (d) ratio of dipole moment to hydrophobic moment (RM), and (e) ratio of charged patch area to hydrophobic patch area (RP).

Figure 4 presents the distributions of the 4 sequence-based descriptors for 122 Fvs from the marketed antibodies. Percent Humanness captures sequence similarity of the marketed antibodies to human ones. The data clusters mainly around two values, ~85% and ~95%, which mostly correspond to humanized antibodies and human-derived ones, respectively. Chemical liabilities show the sum of penalties associated with each chemical degradation sequence motif (details in the Methods section in SI). This is an in-house developed, literature-driven scheme of weights to assess the vulnerability of an Fv to chemical degradation. It does not indicate the chemical degradation rate, rather it measures the intrinsic vulnerability of an Fv to chemical degradation because of incidence of various chemical liability motifs in its amino acid sequence. The chemical liability penalties for the Fv regions found in the marketed antibodies average at 195 ± 62, with some Fvs exhibiting high-liability outliers. The normalized aggregation propensity (AP), calculated by TANGO [31] the tendency of antibodies to self-associate via β-strand  mediated aggregation with higher scores indicating greater susceptibility. The average AP for 122 Fvs is 4.3 ± 2.3 per residue with a majority of them being between 2 and 6. The immunogenicity score (IM) is portion of an Fv's sequence covered by potential MHC II T-cell epitopes (explained in detail in SI). The T-cell epitopes binding potentials were predicted by NetMHCPanII [32]. The average IM in this dataset is 0.14 ± 0.04.

Figure 4.

Distributions of four sequence-based antibody descriptors.

Open in a new tab

Distributions of the four sequence-based descriptors used in this profile. The black dashed line shows the median value for each descriptor, green lines show mean value, and red solid lines represent 25 and 75 percentile limits. The sequence-based descriptors include, (a) average percent humanness, (b) chemical liability score (CL), (c) aggregation propensity (AP), and (d) predicted immunogenicity score (IM).

Figure 5 shows the sequence, structure, and total medicine-likeness scores for the 122 Fvs from the marketed antibodies. The mean value for the sequence score of these Fvs is nearly zero with the range being −7.5 to +4 (Table 3, Fig. 5a). The observed mean for the structure score is −1.9 ± 1.6 (range = −7.5 to +1.6, Table 3, Fig. 5b) and majority of the Fvs from marketed antibodies scored in the range of −5 to +2, with only a few outliers scoring below −5, indicating that the Fvs found among the marketed antibodies demonstrate a diverse range of biophysical behaviors. This observation is consistent with the experimental observations made by Tessier et al. [33]. The sum of the sequence and structure scores—the total medicine-likeness score—for the marketed antibodies is shown in Fig. 5c. It has a mean value of −1.9 ± 2.8 (range = −10.3 to +4.0; Table 3).

Figure 5.

Distributions of sequence, structure, and medicine-likeness scores.

Open in a new tab

Distributions of the combined sequence and structure scores. (a) Sequence score, (b) structure score, and (c) total medicine-likeness score. The black dashed line shows the median value for each descriptor, green lines show mean value, and red solid lines represent 25 and 75 percentile limits.

Table 3.

Values of structure, sequence, and total medicine-likeness scores for MOE-generated homology models, the conformational ensemble from MD, and models generated by ABodyBuilder.

Descriptor MOE energy minimized models (mean ± SD) [Min – Max] Conformer ensembles (mean ± SD) [Min – Max] ABodyBuilder (mean ± SD) [Min – Max]
Structure score (−1.89 ± 1.63) [−7.49–1.56] (−1.79 ± 1.77) [−6.59–1.68] (−1.94 ± 1.84) [−8.55–2.02]
Sequence score (0.01 ± 2.10) [−7.48–4.06] (0.01 ± 2.10) [−7.48–4.06] (0.01 ± 2.10) [−7.48–4.06]
Total medicine-likeness score (−1.90 ± 2.83) [−10.30–4.01] (−1.78 ± 2.73) [−8.96–5.25] (−1.93 ± 2.78) [−8.83–3.72]
Open in a new tab

Figure 6a plots the sequence score versus structure scores for the 122 variable regions from the marketed antibodies. It allows us to classify the 122 Fvs into different quadrants and improve our understanding of the relationship between the sequence and structural attributes of the marketed antibodies. The upper-right quadrant (class 1) represents 33 out of 122 (27%) Fvs with above average sequence as well as structure scores. The upper-left quadrant (class 2) again contains 33 of the 122 (27%) Fvs with below average sequence scores but above average structure scores. Such antibodies may possess less desirable sequence attributes, but this does not seem to impact their structural properties. Taken together, these two classes contain a majority (66, 54%) of the 122 Fvs from the marketed antibodies. The lower-right quadrant (class 3) contains 24 (~20%) of the 122 Fvs that possess above average sequence scores, but below average structure scores. In this quadrant, Fvs with good sequence scores may still demonstrate structure-based issues and, therefore, less desirable in vivo performance. Finally, the lower-left quadrant represents class 4 where the 32 out of the 122 (26%) Fvs possess below average sequence, as well as structure scores. These are the least optimal Fvs from the perspective of in vivo performance.

Figure 6.

Scatter plots classifying antibody Fv medicine-likeness scores.

Open in a new tab

Scatter plots of sequence and structure scores across four quadrant classes for the Fvs from (a) marketed and (b) control antibodies. Dashed lines show mean value for structure and sequence scores. The control antibodies include discontinued (circles) and withdrawn (triangles) antibodies. Class stratified total medicine likeness scores for Fvs from all (marketed plus control) antibodies are shown in the panel (c).

In addition to the 122 Fvs from the marketed antibody-based biotherapeutics, we have also collected from literature and modeled 25 Fvs from antibodies that are currently not in the market, to serve as controls in this study. The dataset of 25 control antibodies includes the following: (i) antibodies whose development was discontinued due to insurmountable developability hurdles, (ii) antibodies which were approved but were later withdrawn from the market, and (iii) antibodies shown to have high self-association and high nonspecific binding in experiments performed by Tessier group [33]. In some cases, mutants of discontinued antibodies were later found to possess improved physicochemical properties. Figure 6b shows that the sequence and structure scores for most (20 out of 25 Fvs, 80%) of them fall into either class 3 (7 Fvs, 28%) or 4 (13 Fvs, 52%). Only 5 out of these 25 (20%) Fvs fall in classes 1 (2 Fvs, 8%) and 2 (3 Fvs, 12%), indicating their discontinuation or withdrawal may have been due to factors other than intrinsic sequence and structural attributes. Finally, Fig. 6c presents distributions of the total medicine-likeness score across all 4 classes for the 122 Fvs from marketed antibodies and the 25 ones from the control antibodies. These trends observed in this figure further support the utility of this sequence versus structural scores classification framework.

Discussion

Within only a few decades, antibody-based biotherapeutics have made a major impact on medicine by serving myriads of previously unmet medical needs. However, development of novel antibody-based biotherapeutics via recombinant DNA technology remains a complex, costly, and time-consuming process with developability of the newly discovered drug candidates being a major consideration [34, 35]. The advent of mRNA technology represents a paradigm shift for the discovery, development, and delivery of antibody-based biotherapeutics. The mRNA technology delivers the “information” to produce biotherapeutics within the human body. This information enables cells in the human body to produce and secrete the biotherapeutic into the extracellular milieu and fulfill its therapeutic purpose. Does developability, which is a major concern for the recombinantly produced antibody-based biotherapeutics, remain relevant to the mRNA-delivered ones? The answer to this question is undeniably, yes in our minds. This is because developability refers to the ability of a newly discovered biotherapeutic “drug candidate” to overcome “all” the hurdles to its product and clinical development, win required regulatory approvals and become a biotherapeutic “drug product” available in the market. Therefore, developability involves both in vitro and in vivo performance for biotherapeutic drug candidates. In the case of mRNA-delivered therapeutic antibodies, developability remains an important consideration for their in vivo performance. However, the onus for assuring the in vitro developability is now shifted to the mRNA molecules that encode the therapeutic antibody drug products for in vivo delivery.

At present, there are no approved mRNA-delivered therapeutic antibodies available in the market. However, several clinical trials involving these modalities are currently ongoing. Therefore, it is important to develop an initial medicine-likeness profile for the mRNA-delivered therapeutic antibodies proactively now rather than retrospectively. Note that the recombinantly produced therapeutic antibodies started to become available in the market around the late 1980s. However, the concept of developability did not gain traction until the 2010s and the first profile to computationally assess developability of antibody drug candidates, the Developability Index [36] was published by Trout group in 2012. We do not have to wait another three decades in case of mRNA-delivered antibody therapeutics. Much can be already learned towards this goal from the growing literature on the contribution of intrinsic physicochemical attributes of the recombinantly produced therapeutic antibodies towards their in vivo performance. For example, charge, hydrophobicity, and pI of an antibody drug play significant roles in determining its Pharmacokinetics/Pharmacodynamics (PK/PD) and rate of clearance [37]. With these thoughts in mind, we have developed a medicine-likeness profile for the in vivo performance of mRNA-delivered therapeutic antibodies based on intrinsic sequence and structural descriptors computed using the Fvs of the recombinantly produced ones currently in the market. This report is a proactive attempt to establish a medicine-likeness profile in anticipation of the first approval(s) of mRNA-delivered antibody therapeutics. This profile should be viewed as an initial model rather than a fully validated one. Note that we use the word “Medicine-likeness” purely in a physicochemical sense and do not imply any functional connotations.

To construct the above-mentioned profile, we have used four sequence-based descriptors in addition to using the five structure-based ones, we have described earlier [11]. This is a major improvement upon our previous works [11, 12, 24, 27] as well as that of others [36, 38], which have relied mostly on the structural attributes. The exploration of the sequence scores versus structural scores for the medicine-likeness is unique to this developability analysis. Balance between these two sets of descriptors could be important for differentiating antibodies that are likely to sail through product and clinical development from those likely to face challenges as shown in Fig. 6. The earlier versions of our medicine-likeness profile for the recombinantly produced antibodies involved calculations at three different environmental conditions, namely, physiological, formulation, and low pH hold for the viral inactivation step during manufacturing. The current version reported here involves only the physiological condition (pH 7.4 and ionic strength 150 mM).

In this initial version of the medicine-likeness profile for the mRNA-delivered therapeutic antibodies, we have intentionally used equal weighting for all nine descriptors. Putting weights based on the relative importance of these descriptors is desirable from a data science perspective. However, there are many caveats to this. First, developability is multifactorial in nature, and there is no single experimental measure that can tell us whether a given biotherapeutic drug candidate is developable or not. Second, we are still in initial stages of understanding developability, and there is very little experimental data in the public domain even for the recombinantly produced therapeutic antibodies. Third, our community tracks only the therapeutic antibodies that win approval from the regulatory authorities, and there is no systematic way to report and track ones whose development is discontinued, or those that do win approvals but are withdrawn from the market later on. We attempted to build a dataset of such antibodies via literature surveys and use them as a control set. However, it is by no means an exhaustive database. Fourth, a decision to discontinue the development of a therapeutic antibody in the clinical phases is often based on multidimensional risk versus benefit analyses performed by the regulators on case by cases basis. Given these limitations, any assumptions on the relative importance of different descriptors towards the in vivo developability of the mRNA-delivered antibody therapeutics will be arbitrary.

Here, we have focused on using nonredundant descriptors that may be related to various aspects of the in vivo developability of mRNA-delivered therapeutic antibodies. This profile does not claim a causal relationship between the intrinsic sequence-structure characteristics and regulatory approval of the antibody drug candidates. However, it is encouraging to find that the antibodies with literature reports of discontinued development and the withdrawn ones occupy less favorable regions of our medicine-likeness profile (Fig. 6).

The current profile studies only the Fvs of the marketed antibodies. However, this approach can potentially be expanded to study full-length antibodies. Another opportunity to further improve this profile is the chemical liabilities. Currently, only well-studied chemical degradation motifs are used for this calculation. These motifs can degrade an antibody both in vivo and in vitro conditions. However, several additional chemical degradations such as phosphorylation, glycation, citrullination, and so on, can also occur in vivo [39]. These liabilities were not studied in this work because in vivo chemical degradation of biotherapeutics has not been extensively studied yet in literature. Despite these limitations, the medicine-likeness profile shall be a valuable tool for discovering and developing mRNA-delivered antibody-based biotherapeutics.

Materials and methods

The schematics of the methods followed in this report are shown in Fig. 2 and complete details for the methods followed in this work are provided in the Supporting Information (SI) accompanying this manuscript. Here, we provide brief descriptions.

Collection and validation of antibody sequences

We curated a dataset of antibody-based biotherapeutics approved by the FDA as of February 2024. Variable-region amino acid sequences were primarily obtained from WHO INN lists, and missing older sequences were retrieved from the FDA GSRS database. All sequences were cross-validated against Thera-SAbDab and FDA GSRS databases. Bispecific antibodies contribute two Fv regions to our dataset. Only unique Fv sequences were retained to yield a nonredundant set of 122 Fv sequences.

Antibody modeling and structure-based descriptor calculations

Homology-based structural models of the 122 Fv regions of the marketed antibody-based biotherapeutics were computed using the Antibody Modeler module in MOE (version 2022.02) software, followed by a protonation and energy minimization step. Five structure-based descriptors were then computed to capture stability, electrostatics, and surface properties, namely VH-VL buried surface area (BSAVH-VL), structure-based isoelectric point (pIFv_3D), HI, ratio of moments (RM; electrostatic dipole/hydrophobic moment), and patch ratio (RP; charged/hydrophobic patch area).

Molecular dynamics simulations and ensemble descriptors

To account for conformational flexibility, we performed all-atom molecular dynamics (MD) simulations for each marketed Fv using GROMACS 2023.2 with the AMBER99SB-ILDN force field and TIP3P water. Each Fv was solvated in water to ionic strength of 150 mM NaCl, equilibrated under NVT and NPT, and simulated for 200 ns. Trajectory convergence was assessed via Cα RMSD (ΔRMSD <0.2 nm over the last 50 ns). For all Fv constructs, the light and heavy chain C-termini were capped to account for truncation relative to constant domains. After discarding the first 20 ns, conformational snapshots were extracted every 0.5 ns from the MD production runs. These snapshots were energy-minimized in MOE, and structural descriptors were computed and averaged to yield ensemble values for the respective structure-based descriptors.

Sequence-based descriptor calculations

Four sequence-based descriptors were computed:

  • (i) Percent humanness was defined as the average identity of heavy- and light-chain variable genes to their respective nearest human germlines as determined by IgBLAST [40].

  • (ii) Aggregation propensity was computed using TANGO [41] at pH 7.4 and 150 mM salt and was normalized by sequence length.

  • (iii) Predicted MHC class II epitope content (intrinsic immunogenicity potential) was computed using NetMHCIIpan [32] from IEDB (www.iedb.org) across 27 representative HLA-II alleles. Immunogenicity potential was defined as the fraction of residues within promiscuous predicted epitopes meeting the specified binding criteria.

  • (iv) Chemical liabilities were scored using a sequence-based penalty scheme restricted to motifs susceptible to degradation at the physiological pH (e.g. glycosylation, non-canonical cysteines, deamidation, oxidation, and fragmentation), with positional weighting based on whether motifs occur in framework regions versus CDRs and within CDR core versus termini.

Normalization of sequence and structure-based descriptors

Raw descriptor values were converted to Z-scores. Z-scores with values >|3| were capped at ±3 to limit outlier bias. Sequence and structure scores were computed as sums of the four sequence-based and five structure-based transformed Z-scores, respectively, using descriptor-specific sign conventions reflecting whether higher, lower, or deviation-from-mean values are favorable to developability. The total medicine-likeness score was computed as the sum of sequence and structure scores.

Key Points

  • The ability to deliver antibody-based biotherapeutics via mRNA will open new avenues for their clinical applications.

  • Developability remains relevant to mRNA-delivered antibody-based biotherapeutics.

  • We have developed a computational profile to assess the in vivo developability of mRNA-delivered antibodies by learning from sequence and structural properties of the variable regions found in the recombinantly produced Food and Drug Administration-approved therapeutic antibodies that are currently available in the market.

  • The availability of this profile has implications towards the discovery and design of mRNA-delivered antibodies.

Supplementary Material

SI_Medicine_LIkeness_Revision2_bbag284
si_medicine_likeness_revision2_bbag284.pdf (1.1MB, pdf)

Acknowledgements

Authors acknowledge helpful discussions with numerous Moderna colleagues on the topics of developability and mRNA delivery of biotherapeutics.

Contributor Information

Farbod Mahmoudinobar, Molecular Design and Modeling Group, Computational Sciences, Moderna, Inc., 325 Binney Street, Cambridge, MA 02142, United States.

Grace Meng, Molecular Design and Modeling Group, Computational Sciences, Moderna, Inc., 325 Binney Street, Cambridge, MA 02142, United States.

J Wade Davis, Molecular Design and Modeling Group, Computational Sciences, Moderna, Inc., 325 Binney Street, Cambridge, MA 02142, United States.

Sandeep Kumar, Molecular Design and Modeling Group, Computational Sciences, Moderna, Inc., 325 Binney Street, Cambridge, MA 02142, United States.

Author contributions

SK and FM conceived the research project. GM and FM performed the data curation and analysis. All authors contributed towards the manuscript writing.

Competing interest

Farbod Mahmoudinobar, Grace Meng, Wade J. Davis, and Sandeep Kumar are current or former employees of Moderna, Inc.

Funding

This study was supported by Moderna, Inc.

Data availability

All the sequences used in this study were obtained from publicly available sources, including WHO INN lists, FDA GSRS and TheraSabDab and are cited within the manuscript, no proprietary sequences were used. The computational descriptors are calculated using established and publicly available tools (IgBlast, TANGO, NetMHCIIpan, and Gromacs), and structural modeling was performed using MOE as detailed in SI section. The scripts and analyses workflows used to assemble the dataset, compute descriptors, normalization, and score generation are all explained in detail in SI and can be easily reproduced. The scripts and other material related with this work can be accessed from https://github.com/modernatx/med-lp.

References

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

SI_Medicine_LIkeness_Revision2_bbag284
si_medicine_likeness_revision2_bbag284.pdf (1.1MB, pdf)

Data Availability Statement

All the sequences used in this study were obtained from publicly available sources, including WHO INN lists, FDA GSRS and TheraSabDab and are cited within the manuscript, no proprietary sequences were used. The computational descriptors are calculated using established and publicly available tools (IgBlast, TANGO, NetMHCIIpan, and Gromacs), and structural modeling was performed using MOE as detailed in SI section. The scripts and analyses workflows used to assemble the dataset, compute descriptors, normalization, and score generation are all explained in detail in SI and can be easily reproduced. The scripts and other material related with this work can be accessed from https://github.com/modernatx/med-lp.

Articles from Briefings in Bioinformatics are provided here courtesy of Oxford University Press

RESOURCES