Spotlight on antibody developability

Monoclonal antibodies (mAbs) have become vital biotherapeutics for the treatment of cancer, autoimmunity and infectious diseases. The main reason behind their fame and success is their ability to precisely bind to their target and initiate a subsequent immune response. Thus, since the development of the first antibody drug, antibody therapeutics became one of the fastest-growing medicinal classes in the pharmaceutical field, with a forecasted global market value of more than $300 billion in 2025 (Figure 1).

However, developing antibody therapeutics is not a straightforward endeavor. For a mAb molecule to be successful, it requires both adequate binding affinity to its target and a favorable "developability" profile.

Developability is an umbrella term that has been used in scientific literature to describe the suitability of an antibody drug to be manufactured, stored and safely administered  to patients. The gold standard for measuring developability is to perform lab-based assays on antibody candidates to measure the values of their developability parameters (DPs). Nevertheless, recent advancements in deep learning and computational approaches have made it possible to estimate DP values at scale, starting from the sequence or the structure of the antibody as an input (Figure 2).

Knowledge gaps and how to fill them

Many studies have focused on extracting developability rules from the limited number of mAbs which made it into the clinical market, considering them as developability idols. But, clinically approved mAbs are very limited in number (100s). Also, lab-based developability measurements and experimental antibody structures are scarce. Additionally, there have been attempts to draw a hard line between natural and human-engineered antibodies in regards to developability. These factors have  hindered large-scale antibody developability studies to assess  the DP plasticity (redundancy, sensitivity and predictability). 

However, evidence suggests that human-engineered antibodies share high sequence similarity with native antibodies, and thus are not entirely dissimilar when compared to them.  Indeed, some researchers have argued for  leveraging the native antibody repertoire diversity to extract developability rules for antibody therapeutics design. Such an approach could leverage the higher abundance and biological compatibility of native antibodies (Figure 3). Additionally, with the increasing availability of native antibody sequences and the development of antibody-specific structure prediction tools, it is possible to assemble large developability datasets to assess DP plasticity. In our paper we assemble a large dataset of native antibodies (>2M Fv sequences), predict their structures and calculate their developability parameters. We leverage the scale of this dataset to conduct DP value plasticity analyses as well as study the effect of sequence similarity on developability profile similarity. 

Highlights and significance of the study 

Developability parameter redundancy may be reduced through analysis of intercorrelations

We show that the number of DPs required to focus on in antibody developability assessment may be summarized, using correlation network and redundancy reduction analyses. Specifically, we present the ABC-EDA algorithm as an effective way to determine minimal sets of DPs that are representative of overall developability. Using this algorithm, we report higher redundancy among sequence DPs compared to structure DPs.

Relevance of species- and chain-specific developability spaces for antibody discovery efforts

We show that the species of origin and the chain type of an antibody are key determinants of its developability. Furthermore, within the same chain type and species, we show that different antibody isotypes displayed similar developability trends, incentivising the exploration of the vast native antibody space for therapeutic purposes beyond their isotype annotation. 

Developability similarity is only loosely related to antibody sequence similarity

We show that even when a group of antibodies harbors high sequence similarity, their developability profiles are only loosely correlated. Such finding highlights the degrees of freedom for therapeutic antibody candidate engineering to optimize its developability with minimal changes introduced to its sequence.

Sequence-based DPs are more predictable than structure-based DPs 

Using a non-exhaustive machine learning algorithm (multiple linear regression) and a general-purpose protein language model (ESM-1v), we found that sequence DPs are more predictable than structure DPs. Within the scope of our analysis, we argue that the higher difficulty of predicting the values of structure-DPs is linked to/associated with the absence of effective representation of antibody structure and the lack of intercorrelations among structure DPs.

Human-engineered antibodies are contained within the natural developability landscape

We showed that human-engineered antibodies (including therapeutic mAbs) are localized within the developability space of natural antibodies. This further highlights the futility of attempting complete separation between natural and human-engineered antibodies. 

Structure-based DP values are reliant on the choice of structure prediction tool

Generally, we found that structure prediction tools only capture a snapshot of the antibody dynamic continuum. This results in tool-dependent disparities among calculated DP values, and stresses the ongoing need to integrate molecular dynamic simulations in developability studies.

Vision and what’s next? 

Computational developability profiling will be increasingly useful once the real-world relevance of computational DPs has been further established. Few issues can be improved or better addressed to increase the generalisability of computational antibody profiling. 

For instance, better availability of negative controls - representing the mAbs that fail to make it into the market due to developability hurdles - could be extremely helpful in establishing meaningful rules and screening tools for computational antibody developability. In general, further collaborations and data sharing between big pharma and academia would benefit the field. Additionally, more accurate and consistent prediction of structural DPs could aid streamlining in silico structural developability assessments. To this end, ML-aided molecular dynamic simulations could offer a high-throughput solution to this issue.