Building the Next Generation of Biologics: Inside the Future of Protein Engineering

Antibodies have led the biologics revolution, with more than 160 monoclonal antibodies approved by the FDA, making up nearly half of the top 20 drugs by market share. But the protein engineering field is rapidly evolving beyond its familiar Y-shaped form.

More than 200 antibody-drug conjugates (ADCs) are now in clinical stages, and bispecific antibodies have surged to 19 global approvals with total sales surpassing $12 billion in 2024. With the FDA increasingly supporting novel modalities in 2025, protein engineering is entering an era where bispecifics, ADCs, and novel binders routinely reach the clinic.

This next wave of biologics brings exceptional structural diversity and a new kind of challenge. For many discovery teams, the bottleneck has moved from imagination to implementation. It’s no longer just about identifying the right target but also finding practical ways to produce and scale molecules that often defy convention.

As these modalities proliferate, discovery groups are turning to partners that can integrate protein engineering, structure–function analysis, and early manufacturability. Viva Biotech, a globally integrated CRO specializing in complex biologics and structural biology, reports rising demand for bispecifics, ADCs, and other next-generation formats, including an increasing number of novel antibody and conjugate programs entering early development.

And as those pressures intensify, early discovery teams are confronting an increasingly familiar obstacle: molecular complexity.

“The biggest problem with these biologics is really the heterogeneity in actually generating these molecules,” said Paul Wan, vice president of early discovery and business development at Viva Biotech. “You’re getting mispaired chains in bispecifics or trispecifics and also different conjugation variants when you’re looking at ADCs.”

Variable drug-to-antibody ratios (DARs) in ADCs have been a persistent cause of clinical delays, a recurring theme in discussions at the American Association for Cancer Research meeting this year.

The field has responded with platform-level engineering approaches, Wan noted. “These can actually standardize correct chain forming or chain pairing for bispecifics, for example. Protein scientists use techniques like controlled Fab-arm exchange, knobs-into-holes, which is quite a traditional method, and crossMabs, and also using site-specific mutations to push the correct formation of these chains.”

For ADCs, precise conjugation methods now enable improved DAR control. The incorporation of non-natural amino acids adds greater chemical flexibility and expands the range of possible conjugations.

Meanwhile, AI and machine learning are increasingly used to prescreen for downstream developability issues, allowing teams to avoid aggregation, glycosylation, and turbidity hotspots before committing resources to manufacturing scale-up.

Parallel multistep purification now enables teams to express and screen hundreds of constructs simultaneously, with inline quality-control tools confirming construct quality in real time. ADC industry analysis for H1 2025 revealed a 30% spike in adoption of novel formats, with laboratories reporting up to 3–5x faster construct turnover through automation and parallel purification systems.

Automated construct design enables iterative refinement across multiple protein formats. Jerry Zhang, director of biology at Viva Biotech, explained that this approach is particularly valuable for structural biology: “High-throughput protein production can allow us to try different variations, and potentially it can help us to find the best construct that we are looking for that we can use for protein production or some structural biology or even some assays.”

The scale has expanded to the point where some groups now express thousands of constructs in fully automated workflows. Some laboratories have pushed automation to express up to 2,500 antibodies simultaneously, rapidly screening for optimal formats including bispecifics and tetraspecifics—a throughput level that would have been unthinkable a decade ago.

Yet as predictive tools improve, the fundamental approach may shift.

“AI will be able to predict which construct to use, which buffers to use automatically,” Wan said. “So, how will the evolution of high-throughput screening really develop as AI becomes more prominent in predicting one or two constructs or one or two protocols or expression systems rather than actually screening these huge numbers? That will be interesting to see in the future.”

AI has moved beyond structure prediction to become embedded in the full protein-engineering cycle, uniting sequence, structure, and manufacturability. The model architectures have evolved from conventional machine learning to deep neural networks, graph neural networks, and now transformer and diffusion-based models.

“Conventionally, there are several different ways to extract protein features and to present these proteins. These are mostly sequence-based,” said Yue Qian, executive director of computational chemistry and artificial intelligence platform at Viva Biotech. “But nowadays, since we have protein large language models, they provide much richer approaches to better describe these proteins. And on top of that is the recent advance of the AlphaFold-like structure prediction tools.”

This progression enables more sophisticated downstream predictions, such as protein stability, solubility, and even manufacturability. “We’ve already seen publications in the past couple of years talking about how they can improve GPCR stability and solubility. And we’ve also seen pretty accurate models to predict the protein yields based on the sequence,” Qian noted. Industry reports indicate that AI-driven protein structure prediction accuracy has improved by nearly 30% over the past two years, catalyzing broader adoption for early stage engineering.

To help bridge the gap between design and manufacturability, Viva Biotech uses a three-part AIDD platform aligned to the typical protein-engineering workflow. V-Scepter handles early parameterization for complex proteins, V-Orb brings physics-based modeling to predict binding and stability, and V-Mantle uses generative AI to design and refine new constructs. Together, these tools help teams move from concept to manufacturable candidates with greater confidence.

The integration of computational prediction with direct wet-lab validation enables rapid iteration cycles. Critically, Viva’s models are trained on both positive and negative experimental outcomes from in-house data, addressing a common limitation in AI-driven protein engineering: models trained exclusively on successful structures often fail to predict what won’t work.

However, even with these advances, membrane protein expression remains unpredictable for certain target classes, and yield predictions are still expression system-dependent. “We need to be very cautious about which system we’re looking at,” Wan said. The challenge of transferring learnings across expression platforms, from mammalian to bacterial to insect cells, is an ongoing area of development.

Next-generation protein degraders and binders demand equal mastery of chemistry and protein biology and represent a natural application area for the convergence of AI-guided design and high-throughput validation. More than 120 PROTACs and 60 molecular glues are in clinical or preclinical phases worldwide for 2025, with recent licensing deals highlighting cross-industry momentum.

“There’s a discipline of ‘linkology’, which is the generation of these huge linker-like libraries,” Wan explained. “And a lot of these linkers actually already incorporate various drug-like properties. So they’re already quite ADME and DMPK-friendly.” Direct-to-biology systems enable rapid screening of chemical reactions without purification, allowing hundreds of thousands of reactions to be screened within weeks.

Validating PROTACs requires confirming ternary complex formation: the protein of interest, the PROTAC, and the E3 ligase all bound together. Proximity biosensors using FRET enable rapid screening for these complexes, while CRISPR-based cellular assays with NanoBRET or HiBiT tags allow real-time visualization of protein degradation in cells.

Molecular glues present a greater challenge. Unlike PROTACs, which have distinct binding ligands for target and E3 ligase connected by a linker, molecular glues are single entities that induce novel protein-protein interactions. Large DNA-encoded libraries (DEL) and affinity selection mass spectrometry(ASMS) techniques enable screening for these more subtle interactions, though hit rates remain lower than for traditional binding molecules.

AI-driven mini-protein design represents another frontier, generating de novo proteins that bind to specific targets and create entirely new binding surfaces. Combining computational modeling with advanced structural biology has become central to these efforts.

For instance, researchers at Viva Biotech integrate AI-guided protein design with cryo-EM and expression optimization to tackle previously unsolved structures, such as challenging protein-protein interactions and PROTAC ternary complexes, according to Qian. These advances have enabled the structure determination of highly flexible structures, a breakthrough enabled by strategically engineered proteins that stabilize target complexes.

The field’s next major challenge is consistency and data uniformity. Despite decades of accumulated structural data, reliable prediction of expression and manufacturability remains elusive. A lack of standardized datasets is commonly cited as a top barrier to AI deployment for biologics, prompting new public–private consortia to push for benchmarking and real-world validation, such as the FAIR principles (findable, accessible, interoperable, reusable).

“Protein production and protein engineering are huge tasks, and each data point takes significant effort to generate,” said Qian. “The challenge is consistency. Different expression systems or experimental conditions can make data hard to compare directly.”

To help close this gap, companies are building training datasets that better reflect biological reality. At Viva Biotech, the protein engineering and AI teams work closely to deliver the correct protein reagents to projects, integrating both positive and negative in-house data into AI/ML models to improve predictive realism. Leveraging this proprietary dataset helps bridge in silico design and true biological behavior, enhancing manufacturability predictions.

“Protein engineering needs not just better models, but data that truly reflects biology,” she added.

Looking ahead, the next stage of protein engineering will increasingly rely on AI-driven integration of sequence, structure, and experimental feedback. At Viva Biotech, these approaches are coupled with structure-based drug design (SBDD) to anchor computational predictions in atomic-level insight, helping bridge in silico design with real biological behavior and supporting more reliable advancement of complex biologics.

This article was written in partnership with Viva Biotech.