Smarter mAb Development with AI.

Ghosh Rajat Kumar: Good evening and good morning, everyone. Biologics Forum began in 2014 as a place for scientific discussion around biologics development and process characterization. The community has grown to approximately 850 members. This webinar continues the forum's expert-led discussion series and focuses on the use of AI in biologics development.

Ghosh Rajat Kumar: Rajiv, thank you for your time and for bringing this topic to the forum. The central idea—made in India and made for the world—is an exciting one. Over to you.

Rajiv Sondhi: Today we will focus on a practical question: how can we improve the scientific decisions that shape mAb development, from cell culture and experiments through scale-up? This is not an introduction to AI, and we will not debate whether AI belongs in pharma. We will examine where it changes the decision flow and where it acts as an enabler.

Rajiv Sondhi: We will walk through an illustrative case study and leave time for questions. The live session also included a BioMedAna demonstration; detailed product workflows are reserved for personalized sessions.

Rajiv Sondhi: Three signals show where the industry is heading. First, AI is becoming infrastructure. Second, laboratory systems are moving toward closed-loop learning, where models can help design and execute experiments and use the resulting evidence to improve the next decision.

Rajiv Sondhi: Third, regulators are publishing principles for responsible AI across development and manufacturing. The question is shifting from whether AI will be used to how it will be connected, validated and governed.

Rajiv Sondhi: Monoclonal antibodies are a critical biological modality that has delivered major clinical value and continues to do so. Every new program still carries demanding questions around productivity, product quality, manufacturability and development time. The opportunity is concrete: help teams bring stronger and more reliable processes forward while accelerating the development cycle.

Rajiv Sondhi: The test is whether an analysis changes the next scientific decision or merely creates another dashboard. BioMedAna connects evidence, models and assistance around the next action a team needs to consider.

Rajiv Sondhi: Biology is dynamic and nonlinear. Data matters, but so do viability, aggregation, robustness, scalability, cost, productivity and yield. A decision that looks good against one endpoint or at one scale may create a new risk elsewhere.

Rajiv Sondhi: The goal is not simply to predict an outcome. It is to recommend an informative experiment and identify an operating region that improves the process as a whole. mAb development needs a holistic view rather than a piecemeal optimization of individual responses.

Rajiv Sondhi: AI can support decisions from discovery and wet-lab development through process development, PPQ, scale-up, technology transfer and commercial manufacturing. The use case changes at each stage: molecule or clone selection early on, experimental and operating-condition decisions during process development, quality during PPQ, and engineering compatibility, monitoring and control during scale-up and transfer.

Rajiv Sondhi: The common requirement is continuity. Evidence should carry forward even when each phase uses different tools. The starting point is never the model; it is connected process data, biological data, experimental records and interface systems with scientific meaning and lineage.

Rajiv Sondhi: Predictive, mechanistic and hybrid models should answer decision questions: what is driving the result, which experiment should run next, what risk may emerge at scale, and what should the scientist review? Governance, uncertainty and explainability must sit across the entire evidence-to-decision chain.

Rajiv Sondhi: Consider a CHO fed-batch process with controllable variables such as temperature, pH and feed, and outcomes including titer, viability, glycosylation and aggregation. A purely exploitative strategy chooses the condition with the highest predicted performance. A purely exploratory strategy chooses the condition with the greatest uncertainty.

Rajiv Sondhi: A balanced strategy considers expected process improvement and information gain together. The selected run may not have the highest predicted titer, but it can have the highest total value once learning is included. The model recommends; the scientist reviews and approves.

Rajiv Sondhi: The highest-titer condition is not automatically the best process. A useful operating region must balance productivity, viability, aggregation, quality, robustness and manufacturability. Moving in one direction may improve titer while weakening quality.

Rajiv Sondhi: A Pareto frontier makes those trade-offs visible. AI can help identify candidate regions where several objectives are jointly acceptable and filter noise so scientists can focus on the signal. The target is a stable region that tolerates normal process variation, not a condition that succeeds only under ideal circumstances.

Rajiv Sondhi: A model can observe process history, cell state, metabolism and batch context. A prediction made at the end of a run has limited value because the outcome is already fixed. A forecast becomes useful when it arrives early enough to change sampling, feeding, gas strategy or the investigation plan.

Rajiv Sondhi: A confident line beyond the observed evidence is not enough. The team needs to know when the forecast becomes uncertain and what additional evidence could reduce that uncertainty. That allows scientists to change appropriate process levers while there is still time to affect the outcome.

Rajiv Sondhi: Traditional scale-up relies on rules, correlations and expert experience, and that expertise remains essential. AI adds a learning layer by combining biological signals—such as viable cell density and metabolism—with engineering descriptors such as mixing time, vessel geometry and gas transfer.

Rajiv Sondhi: The practical framework is transfer, adapt and relearn. Stable knowledge can be reused; scale-dependent relationships must be adjusted; and targeted experiments can reduce uncertainty. A model should not start from zero at every scale when relevant biological and process knowledge already exists.

Rajiv Sondhi: Scientists define the objectives, constraints and acceptable risk. AI and automation integrate the current evidence, predict outcomes and rank candidate actions. Once an experiment is reviewed, selected and executed, the new evidence updates the process understanding and the model.

Rajiv Sondhi: Each experiment therefore does two jobs: it improves the immediate process and reduces uncertainty for the next decision. Data context, models, agents and governed workflows should work together around one strong scientific recommendation, with the expert retaining the approval boundary.

Rajiv Sondhi: Adoption should follow a crawl–walk–run progression. Start with a focused problem and a connected evidence foundation rather than attempting the most complex hybrid twin first. Demonstrate a decision improvement, then expand the models, scenarios and automation as confidence grows.

Rajiv Sondhi: The objective is measurable scientific and operational value, not technology for its own sake. Teams should select use cases where better evidence, earlier prediction or a more focused experiment can influence a real development decision.

Rajiv Sondhi: The case begins with a CHO fed-batch program showing strong productivity but uncertainty around glycosylation and aggregation. The next experiment is chosen to test the drivers of those CQAs while preserving the productivity signal. The result is a more robust operating region rather than a single peak-performance point.

Rajiv Sondhi: Development-scale confirmation is followed by variability at pilot scale. The investigation considers mixing, local concentration gradients, gas transfer and other scale-dependent exposures. A hybrid model combines biological observations, process history and engineering descriptors to estimate where risk may emerge at 2,000 liters.

Rajiv Sondhi: The model indicates which relationships can transfer, which require adaptation and where a targeted scale-down experiment can generate the missing evidence. Each new run strengthens both process understanding and the next scale-up decision.

Rajiv Sondhi: Digital twins, AI/ML platforms and agentic systems are often implemented separately. A useful workflow connects them so teams can orient to a program, detect what changed, investigate contributing evidence, evaluate scenarios and recommend a next action.

Rajiv Sondhi: The BioMedAna Hub provides connected scientific context and lineage; modeling and simulation evaluate scenarios; agents support investigation and explanation; and the operating layer provides governance and control. The live webinar included a high-level product demonstration, while detailed workflows remain reserved for tailored sessions.

Audience question: Can the approach be adopted for microbial products, and how does it integrate with existing DCS or SCADA systems?

Rajiv Sondhi: The workflow can be adapted to microbial and other biologic modalities. Prebuilt connectors can bring data from SCADA, DCS, bioreactors, historians, ELNs and LIMS into the Hub. Once the evidence is contextualized, modeling and analysis can operate on that connected foundation. The specific integration pattern depends on the source system, protocol and validation requirements.

Audience question: Can organizations bring their own models?

Rajiv Sondhi: A language-model wrapper alone is not sufficient for complex biological decisions. Physics, biology, process data, machine learning and agent-supported reasoning need to work together. BioMedAna uses a mixture-of-experts approach that can coordinate in-house and external models according to the problem.

Rajiv Sondhi: The recommendation is only part of the output. Teams need to understand how it was produced and whether the supporting evidence is defensible. That is particularly important when probabilistic models can answer differently as the question and context change.

Audience question: Can an organization use an existing biologics data lake? Rajiv explained that a governed warehouse or lake can accelerate adoption because teams may not need to rebuild every source-system connection. The remaining work is to map the data into the scientific context required by the decision workflow.

Rajiv Sondhi: Models can be tuned for a client's process and use case while keeping client-specific evidence private and controlled. Validation can begin retrospectively: compare outputs against previous runs, manual analyses and known outcomes, then investigate where and why results differ.

Rajiv Sondhi: Confidence comes from the explanation and supporting facts, not only from a score. Scientists may accept or reject a recommendation, but they should be able to inspect its reasoning, evidence and uncertainty.

Audience question: What models are integrated, and can process models be linked to cost-of-goods models?

Rajiv Sondhi: BioMedAna combines in-house machine-learning models, selected open-source models and language models used for agent-supported tasks. The Hub, Twin and process-simulation capabilities are designed to work together so scientific evidence, predictions and scenarios remain connected. The precise model and COGS configuration depends on the organization's use case and data environment.

Rajiv Sondhi: Organizations can deploy in the cloud or on premises and can begin with a smaller approved user group or a broader enterprise model. Access is managed through organizational accounts.

Rajiv Sondhi: Cloud deployments can align with regional data-residency requirements. The platform can operate across major cloud environments or in colocated and on-premises infrastructure, with the deployment pattern selected alongside the customer's technology and governance teams.

Rajiv Sondhi: Participants are encouraged to request complimentary access, try the platform and provide feedback. Teams seeking answers to a specific scientific question or a demonstration tailored to their process and use case can contact BioMedAna directly.

Ghosh Rajat Kumar: The session made a complex topic practical, relevant and exciting. AI is not only about automation; it can help scientists and engineers make smarter decisions, uncover insights faster and ultimately bring better medicines to patients sooner.