Scaling up AI across the life sciences value chain

Research and development

Drug discovery involves a series of laborious, manual, and capital-intensive processes that can be automated or augmented through the use of AI. Biopharma companies are already beginning to experiment with AI to speed up drug discovery activities (see case study 1). In the next five years, companies could increasingly use AI models to identify and validate targets, design molecules, synthesize and test these molecules in silico, and feed data back into these models to improve their predictive capabilities. This could vastly accelerate the traditionally slow drug discovery process.

The growing abundance of genotypic and phenotypic data has created a need for techniques that can transform and expedite target identification and validation. AI imaging capabilities can detect cell morphology changes that humans cannot see through a microscope. AI applied to knowledge graphs could help understand complex relationships between compounds, genes, diseases, and proteins. Validating targets through functional genomics experiments could make extensive use of AI to interpret phenotypic changes in cells via imaging, and integrate disparate readouts such as expression assays, cell images, and epigenetics (noncoding changes that affect expression). To prevent expensive failures in late-stage drug development, generating genetic evidence for target identification will become an increasingly important strategic priority.¹

In the molecular design space, several startups are already pioneering the use of generative modeling for small molecule design and protein engineering through partnerships with large biopharma companies. At the same time, biopharma companies are also building their own capabilities to augment medicinal chemists in their design work. Over the next five years, generative modeling² could become an important part of computational chemistry toolkits, enabling companies to explore novel spaces and broaden the pool of drug candidates.

Today, creating trial artifacts (such as case report forms and study reports) requires manually inputting data into multiple systems. This leads to inconsistencies and errors and the need for rework, slowing down trial execution. AI-driven digital data flow solutions could integrate trial data from multiple source systems and documents to create standardized digital data elements for transmission to downstream systems. These data elements can then be used to auto-populate required reports and analyses and generate content for trial artifact creation. Some organizations are already exploring the feasibility of using AI to better manage clinical trial data (see case study 2).

Targeting patient populations to demonstrate efficacy will become an important aspect of clinical trial design over the next decade, speeding up the development process and reducing the number of failures. AI will be increasingly leveraged to normalize data from different platforms (e.g., gene expression data) and integrate multiple data points (genotypic, imaging, clinical records, and epidemiological) for patient stratification and identification of patients most likely to improve the probability of trial success. AI-enabled companion diagnostics tools (e.g., those used to screen for biomarkers) could be used to better understand the profile of response during clinical research and then give clinical decision support upon the approval/commercialization of the medicine.

Manufacturing

Fixing manufacturing issues today requires laborious manual intervention to access multiple systems, with action taken only after problems occur. Applying AI to manufacturing data could help predict process bottlenecks, identify quality control issues, and proactively suggest corrective actions. Applying AI technologies can help reduce manual oversight in manufacturing operations and allow tighter control of quality and operating costs through:

• Proactive quality control by applying machine learning (ML) to assess manufacturing data from multiple batches, product lines to identify process discrepancies, and predict quality issues. This can direct staff to investigate only those batches most likely to have quality issues, saving time and resources.
• Coupling AI with robotic process automation (RPA) to automate batch release and documentation can generate a comprehensive and auditable data trail to meet regulatory and compliance requirements.
• AI-driven simulations and modeling that assess various parameters during the manufacturing process to enable corrective actions to optimize yield and output. Our research suggests AI is uniquely suited to assist in improving antibody bioreactor productivity. Applied to imaging data of modified cell cultures, AI could help predict which cell lines are likely to fail early and improve the productivity of the bioreactor.
• AI-enabled predictive maintenance activity could reduce machine downtime, production disruptions, and loss of expensive API materials. This will help reduce operating costs.

Enabling these applications calls for investing in aggregating data from multiple manufacturing systems and thoughtful placement of sensors across the manufacturing floor. Some early movers are already beginning to report benefits from applying AI to manufacturing activities (see case study 3).

Supply chain

Today, many companies respond reactively to supply chain disruptions and are slow to adjust inventory and production levels. With the pandemic exposing the fragility of the biopharma supply chains and new product types (e.g., cell therapies) with complex logistical requirements entering the market, there is an urgent need to improve supply chain visibility and adaptability.

Early movers are already beginning to tap the potential of AI to automate demand and supply planning (see case study 4). AI could automate analysis of aggregated manufacturing, supply chain, and marketing data to predict demand and supply, recommend the next best action to supply chain operators, and even autonomously perform certain activities. Such self-healing AI supply chain solutions could improve the ability to dynamically respond to changes in market demand and supply availability, enable quick recovery from disruptions, and improve decision-making related to product distribution and new product introductions.

AI will also become an important enabler for the industry to transition away from the use of linear supply chains to dynamic digital supply networks (DSNs). DSNs leverage digital technologies such as sensors, RPA, blockchain, advanced analytics, and AI to collate data from multiple sources and locations in real time to optimize production and distribution (for more see Deloitte’s publications on digital supply networks and intelligent supply chain).

Commercial

Current biopharma consumer engagement strategies across channels (TV, social media, print, web, sales representatives) tend to address broad groups rather than precisely target customers based on individual needs and preferences. This often results in irrelevant and unwanted engagement, leading to low returns on marketing spending.

Omnichannel AI engagement solutions applied to analyze multiple data sources (including socioeconomic, demographic, location, medical history, and sales data) can predict how, when, and with what types of messages to engage with patient and health care practitioners (HCPs). This could enable personalized just-in-time content such as suggestions on a patient’s smartphone when they enter a pharmacy or read a WebMD article. While engaging HCPs, omnichannel AI could provide recommendations to marketing and sales representatives on next best actions, channels, and personalized content to engage HCPs. A few companies are already experimenting with AI to optimize interactions with HCPs (see case study 5).

Beyond marketing, AI can be used to analyze patient complaints, medical inquiries, and social media data to incorporate the voice of the patient into product iterations (see case study 6). Natural language processing (NLP) applied to structured and unstructured data across silos can create a “voice of the market” feedback loop for key business areas such as medical affairs, market access, sales, and brand teams. Our research also suggests conversational AI technologies are being used increasingly to advance patient literacy by equipping patients with de-jargonized insights to help them understand treatment options and trade-offs.

Biopharma companies routinely interact with HCPs through contractual arrangements such as paying a standard fee for participation in an event or provision of services. These contracts are subject to regulatory and legislative scrutiny. Companies could increasingly use software bots, NLP, and machine learning to review contracts to ensure they comply with internal norms and external regulations. Current systems can accurately and confidently make “approve/reject” decisions for more than 70% of contracts.

AI solutions are also being used as part of cyber strategy.⁹ Modern cyberattacks can circumvent traditional, rule-based security controls by learning detection rules. Smart cyber technologies will be increasingly needed to protect the growing volume of data biopharma companies own and have access to. AI-based approaches can improve threat intelligence and prediction, and enable faster attack detection and response, augmenting efforts by cybersecurity experts (for more, see Deloitte’s publication on Smart Cyber: How AI can manage cyber risk).¹⁰

AI investments by life sciences organizations varied in 2019. More than 40% of survey respondents reported their organizations invested US$20–50 million, while 21% invested more than US$50 million on AI initiatives. More than half of the survey respondents expected investments in AI to continue to increase in 2020. We believe investments in R&D are likely to continue to accelerate because of its central role in developing new products and intellectual property. As mentioned earlier, we also expect investment in AI to grow across the value chain as more executives recognize how critical digital transformation is to their companies’ future.

Unlike ERP systems that begin showing returns in a few months or a year, AI implementations require time to ramp up and provide benefits; 46% of respondents say it takes longer than expected to receive payback from investments in AI initiatives. Training an AI model is an ongoing process, and as the model is trained over time, returns on investment gradually increase. Executives therefore should account for model learning while calculating payback periods on AI projects.

Top priorities include enhancing existing products and services, creating new products and services, and increasing process efficiency

Forty-three percent of survey respondents reported that AI implementations resulted in making processes more efficient. When asked about the outcomes companies are trying to achieve through AI use, surveyed executives cited enhancing existing products and services (28%) as a top priority. This was closely followed by creating new products and services (27%) and increasing process efficiencies (26%). While we agree that AI can enhance existing products and make business more efficient, we expect AI use in creating new products to grow as companies embrace more advanced AI applications and make it a central part of their business transformation.

Top challenges include identifying high-value use cases, AI integration, and data and risk issues

Survey respondents pointed to multiple challenges associated with AI implementations, including poor-quality data, siloed data systems and integrating AI into legacy systems, and risk management. Top challenges for life sciences companies include identifying use cases with the greatest business value (30%), integrating AI into the organization (28%), and data challenges (28%). Our interviewees also said finding talent with industry expertise and AI skill sets is difficult.

Surveyed respondents also pointed to the lack of explainability of AI algorithms or the AI black box issue as a major concern. To address this, there is a need to demonstrate that AI during its learning process has not deviated from providing a valid or intended output. This will help not only gain regulatory approval to expand the use of AI but also lead to greater trust in and use of AI systems across the industry.