When AI Scales Faster than Systems
With nine out of ten candidates entering clinical trials ultimately failing, the pharmaceutical industry faces structural challenges that surface-level interventions cannot solve. Despite investments in digital transformations, the cost of developing a new drug has crossed the $2 billion per asset mark, while Phase-III cycle times are now 12% longer than before. At the same time, the industry’s AI value is still uneven, with only 23% of life sciences organizations achieving high or very high ROI from AI, with the majority expecting returns to remain flat in the near term.
The gap is leading to the introduction of AI-led, in silico methods to reduce wet-lab iterations early. Teams want to focus experimental validation on the most promising drug targets, including complex targets where conventional modeling and infrastructure quickly hit their limits.
This was where a Europe-based global pharmaceutical company struggled. It needed to model the 3D structure of a very large and complex protein (over 4,000 building blocks) to advance drug discovery research. Built on AlphaFold and Boltz‑2, both AI systems that predict a protein’s 3D structure from its sequence, the protein was so complex that the standard GPU machines ran out of capacity. As a result, the runs failed or produced results for only partial protein length that couldn’t be used, slowing target validation and downstream discovery work.
Across the industry, preclinical research remains heavily dependent on wet‑lab experimentation, while the AI toolchain is often fragmented and difficult to operationalize without specialist tech and domain talent and significant compute, all already hard to access. Despite rising competition and leadership pressure to demonstrate ROI, programs continue to stall in the pilot‑to‑production gap due to operational issues. These include as insufficient data readiness, limited integration into end‑to‑end scientific workflows and operating models that do not yet align R&D need, data science and IT for scaled execution.
Pi‑AFDD on Google Cloud: Operationalizing AI-led Drug Discovery, End to End
To address this challenge, Persistent partnered with Google Cloud to operationalize Persistent’s Integrated AI for Drug Discovery (Pi‑AFDD), an end-to-end, AI-driven workbench for early-stage drug discovery. As described here, Persistent’s Pi‑AlphaFold Drug Discovery, Pi‑AFDD brings together target discovery, protein structure prediction, binding-site analysis, virtual screening and smart reporting in a unified workflow, helping teams move from disease hypothesis to validated in silico evidence with greater speed, traceability and repeatability.
By integrating advanced AI models, computational tools and governed cloud execution, Pi‑AFDD reduces manual handoffs, minimizes experimental effort and enables faster, data-led decisions across target identification, structure-based analysis and hit discovery.
To address the client’s protein‑modelling challenge, Persistent and Google Cloud provisioned specialised high‑performance infrastructure:
- Advanced GPU‑accelerated environments enabled successful prediction of large multi‑chain protein complexes
- High‑memory, compute‑intensive configurations enabled modelling of protein structures incorporating metal ions—capabilities that were not achievable with prior setups
What had previously been considered a “heavy lift” became feasible through the combination of Pi‑AFDD, deep computational biology, engineering expertise and Google Cloud’s scalable HPC capabilities.
Making AI’s Impact Upstream
Pi‑AFDD demonstrates significant potential impact across preclinical drug discovery. Industry benchmarks reinforce the scale of opportunity, with AI-first leaders reducing early discovery and candidate identification from four to five years to ~eight months in some cases.
Persistent’s proven approach is pillared on:
- Demonstrated ability to solve “hard” protein-modelling problems: Accuracy in predicting extremely large, complex transmembrane proteins (>4,000 amino acids) and multi‑chain structures (AlphaFold‑Multimer), including modelling with metal ions where prior setups failed.
- Deep computational biology and AI domain expertise: Strong research and pharma experience in know-how to operationalize advanced models in real discovery workflows (not just experimentation).
- Cloud-scale HPC enablement on Google Cloud: Expertise in selecting and provisioning specialized infrastructure (e.g., a2‑ultragpu‑1g, m4‑ultramem‑224, c4d‑standard‑192) to remove compute bottlenecks and make previously infeasible workloads practical.
- End-to-end delivery with hyperscaler partnership: Joint execution with Google Cloud from idea to pilot, combining architecture guidance, scalable services and production-ready delivery—critical because value is increasingly captured by teams that can move beyond pilots into scaled adoption.
- Automation that improves throughput and traceability: Smart report builder streamlines documentation and insight generation across the workflow, supporting repeatability and faster stakeholder reviews.
For the Pharma client, we delivered:
- In silico target validation of challenging multimeric proteins, reducing the dependencies on experimental structure
- More than 50% estimated reduction in preclinical timelines by enabling identification of suitable ligand binding sites and accelerating virtual screening of molecule library
- More than 50% estimated cost reduction through minimising experimental studies
By combining Pi‑AFDD with deep domain knowledge in computational biology, strong engineering and proven Google Cloud delivery expertise, Persistent enables pharma companies to overcome challenges in 3-D protein structure modelling, virtual screening simulation thereby, de‑risking early research and accelerating the path from discovery to development.
