How AI is Revolutionizing Drug Discovery: Baidu’s Large‑Scale Bio‑Computing Models

1. Trends and AI challenges in biopharma

From a global perspective, the biopharma industry is experiencing rapid growth, yet it faces the “anti‑Moore’s law” problem: billions of dollars are no longer sufficient to develop a new drug. This drives strong demand for new technologies such as AI.

In China, policy support and technological advances have created new opportunities. Leading pharmaceutical companies have nearly doubled R&D investment from 2019 to 2021, and the number of Class‑1 new drugs approved in 2021 was almost three times that of 2018.

AI breakthroughs such as AlphaFold2 (AF2) in protein structure prediction have boosted confidence in AI, while the proliferation of super‑computing and intelligent‑computing infrastructure provides favorable conditions for AI adoption in biopharma.

Compared with traditional high‑throughput experiments and simulation‑based drug design, AI offers higher efficiency, but it also faces challenges: high‑cost, high‑precision experimental data, batch effects, large amounts of unlabeled data, complex multi‑task ADMET prediction, and the need for deep domain knowledge and powerful computing resources.

2. Building Bio‑Computing Large Models

Baidu constructs a “data‑and‑first‑principles dual‑driven” pre‑training model (Wenxin Bio‑Computing) that learns massive biochemical knowledge at the atomic level, enabling better extrapolation and generalization for compounds, proteins, and RNA.

By integrating first‑principles physics/chemistry into affinity prediction tasks, Baidu trains a high‑extrapolation affinity model that improves small‑molecule and peptide/protein design.

The base model can be fine‑tuned with a small amount of high‑precision data for specific downstream tasks.

Baidu’s HelixGEM series exemplify this approach. HelixGEM‑1, published in Nature Machine Intelligence, was trained on 20 million data points and is the first neural network to incorporate 3‑D geometric conformations of compounds, achieving state‑of‑the‑art performance on 14 drug‑related benchmarks.

HelixGEM‑2 further incorporates many‑body and long‑range interactions inspired by quantum‑mechanical simulations, becoming the first model to consider atomic multi‑body and long‑range relations, and achieving SOTA on PCQM‑4mv2 and Lit‑PCBA datasets.

HelixADMET, built on HelixGEM‑1 with multi‑task learning and knowledge transfer, outperforms other methods by over 4 % on ADMET prediction, as reported in Bioinformatics.

In the protein domain, Baidu re‑trained AlphaFold2 (AF2) to create HelixFold, which surpasses AF2 by 1–2 percentage points in accuracy and doubles inference speed on GPUs; on domestic DCU hardware, a full training round finishes in 2.6 days.

To address AF2’s slow MSA retrieval, Baidu and Baidu Biotech developed HelixFold‑Single, a language‑model‑based single‑sequence predictor that accelerates inference by hundreds of times while matching AF2’s accuracy, and even exceeds it on antibody structure prediction.

For protein‑protein interaction (PPI), Baidu introduced a multimodal pre‑training technique that encodes sequence, structure, and function using atomic point‑cloud topology, achieving 5–10 % improvements over previous methods on cross‑species PPI, antigen‑antibody affinity, and mutation‑driven binding prediction.

3. Industrial Practice of PaddleHelix

Baidu applied these large‑model technologies in real drug‑discovery pipelines. In a PPI disruption project targeting CDK4/6‑Cyclin D1, virtual screening of 7.8 million molecules yielded 110 candidates, of which 40 were tested and 6 successfully disrupted the interaction.

A conventional small‑molecule screening of 200 k compounds identified a novel scaffold after testing seven compounds.

In protein research, collaborations with Chinese agricultural institutes used the structure‑prediction models to study wheat protein regulation, and a major hospital employed the models to analyze a >4000‑residue protein for disease mechanisms.

Baidu’s PaddleHelix platform now offers an end‑to‑end bio‑computing solution, from algorithms and cloud scheduling to cloud‑integrated products, and provides open‑source access to many models.

In mRNA vaccine design, the LinearDesign algorithm, developed early in the COVID‑19 pandemic, optimizes both minimum free energy and codon adaptation index, achieving higher expression and immunogenicity; the method has been commercialized by Sanofi.

Overall, Baidu continues to build a data ecosystem and cloud‑native bio‑computing services to empower the biopharma industry.