Source-linked AI summary
MolParser-Mobile: Ultrafast OCSR System for Large-Scale Chemical Literature Mining
Xi Fang, Haocheng Lu, Han Lyu, Chengxiang Luo, Linfeng Zhang, Guolin Ke
TL;DR
OCSR inference throughput remains a bottleneck for large-scale chemical literature mining. MolParser-Mobile redesigns the pipeline with compact models and large-scale data, reaching 1,520 molecules per second on one RTX 4090D GPU while maintaining competitive and sometimes superior accuracy.
Problem
OCSR inference throughput remains a bottleneck for web-scale chemical literature mining, despite advances in recognition accuracy.
Source: §1 · pp. 1–2Method
MolParser-Mobile redesigns the OCSR pipeline with MolDetv2, the MolGallery dataset, and an AutoML-optimized lightweight encoder-decoder trained on over 18 million images.
Source: §1 · p. 2Results
1,520 molecules per second on one NVIDIA RTX 4090D GPU, with 9.98M parameters and competitive or sometimes superior recognition accuracy.
Source: §1 · p. 2Takeaways & Limitations
At this throughput, one billion molecular images can be parsed in approximately one day using eight consumer-grade GPUs.
Source: §1 · p. 2MolGallery pseudo-labels may contain a small number of residual errors, so subsequent fine-tuning on high-quality human-annotated datasets further improves recognition accuracy.
Source: §3 · p. 4
Abstract
from arXiv · showhide
Optical Chemical Structure Recognition (OCSR) is a fundamental component of chemical literature mining, enabling molecular database construction, reaction extraction, and AI-driven scientific discovery. Despite substantial progress in recognition accuracy with recent deep learning-based methods, inference throughput remains a critical bottleneck that limits web-scale deployment. To address this challenge, we propose MolParser-Mobile, an AutoML-optimized lightweight end-to-end OCSR framework. MolParser-Mobile contains only 9.98M parameters, while reaching a throughput of 1,520 molecules per second on a single NVIDIA RTX 4090D GPU. Despite its compact design, it maintains competitive and, on several benchmarks, superior recognition accuracy.
1 Introduction
MolParser-Mobile addresses the inference-throughput bottleneck in OCSR that constrains large-scale chemical literature mining. It redesigns the pipeline and reports high throughput with competitive or superior recognition accuracy.
1 Introduction
- Prior OCSR systems remain inference-limited: traditional systems take seconds per molecule, while optimized end-to-end systems such as MolParser reach only tens of molecules per second.At one molecule per second, recognizing one billion images would consume more than 11,000 GPU-days.Source: §1 · pp. 1–2
- Large-scale chemical mining requires OCSR that balances throughput, lightweight deployment, recognition performance, and robustness across real-world document layouts.The paper argues that meeting these requirements calls for redesigning localization, training data, and architecture rather than merely compressing existing models.Source: §1 · p. 2
- MolParser-Mobile redesigns the OCSR pipeline around MolDetv2 and MolDet-Bench, the 10-million-sample MolGallery dataset, and an AutoML-optimized encoder-decoder trained on over 18 million images.MolGallery is built from chemical literature using multi-model consistency-based pseudo-labeling.Source: §1 · p. 2
- With 9.98M parameters, MolParser-Mobile reaches 1,520 molecules per second on one RTX 4090D GPU, over 38× faster than MolParser 1.0, with competitive and sometimes superior accuracy.The authors estimate one billion images can be parsed in approximately one day using eight consumer-grade GPUs.Source: §1 · p. 2
2 Molecule Detection
MolDetv2 introduces compact detectors for broad-scale molecular depictions and complete document pages, evaluated with complementary localization benchmarks. Reported results show strong accuracy–latency trade-offs across those operating regimes.
2 Molecule Detection
- Real-world document localization is a prerequisite for downstream OCSR and is challenged by dense text, multi-panel figures, tables, and reaction schemes.The original MolDet models can falsely detect non-chemical graphics, while compact variants can lose accuracy.Source: §2 · p. 2
- MolDetv2 comprises two 2.6-million-parameter YOLO11n detectors for complementary operating regimes.MolDetv2-General-n targets varied image scales at 640×640; MolDetv2-Doc-n targets small depictions on full pages at 960 × 960.Source: §2 · p. 2
- MolDet-Bench-General covers 799 varied images, while MolDet-Bench-Doc covers 447 PDF pages with 2,178 instances; BioVista adds 11,212 instances from 500 papers.General uses adaptively expanded boxes; the document benchmark uses tight, edge-aligned boxes.Source: §2 · pp. 2–3
- On MolDet-Bench-General, MolDetv2-General-n reaches 0.9872 mAP50 and 0.8776 mAP50:95 at 1.5 ms; on MolDet-Bench-Doc, MolDetv2-Doc-n reaches 0.9936 mAP50 at 3.1 ms.The figure compares accuracy–latency trade-offs across general and document-page inputs.Source: §2 · pp. 3–4, Figure 1 · p. 3
- At the same latency as MolDet-General-n, MolDetv2-General-n improves mAP50 by 3.0 percentage points and mAP50:95 by 7.2 percentage points.MolDetv2-Doc-n has accuracy comparable to larger document-specific variants while reducing latency approximately 44% versus MolDet-Doc-s; Uni-Parser-LD has higher mAP50:95, but takes 10.1 ms versus 3.1 ms.Source: §2 · pp. 3–4
3 MolGallery Dataset
MolGallery is a large-scale real-world molecular-image dataset built to support lightweight OCSR training. Its pseudo-labeling process scales through model agreement, though some residual label errors remain.
3 MolGallery Dataset
- MolGallery is designed to provide the large, diverse training data lightweight OCSR models need to preserve recognition accuracy under aggressive compression.Source: §3 · p. 4
- MolGallery contains approximately 10 million molecular images from authentic documents and online resources, paired with E-SMILES annotations.Its 193 GB corpus captures diverse layouts, rendering styles, scanning artifacts, and real-world noise.Source: §3 · p. 4
- Three OCSR models recognize detector-extracted crops, and MolGallery retains samples only when all three produce identical E-SMILES predictions.The models include two MolParser-1.0 variants and MolScribe.Source: §3 · p. 4
- Model consensus removes ambiguous and low-quality samples while enabling scalable dataset construction, but pseudo-labeling can leave residual errors.MolGallery is primarily intended for large-scale pretraining; subsequent fine-tuning on high-quality human-annotated datasets further improves recognition accuracy.Source: §3 · p. 4
4 MolParser-Mobile
MolParser-Mobile is a compact end-to-end diagram-to-E-SMILES system built around a redesigned recognition pipeline and four training stages. Its evaluation reports 38× higher GPU throughput than MolParser 1.0 while retaining competitive recognition, with gains in chirality and Markush criteria.
4 MolParser-Mobile
- MolParser-Mobile maps molecular diagrams directly to E-SMILES in a 9.98M-parameter end-to-end model trained through pretraining, supervised fine-tuning, on-policy distillation, and DPO calibration.The framework does not reconstruct an atom–bond graph or rely on extra post-processing modules.Source: §4 · pp. 4–5
4.1 From Molecular Diagrams to E-SMILES
- E-SMILES pairs an RDKit-compatible SMILES graph serialization with optional XML-like extensions for Markush groups, abbreviations, dummy atoms, and uncertain ring substitutions.The output is a formal language with syntactic and chemical constraints, so a token error can alter or invalidate a prediction.Source: §4.1 · p. 5
4.2 Architecture
- The decoder models E-SMILES syntax with causal self-attention and queries bidirectional visual memory through separate cross-attention for spatial retrieval.Cross-attention can focus on image regions as generation proceeds, supporting atom tracking, ring closure, and association of group labels with indices.Source: §4.2 · p. 5
4.3 Pre-training
- Pretraining combines over 18M pairs—7.72M mostly synthetic MolParser-7M-Pretrain pairs and 10.3M real-document MolGallery pairs—to cover rendering variation and document noise.The real-document corpus includes crop artifacts, varied drawing styles, and more chiral and Markush cases.Source: §4.3 · p. 6
4.4 Supervised Fine-tuning
- 91.2k human-annotated samples fine-tune noisy pseudo-labels, with a 200k pretraining replay buffer to preserve coverage of rare topologies.The authors report that fine-tuning raises precision on real-literature crops while replay retains rare-topology coverage.Source: §4.4 · p. 6
4.5 On-Policy Distillation
- On-policy distillation trains the 9.98M student on its own sampled trajectories using MolParser 1.5 token-level supervision, matching the student's decoding-error distribution.The approach targets imitation of the expert on long Markush captions and rare stereo patterns.Source: §4.5 · p. 6
4.6 Direct Preference Optimization
- DPO calibration uses 500 preference triples pairing corrected E-SMILES with plausible wrong outputs to target residual label errors, including Markush and chiral cases.The authors report that it aligns output format on complex Markush cases and repairs residual chiral errors after OPD.Source: §4.6 · p. 6
4.7 Efficiency and Recognition Performance
- MolParser-Mobile consistently outperforms MolParser 1.0 on most BioVista and Uni-Parser Bench criteria and achieves 38× GPU throughput improvement (1,520 vs. 39.8 samples/s) using the native PyTorch backend.Source: §4.7 · pp. 6–7, Table 2 · p. 7
5 Future Works
Future work targets more chemically expressive E-SMILES and structure-image grounding, alongside broader pretraining and chemistry-aware or adaptive-resolution improvements. These directions aim to make MolParser more robust, chemically expressive, and scalable across scientific and patent literature.
Representation Expansion
- Planned E-SMILES extensions target complex ring connectivity, repeated substructures, and uncertain or query-like bonds.Source: §5 · p. 7
Structure-Image Grounding
- The planned grounding work aligns predicted atoms and bonds with diagram locations without explicit structural-grounding annotations.Source: §5 · p. 7
Scaling and Robust Pretraining
- Larger, more diverse pretraining corpora are intended to improve robustness across heterogeneous literature, including low-quality scans and complex patent figures.Source: §5 · p. 7
Chemistry-aware Optimization
- Planned chirality-aware reinforcement learning will reward stereochemical correctness and penalize invalid molecular configurations.Source: §5 · p. 8
Adaptive Resolution
- Adaptive resolution will allocate higher image resolution to large or structurally dense molecules while maintaining pipeline efficiency.Source: §5 · p. 8