AI Models Learn to Predict Chemical Reaction Products Using SMILES Sequences

·1 views

Researchers have developed AI approaches to forward reaction prediction, which determines what products emerge from given reactants and reagents. Two main methods are used: sequence-to-sequence models that treat reaction SMILES strings as a translation problem, and graph-based models that predict which chemical bonds change during a reaction. Philippe Schwaller and colleagues introduced the Molecular Transformer in 2019, demonstrating that token probability scores can help prioritize predictions requiring expert review. A key challenge is atom-mapping — correctly tracking where each atom ends up — since models scored only on product strings may learn pattern matching rather than true chemical transformations. Inconsistent handling of reactants versus agents across datasets also makes published benchmark accuracies difficult to compare directly.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Batch vs Streaming in Data Engineering: How to Choose the Right Approach

In data engineering, choosing between batch and streaming processing is an operational decision with real consequences for cost, complexity, and reliability. Batch processing suits use cases where latency of minutes or hours is acceptable, such as incremental ETL pipelines and business dashboards, offering simplicity and lower infrastructure costs. Streaming is better suited for time-sensitive scenarios like fraud detection or operational alerts, where data loses value within seconds, but it comes with higher complexity and continuous infrastructure overhead. A practical hybrid approach — using batch for dashboards and streaming only for fraud detection — can reduce costs while maintaining system reliability. Experts warn against common pitfalls such as assuming everything needs to be real-time, underestimating streaming complexity, and selecting technology before clearly defining the problem.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Dashboard Audit Reveals One Metric Is Reliable, Another Fluctuates by 1,000 Places

A developer tracking their browser extension discovered that a third-party dashboard reports two very different types of data quality from the same page. The extension's install count proved consistent, returning the same figure across multiple readings of the same date, while the rank metric shifted by roughly a thousand places between readings. This discrepancy only became visible after the developer switched to logging one row per reading instead of one per day, which had been masking the disagreement. The finding led to a practical rule: rank movements under a thousand places carry no meaningful signal and should not be acted upon. The developer concludes that reliability is a property of individual data fields, not of sources as a whole, and recommends repeated readings before trusting any metric used for decision-making.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Closing Disclosure Extraction Requires Regulation-Anchored Schema Design

A Closing Disclosure is a federally regulated mortgage form introduced by the Consumer Financial Protection Bureau under Regulation Z, with a fixed layout that makes structure inference unnecessary and unreliable. Because every section of closing costs is assigned a stable letter designation, extraction pipelines can anchor to those letters rather than searching for fee names, enabling checkable subtotals and meaningful detection of empty sections. The critical challenge lies in the closing cost details page, where each line item can carry up to five distinct amounts across borrower, seller, and third-party payment columns. A schema that captures only one amount per line will silently misattribute costs — for example, recording a seller-paid title premium as borrower-paid — and because totals are extracted rather than derived, the errors go undetected. The correct data model maps each line item to a column-keyed amount structure tied to its lettered section, allowing totals to be independently verified against their printed subtotals.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Extracting Clinical Trial Eligibility Criteria Is Harder Than It Looks

Clinical trial protocols contain eligibility criteria written for human readers, making automated extraction structurally complex and error-prone. Two key source documents exist — the full protocol and the trial registry entry — and they are not interchangeable, with the registry often omitting important qualifications found in the protocol. Inclusion and exclusion criteria must be kept as separate lists rather than merged with a polarity flag, since mechanically negating criteria containing internal negations can cause downstream rule engines to evaluate conditions incorrectly. Bundled criteria, where a single numbered item contains multiple evaluable conditions joined by 'and', must be split into atomic units to be reliably assessed against individual patients. Multi-arm trials add further complexity, as arm-specific criteria flattened into a single list produce a criteria set that no individual trial arm actually holds.

0 comments Read more at DEV Community